Skip to content

GitLab

Commit 54a066bf authored by mashun1's avatar mashun1
Browse files

ootdiffusion

parents
Pipeline #1004 canceled with stages
Hide whitespace changes
Inline Side-by-side
Showing with 11086 additions and 0 deletions
ootd/pipelines_ootd/transformer_garm_2d.py 0 → 100644
View file @ 54a066bf
# Copyright 2023 The HuggingFace Team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# Modified by Yuhao Xu for OOTDiffusion (https://github.com/levihsu/OOTDiffusion)
from dataclasses import dataclass
from typing import Any, Dict, Optional
import torch
import torch.nn.functional as F
from torch import nn
from .attention_garm import BasicTransformerBlock
from diffusers.configuration_utils import ConfigMixin, register_to_config
from diffusers.models.embeddings import ImagePositionalEmbeddings
from diffusers.utils import USE_PEFT_BACKEND, BaseOutput, deprecate
# from diffusers.models.attention import BasicTransformerBlock
from diffusers.models.embeddings import CaptionProjection, PatchEmbed
from diffusers.models.lora import LoRACompatibleConv, LoRACompatibleLinear
from diffusers.models.modeling_utils import ModelMixin
from diffusers.models.normalization import AdaLayerNormSingle
@dataclass
class Transformer2DModelOutput(BaseOutput):
"""
The output of [`Transformer2DModel`].
Args:
sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)` or `(batch size, num_vector_embeds - 1, num_latent_pixels)` if [`Transformer2DModel`] is discrete):
The hidden states output conditioned on the `encoder_hidden_states` input. If discrete, returns probability
distributions for the unnoised latent pixels.
"""
sample: torch.FloatTensor
class Transformer2DModel(ModelMixin, ConfigMixin):
"""
A 2D Transformer model for image-like data.
Parameters:
num_attention_heads (`int`, *optional*, defaults to 16): The number of heads to use for multi-head attention.
attention_head_dim (`int`, *optional*, defaults to 88): The number of channels in each head.
in_channels (`int`, *optional*):
The number of channels in the input and output (specify if the input is **continuous**).
num_layers (`int`, *optional*, defaults to 1): The number of layers of Transformer blocks to use.
dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
cross_attention_dim (`int`, *optional*): The number of `encoder_hidden_states` dimensions to use.
sample_size (`int`, *optional*): The width of the latent images (specify if the input is **discrete**).
This is fixed during training since it is used to learn a number of position embeddings.
num_vector_embeds (`int`, *optional*):
The number of classes of the vector embeddings of the latent pixels (specify if the input is **discrete**).
Includes the class for the masked latent pixel.
activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to use in feed-forward.
num_embeds_ada_norm ( `int`, *optional*):
The number of diffusion steps used during training. Pass if at least one of the norm_layers is
`AdaLayerNorm`. This is fixed during training since it is used to learn a number of embeddings that are
added to the hidden states.
During inference, you can denoise for up to but not more steps than `num_embeds_ada_norm`.
attention_bias (`bool`, *optional*):
Configure if the `TransformerBlocks` attention should contain a bias parameter.
"""
@register_to_config
def __init__(
self,
num_attention_heads: int = 16,
attention_head_dim: int = 88,
in_channels: Optional[int] = None,
out_channels: Optional[int] = None,
num_layers: int = 1,
dropout: float = 0.0,
norm_num_groups: int = 32,
cross_attention_dim: Optional[int] = None,
attention_bias: bool = False,
sample_size: Optional[int] = None,
num_vector_embeds: Optional[int] = None,
patch_size: Optional[int] = None,
activation_fn: str = "geglu",
num_embeds_ada_norm: Optional[int] = None,
use_linear_projection: bool = False,
only_cross_attention: bool = False,
double_self_attention: bool = False,
upcast_attention: bool = False,
norm_type: str = "layer_norm",
norm_elementwise_affine: bool = True,
norm_eps: float = 1e-5,
attention_type: str = "default",
caption_channels: int = None,
):
super().__init__()
self.use_linear_projection = use_linear_projection
self.num_attention_heads = num_attention_heads
self.attention_head_dim = attention_head_dim
inner_dim = num_attention_heads * attention_head_dim
conv_cls = nn.Conv2d if USE_PEFT_BACKEND else LoRACompatibleConv
linear_cls = nn.Linear if USE_PEFT_BACKEND else LoRACompatibleLinear
# 1. Transformer2DModel can process both standard continuous images of shape `(batch_size, num_channels, width, height)` as well as quantized image embeddings of shape `(batch_size, num_image_vectors)`
# Define whether input is continuous or discrete depending on configuration
self.is_input_continuous = (in_channels is not None) and (patch_size is None)
self.is_input_vectorized = num_vector_embeds is not None
self.is_input_patches = in_channels is not None and patch_size is not None
if norm_type == "layer_norm" and num_embeds_ada_norm is not None:
deprecation_message = (
f"The configuration file of this model: {self.__class__} is outdated. `norm_type` is either not set or"
" incorrectly set to `'layer_norm'`.Make sure to set `norm_type` to `'ada_norm'` in the config."
" Please make sure to update the config accordingly as leaving `norm_type` might led to incorrect"
" results in future versions. If you have downloaded this checkpoint from the Hugging Face Hub, it"
" would be very nice if you could open a Pull request for the `transformer/config.json` file"
)
deprecate("norm_type!=num_embeds_ada_norm", "1.0.0", deprecation_message, standard_warn=False)
norm_type = "ada_norm"
if self.is_input_continuous and self.is_input_vectorized:
raise ValueError(
f"Cannot define both `in_channels`: {in_channels} and `num_vector_embeds`: {num_vector_embeds}. Make"
" sure that either `in_channels` or `num_vector_embeds` is None."
)
elif self.is_input_vectorized and self.is_input_patches:
raise ValueError(
f"Cannot define both `num_vector_embeds`: {num_vector_embeds} and `patch_size`: {patch_size}. Make"
" sure that either `num_vector_embeds` or `num_patches` is None."
)
elif not self.is_input_continuous and not self.is_input_vectorized and not self.is_input_patches:
raise ValueError(
f"Has to define `in_channels`: {in_channels}, `num_vector_embeds`: {num_vector_embeds}, or patch_size:"
f" {patch_size}. Make sure that `in_channels`, `num_vector_embeds` or `num_patches` is not None."
)
# 2. Define input layers
if self.is_input_continuous:
self.in_channels = in_channels
self.norm = torch.nn.GroupNorm(num_groups=norm_num_groups, num_channels=in_channels, eps=1e-6, affine=True)
if use_linear_projection:
self.proj_in = linear_cls(in_channels, inner_dim)
else:
self.proj_in = conv_cls(in_channels, inner_dim, kernel_size=1, stride=1, padding=0)
elif self.is_input_vectorized:
assert sample_size is not None, "Transformer2DModel over discrete input must provide sample_size"
assert num_vector_embeds is not None, "Transformer2DModel over discrete input must provide num_embed"
self.height = sample_size
self.width = sample_size
self.num_vector_embeds = num_vector_embeds
self.num_latent_pixels = self.height * self.width
self.latent_image_embedding = ImagePositionalEmbeddings(
num_embed=num_vector_embeds, embed_dim=inner_dim, height=self.height, width=self.width
)
elif self.is_input_patches:
assert sample_size is not None, "Transformer2DModel over patched input must provide sample_size"
self.height = sample_size
self.width = sample_size
self.patch_size = patch_size
interpolation_scale = self.config.sample_size // 64 # => 64 (= 512 pixart) has interpolation scale 1
interpolation_scale = max(interpolation_scale, 1)
self.pos_embed = PatchEmbed(
height=sample_size,
width=sample_size,
patch_size=patch_size,
in_channels=in_channels,
embed_dim=inner_dim,
interpolation_scale=interpolation_scale,
)
# 3. Define transformers blocks
self.transformer_blocks = nn.ModuleList(
[
BasicTransformerBlock(
inner_dim,
num_attention_heads,
attention_head_dim,
dropout=dropout,
cross_attention_dim=cross_attention_dim,
activation_fn=activation_fn,
num_embeds_ada_norm=num_embeds_ada_norm,
attention_bias=attention_bias,
only_cross_attention=only_cross_attention,
double_self_attention=double_self_attention,
upcast_attention=upcast_attention,
norm_type=norm_type,
norm_elementwise_affine=norm_elementwise_affine,
norm_eps=norm_eps,
attention_type=attention_type,
)
for d in range(num_layers)
]
)
# 4. Define output layers
self.out_channels = in_channels if out_channels is None else out_channels
if self.is_input_continuous:
# TODO: should use out_channels for continuous projections
if use_linear_projection:
self.proj_out = linear_cls(inner_dim, in_channels)
else:
self.proj_out = conv_cls(inner_dim, in_channels, kernel_size=1, stride=1, padding=0)
elif self.is_input_vectorized:
self.norm_out = nn.LayerNorm(inner_dim)
self.out = nn.Linear(inner_dim, self.num_vector_embeds - 1)
elif self.is_input_patches and norm_type != "ada_norm_single":
self.norm_out = nn.LayerNorm(inner_dim, elementwise_affine=False, eps=1e-6)
self.proj_out_1 = nn.Linear(inner_dim, 2 * inner_dim)
self.proj_out_2 = nn.Linear(inner_dim, patch_size * patch_size * self.out_channels)
elif self.is_input_patches and norm_type == "ada_norm_single":
self.norm_out = nn.LayerNorm(inner_dim, elementwise_affine=False, eps=1e-6)
self.scale_shift_table = nn.Parameter(torch.randn(2, inner_dim) / inner_dim**0.5)
self.proj_out = nn.Linear(inner_dim, patch_size * patch_size * self.out_channels)
# 5. PixArt-Alpha blocks.
self.adaln_single = None
self.use_additional_conditions = False
if norm_type == "ada_norm_single":
self.use_additional_conditions = self.config.sample_size == 128
# TODO(Sayak, PVP) clean this, for now we use sample size to determine whether to use
# additional conditions until we find better name
self.adaln_single = AdaLayerNormSingle(inner_dim, use_additional_conditions=self.use_additional_conditions)
self.caption_projection = None
if caption_channels is not None:
self.caption_projection = CaptionProjection(in_features=caption_channels, hidden_size=inner_dim)
self.gradient_checkpointing = False
def forward(
self,
hidden_states: torch.Tensor,
spatial_attn_inputs = [],
encoder_hidden_states: Optional[torch.Tensor] = None,
timestep: Optional[torch.LongTensor] = None,
added_cond_kwargs: Dict[str, torch.Tensor] = None,
class_labels: Optional[torch.LongTensor] = None,
cross_attention_kwargs: Dict[str, Any] = None,
attention_mask: Optional[torch.Tensor] = None,
encoder_attention_mask: Optional[torch.Tensor] = None,
return_dict: bool = True,
):
"""
The [`Transformer2DModel`] forward method.
Args:
hidden_states (`torch.LongTensor` of shape `(batch size, num latent pixels)` if discrete, `torch.FloatTensor` of shape `(batch size, channel, height, width)` if continuous):
Input `hidden_states`.
encoder_hidden_states ( `torch.FloatTensor` of shape `(batch size, sequence len, embed dims)`, *optional*):
Conditional embeddings for cross attention layer. If not given, cross-attention defaults to
self-attention.
timestep ( `torch.LongTensor`, *optional*):
Used to indicate denoising step. Optional timestep to be applied as an embedding in `AdaLayerNorm`.
class_labels ( `torch.LongTensor` of shape `(batch size, num classes)`, *optional*):
Used to indicate class labels conditioning. Optional class labels to be applied as an embedding in
`AdaLayerZeroNorm`.
cross_attention_kwargs ( `Dict[str, Any]`, *optional*):
A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
`self.processor` in
[diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
attention_mask ( `torch.Tensor`, *optional*):
An attention mask of shape `(batch, key_tokens)` is applied to `encoder_hidden_states`. If `1` the mask
is kept, otherwise if `0` it is discarded. Mask will be converted into a bias, which adds large
negative values to the attention scores corresponding to "discard" tokens.
encoder_attention_mask ( `torch.Tensor`, *optional*):
Cross-attention mask applied to `encoder_hidden_states`. Two formats supported:
* Mask `(batch, sequence_length)` True = keep, False = discard.
* Bias `(batch, 1, sequence_length)` 0 = keep, -10000 = discard.
If `ndim == 2`: will be interpreted as a mask, then converted into a bias consistent with the format
above. This bias will be added to the cross-attention scores.
return_dict (`bool`, *optional*, defaults to `True`):
Whether or not to return a [`~models.unet_2d_condition.UNet2DConditionOutput`] instead of a plain
tuple.
Returns:
If `return_dict` is True, an [`~models.transformer_2d.Transformer2DModelOutput`] is returned, otherwise a
`tuple` where the first element is the sample tensor.
"""
# ensure attention_mask is a bias, and give it a singleton query_tokens dimension.
# we may have done this conversion already, e.g. if we came here via UNet2DConditionModel#forward.
# we can tell by counting dims; if ndim == 2: it's a mask rather than a bias.
# expects mask of shape:
# [batch, key_tokens]
# adds singleton query_tokens dimension:
# [batch, 1, key_tokens]
# this helps to broadcast it as a bias over attention scores, which will be in one of the following shapes:
# [batch, heads, query_tokens, key_tokens] (e.g. torch sdp attn)
# [batch * heads, query_tokens, key_tokens] (e.g. xformers or classic attn)
if attention_mask is not None and attention_mask.ndim == 2:
# assume that mask is expressed as:
# (1 = keep, 0 = discard)
# convert mask into a bias that can be added to attention scores:
# (keep = +0, discard = -10000.0)
attention_mask = (1 - attention_mask.to(hidden_states.dtype)) * -10000.0
attention_mask = attention_mask.unsqueeze(1)
# convert encoder_attention_mask to a bias the same way we do for attention_mask
if encoder_attention_mask is not None and encoder_attention_mask.ndim == 2:
encoder_attention_mask = (1 - encoder_attention_mask.to(hidden_states.dtype)) * -10000.0
encoder_attention_mask = encoder_attention_mask.unsqueeze(1)
# Retrieve lora scale.
lora_scale = cross_attention_kwargs.get("scale", 1.0) if cross_attention_kwargs is not None else 1.0
# 1. Input
if self.is_input_continuous:
batch, _, height, width = hidden_states.shape
residual = hidden_states
hidden_states = self.norm(hidden_states)
if not self.use_linear_projection:
hidden_states = (
self.proj_in(hidden_states, scale=lora_scale)
if not USE_PEFT_BACKEND
else self.proj_in(hidden_states)
)
inner_dim = hidden_states.shape[1]
hidden_states = hidden_states.permute(0, 2, 3, 1).reshape(batch, height * width, inner_dim)
else:
inner_dim = hidden_states.shape[1]
hidden_states = hidden_states.permute(0, 2, 3, 1).reshape(batch, height * width, inner_dim)
hidden_states = (
self.proj_in(hidden_states, scale=lora_scale)
if not USE_PEFT_BACKEND
else self.proj_in(hidden_states)
)
elif self.is_input_vectorized:
hidden_states = self.latent_image_embedding(hidden_states)
elif self.is_input_patches:
height, width = hidden_states.shape[-2] // self.patch_size, hidden_states.shape[-1] // self.patch_size
hidden_states = self.pos_embed(hidden_states)
if self.adaln_single is not None:
if self.use_additional_conditions and added_cond_kwargs is None:
raise ValueError(
"`added_cond_kwargs` cannot be None when using additional conditions for `adaln_single`."
)
batch_size = hidden_states.shape[0]
timestep, embedded_timestep = self.adaln_single(
timestep, added_cond_kwargs, batch_size=batch_size, hidden_dtype=hidden_states.dtype
)
# 2. Blocks
if self.caption_projection is not None:
batch_size = hidden_states.shape[0]
encoder_hidden_states = self.caption_projection(encoder_hidden_states)
encoder_hidden_states = encoder_hidden_states.view(batch_size, -1, hidden_states.shape[-1])
for block in self.transformer_blocks:
if self.training and self.gradient_checkpointing:
hidden_states, spatial_attn_inputs = torch.utils.checkpoint.checkpoint(
block,
hidden_states,
spatial_attn_inputs,
attention_mask,
encoder_hidden_states,
encoder_attention_mask,
timestep,
cross_attention_kwargs,
class_labels,
use_reentrant=False,
)
else:
hidden_states, spatial_attn_inputs = block(
hidden_states,
spatial_attn_inputs,
attention_mask=attention_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
timestep=timestep,
cross_attention_kwargs=cross_attention_kwargs,
class_labels=class_labels,
)
# 3. Output
if self.is_input_continuous:
if not self.use_linear_projection:
hidden_states = hidden_states.reshape(batch, height, width, inner_dim).permute(0, 3, 1, 2).contiguous()
hidden_states = (
self.proj_out(hidden_states, scale=lora_scale)
if not USE_PEFT_BACKEND
else self.proj_out(hidden_states)
)
else:
hidden_states = (
self.proj_out(hidden_states, scale=lora_scale)
if not USE_PEFT_BACKEND
else self.proj_out(hidden_states)
)
hidden_states = hidden_states.reshape(batch, height, width, inner_dim).permute(0, 3, 1, 2).contiguous()
output = hidden_states + residual
elif self.is_input_vectorized:
hidden_states = self.norm_out(hidden_states)
logits = self.out(hidden_states)
# (batch, self.num_vector_embeds - 1, self.num_latent_pixels)
logits = logits.permute(0, 2, 1)
# log(p(x_0))
output = F.log_softmax(logits.double(), dim=1).float()
if self.is_input_patches:
if self.config.norm_type != "ada_norm_single":
conditioning = self.transformer_blocks[0].norm1.emb(
timestep, class_labels, hidden_dtype=hidden_states.dtype
)
shift, scale = self.proj_out_1(F.silu(conditioning)).chunk(2, dim=1)
hidden_states = self.norm_out(hidden_states) * (1 + scale[:, None]) + shift[:, None]
hidden_states = self.proj_out_2(hidden_states)
elif self.config.norm_type == "ada_norm_single":
shift, scale = (self.scale_shift_table[None] + embedded_timestep[:, None]).chunk(2, dim=1)
hidden_states = self.norm_out(hidden_states)
# Modulation
hidden_states = hidden_states * (1 + scale) + shift
hidden_states = self.proj_out(hidden_states)
hidden_states = hidden_states.squeeze(1)
# unpatchify
if self.adaln_single is None:
height = width = int(hidden_states.shape[1] ** 0.5)
hidden_states = hidden_states.reshape(
shape=(-1, height, width, self.patch_size, self.patch_size, self.out_channels)
)
hidden_states = torch.einsum("nhwpqc->nchpwq", hidden_states)
output = hidden_states.reshape(
shape=(-1, self.out_channels, height * self.patch_size, width * self.patch_size)
)
if not return_dict:
return (output,), spatial_attn_inputs
return Transformer2DModelOutput(sample=output), spatial_attn_inputs
ootd/pipelines_ootd/transformer_vton_2d.py 0 → 100644
View file @ 54a066bf
# Copyright 2023 The HuggingFace Team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# Modified by Yuhao Xu for OOTDiffusion (https://github.com/levihsu/OOTDiffusion)
from dataclasses import dataclass
from typing import Any, Dict, Optional
import torch
import torch.nn.functional as F
from torch import nn
from .attention_vton import BasicTransformerBlock
from diffusers.configuration_utils import ConfigMixin, register_to_config
from diffusers.models.embeddings import ImagePositionalEmbeddings
from diffusers.utils import USE_PEFT_BACKEND, BaseOutput, deprecate
# from diffusers.models.attention import BasicTransformerBlock
from diffusers.models.embeddings import CaptionProjection, PatchEmbed
from diffusers.models.lora import LoRACompatibleConv, LoRACompatibleLinear
from diffusers.models.modeling_utils import ModelMixin
from diffusers.models.normalization import AdaLayerNormSingle
@dataclass
class Transformer2DModelOutput(BaseOutput):
"""
The output of [`Transformer2DModel`].
Args:
sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)` or `(batch size, num_vector_embeds - 1, num_latent_pixels)` if [`Transformer2DModel`] is discrete):
The hidden states output conditioned on the `encoder_hidden_states` input. If discrete, returns probability
distributions for the unnoised latent pixels.
"""
sample: torch.FloatTensor
class Transformer2DModel(ModelMixin, ConfigMixin):
"""
A 2D Transformer model for image-like data.
Parameters:
num_attention_heads (`int`, *optional*, defaults to 16): The number of heads to use for multi-head attention.
attention_head_dim (`int`, *optional*, defaults to 88): The number of channels in each head.
in_channels (`int`, *optional*):
The number of channels in the input and output (specify if the input is **continuous**).
num_layers (`int`, *optional*, defaults to 1): The number of layers of Transformer blocks to use.
dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
cross_attention_dim (`int`, *optional*): The number of `encoder_hidden_states` dimensions to use.
sample_size (`int`, *optional*): The width of the latent images (specify if the input is **discrete**).
This is fixed during training since it is used to learn a number of position embeddings.
num_vector_embeds (`int`, *optional*):
The number of classes of the vector embeddings of the latent pixels (specify if the input is **discrete**).
Includes the class for the masked latent pixel.
activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to use in feed-forward.
num_embeds_ada_norm ( `int`, *optional*):
The number of diffusion steps used during training. Pass if at least one of the norm_layers is
`AdaLayerNorm`. This is fixed during training since it is used to learn a number of embeddings that are
added to the hidden states.
During inference, you can denoise for up to but not more steps than `num_embeds_ada_norm`.
attention_bias (`bool`, *optional*):
Configure if the `TransformerBlocks` attention should contain a bias parameter.
"""
@register_to_config
def __init__(
self,
num_attention_heads: int = 16,
attention_head_dim: int = 88,
in_channels: Optional[int] = None,
out_channels: Optional[int] = None,
num_layers: int = 1,
dropout: float = 0.0,
norm_num_groups: int = 32,
cross_attention_dim: Optional[int] = None,
attention_bias: bool = False,
sample_size: Optional[int] = None,
num_vector_embeds: Optional[int] = None,
patch_size: Optional[int] = None,
activation_fn: str = "geglu",
num_embeds_ada_norm: Optional[int] = None,
use_linear_projection: bool = False,
only_cross_attention: bool = False,
double_self_attention: bool = False,
upcast_attention: bool = False,
norm_type: str = "layer_norm",
norm_elementwise_affine: bool = True,
norm_eps: float = 1e-5,
attention_type: str = "default",
caption_channels: int = None,
):
super().__init__()
self.use_linear_projection = use_linear_projection
self.num_attention_heads = num_attention_heads
self.attention_head_dim = attention_head_dim
inner_dim = num_attention_heads * attention_head_dim
conv_cls = nn.Conv2d if USE_PEFT_BACKEND else LoRACompatibleConv
linear_cls = nn.Linear if USE_PEFT_BACKEND else LoRACompatibleLinear
# 1. Transformer2DModel can process both standard continuous images of shape `(batch_size, num_channels, width, height)` as well as quantized image embeddings of shape `(batch_size, num_image_vectors)`
# Define whether input is continuous or discrete depending on configuration
self.is_input_continuous = (in_channels is not None) and (patch_size is None)
self.is_input_vectorized = num_vector_embeds is not None
self.is_input_patches = in_channels is not None and patch_size is not None
if norm_type == "layer_norm" and num_embeds_ada_norm is not None:
deprecation_message = (
f"The configuration file of this model: {self.__class__} is outdated. `norm_type` is either not set or"
" incorrectly set to `'layer_norm'`.Make sure to set `norm_type` to `'ada_norm'` in the config."
" Please make sure to update the config accordingly as leaving `norm_type` might led to incorrect"
" results in future versions. If you have downloaded this checkpoint from the Hugging Face Hub, it"
" would be very nice if you could open a Pull request for the `transformer/config.json` file"
)
deprecate("norm_type!=num_embeds_ada_norm", "1.0.0", deprecation_message, standard_warn=False)
norm_type = "ada_norm"
if self.is_input_continuous and self.is_input_vectorized:
raise ValueError(
f"Cannot define both `in_channels`: {in_channels} and `num_vector_embeds`: {num_vector_embeds}. Make"
" sure that either `in_channels` or `num_vector_embeds` is None."
)
elif self.is_input_vectorized and self.is_input_patches:
raise ValueError(
f"Cannot define both `num_vector_embeds`: {num_vector_embeds} and `patch_size`: {patch_size}. Make"
" sure that either `num_vector_embeds` or `num_patches` is None."
)
elif not self.is_input_continuous and not self.is_input_vectorized and not self.is_input_patches:
raise ValueError(
f"Has to define `in_channels`: {in_channels}, `num_vector_embeds`: {num_vector_embeds}, or patch_size:"
f" {patch_size}. Make sure that `in_channels`, `num_vector_embeds` or `num_patches` is not None."
)
# 2. Define input layers
if self.is_input_continuous:
self.in_channels = in_channels
self.norm = torch.nn.GroupNorm(num_groups=norm_num_groups, num_channels=in_channels, eps=1e-6, affine=True)
if use_linear_projection:
self.proj_in = linear_cls(in_channels, inner_dim)
else:
self.proj_in = conv_cls(in_channels, inner_dim, kernel_size=1, stride=1, padding=0)
elif self.is_input_vectorized:
assert sample_size is not None, "Transformer2DModel over discrete input must provide sample_size"
assert num_vector_embeds is not None, "Transformer2DModel over discrete input must provide num_embed"
self.height = sample_size
self.width = sample_size
self.num_vector_embeds = num_vector_embeds
self.num_latent_pixels = self.height * self.width
self.latent_image_embedding = ImagePositionalEmbeddings(
num_embed=num_vector_embeds, embed_dim=inner_dim, height=self.height, width=self.width
)
elif self.is_input_patches:
assert sample_size is not None, "Transformer2DModel over patched input must provide sample_size"
self.height = sample_size
self.width = sample_size
self.patch_size = patch_size
interpolation_scale = self.config.sample_size // 64 # => 64 (= 512 pixart) has interpolation scale 1
interpolation_scale = max(interpolation_scale, 1)
self.pos_embed = PatchEmbed(
height=sample_size,
width=sample_size,
patch_size=patch_size,
in_channels=in_channels,
embed_dim=inner_dim,
interpolation_scale=interpolation_scale,
)
# 3. Define transformers blocks
self.transformer_blocks = nn.ModuleList(
[
BasicTransformerBlock(
inner_dim,
num_attention_heads,
attention_head_dim,
dropout=dropout,
cross_attention_dim=cross_attention_dim,
activation_fn=activation_fn,
num_embeds_ada_norm=num_embeds_ada_norm,
attention_bias=attention_bias,
only_cross_attention=only_cross_attention,
double_self_attention=double_self_attention,
upcast_attention=upcast_attention,
norm_type=norm_type,
norm_elementwise_affine=norm_elementwise_affine,
norm_eps=norm_eps,
attention_type=attention_type,
)
for d in range(num_layers)
]
)
# 4. Define output layers
self.out_channels = in_channels if out_channels is None else out_channels
if self.is_input_continuous:
# TODO: should use out_channels for continuous projections
if use_linear_projection:
self.proj_out = linear_cls(inner_dim, in_channels)
else:
self.proj_out = conv_cls(inner_dim, in_channels, kernel_size=1, stride=1, padding=0)
elif self.is_input_vectorized:
self.norm_out = nn.LayerNorm(inner_dim)
self.out = nn.Linear(inner_dim, self.num_vector_embeds - 1)
elif self.is_input_patches and norm_type != "ada_norm_single":
self.norm_out = nn.LayerNorm(inner_dim, elementwise_affine=False, eps=1e-6)
self.proj_out_1 = nn.Linear(inner_dim, 2 * inner_dim)
self.proj_out_2 = nn.Linear(inner_dim, patch_size * patch_size * self.out_channels)
elif self.is_input_patches and norm_type == "ada_norm_single":
self.norm_out = nn.LayerNorm(inner_dim, elementwise_affine=False, eps=1e-6)
self.scale_shift_table = nn.Parameter(torch.randn(2, inner_dim) / inner_dim**0.5)
self.proj_out = nn.Linear(inner_dim, patch_size * patch_size * self.out_channels)
# 5. PixArt-Alpha blocks.
self.adaln_single = None
self.use_additional_conditions = False
if norm_type == "ada_norm_single":
self.use_additional_conditions = self.config.sample_size == 128
# TODO(Sayak, PVP) clean this, for now we use sample size to determine whether to use
# additional conditions until we find better name
self.adaln_single = AdaLayerNormSingle(inner_dim, use_additional_conditions=self.use_additional_conditions)
self.caption_projection = None
if caption_channels is not None:
self.caption_projection = CaptionProjection(in_features=caption_channels, hidden_size=inner_dim)
self.gradient_checkpointing = False
def forward(
self,
hidden_states: torch.Tensor,
spatial_attn_inputs = [],
spatial_attn_idx = 0,
encoder_hidden_states: Optional[torch.Tensor] = None,
timestep: Optional[torch.LongTensor] = None,
added_cond_kwargs: Dict[str, torch.Tensor] = None,
class_labels: Optional[torch.LongTensor] = None,
cross_attention_kwargs: Dict[str, Any] = None,
attention_mask: Optional[torch.Tensor] = None,
encoder_attention_mask: Optional[torch.Tensor] = None,
return_dict: bool = True,
):
"""
The [`Transformer2DModel`] forward method.
Args:
hidden_states (`torch.LongTensor` of shape `(batch size, num latent pixels)` if discrete, `torch.FloatTensor` of shape `(batch size, channel, height, width)` if continuous):
Input `hidden_states`.
encoder_hidden_states ( `torch.FloatTensor` of shape `(batch size, sequence len, embed dims)`, *optional*):
Conditional embeddings for cross attention layer. If not given, cross-attention defaults to
self-attention.
timestep ( `torch.LongTensor`, *optional*):
Used to indicate denoising step. Optional timestep to be applied as an embedding in `AdaLayerNorm`.
class_labels ( `torch.LongTensor` of shape `(batch size, num classes)`, *optional*):
Used to indicate class labels conditioning. Optional class labels to be applied as an embedding in
`AdaLayerZeroNorm`.
cross_attention_kwargs ( `Dict[str, Any]`, *optional*):
A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
`self.processor` in
[diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
attention_mask ( `torch.Tensor`, *optional*):
An attention mask of shape `(batch, key_tokens)` is applied to `encoder_hidden_states`. If `1` the mask
is kept, otherwise if `0` it is discarded. Mask will be converted into a bias, which adds large
negative values to the attention scores corresponding to "discard" tokens.
encoder_attention_mask ( `torch.Tensor`, *optional*):
Cross-attention mask applied to `encoder_hidden_states`. Two formats supported:
* Mask `(batch, sequence_length)` True = keep, False = discard.
* Bias `(batch, 1, sequence_length)` 0 = keep, -10000 = discard.
If `ndim == 2`: will be interpreted as a mask, then converted into a bias consistent with the format
above. This bias will be added to the cross-attention scores.
return_dict (`bool`, *optional*, defaults to `True`):
Whether or not to return a [`~models.unet_2d_condition.UNet2DConditionOutput`] instead of a plain
tuple.
Returns:
If `return_dict` is True, an [`~models.transformer_2d.Transformer2DModelOutput`] is returned, otherwise a
`tuple` where the first element is the sample tensor.
"""
# ensure attention_mask is a bias, and give it a singleton query_tokens dimension.
# we may have done this conversion already, e.g. if we came here via UNet2DConditionModel#forward.
# we can tell by counting dims; if ndim == 2: it's a mask rather than a bias.
# expects mask of shape:
# [batch, key_tokens]
# adds singleton query_tokens dimension:
# [batch, 1, key_tokens]
# this helps to broadcast it as a bias over attention scores, which will be in one of the following shapes:
# [batch, heads, query_tokens, key_tokens] (e.g. torch sdp attn)
# [batch * heads, query_tokens, key_tokens] (e.g. xformers or classic attn)
if attention_mask is not None and attention_mask.ndim == 2:
# assume that mask is expressed as:
# (1 = keep, 0 = discard)
# convert mask into a bias that can be added to attention scores:
# (keep = +0, discard = -10000.0)
attention_mask = (1 - attention_mask.to(hidden_states.dtype)) * -10000.0
attention_mask = attention_mask.unsqueeze(1)
# convert encoder_attention_mask to a bias the same way we do for attention_mask
if encoder_attention_mask is not None and encoder_attention_mask.ndim == 2:
encoder_attention_mask = (1 - encoder_attention_mask.to(hidden_states.dtype)) * -10000.0
encoder_attention_mask = encoder_attention_mask.unsqueeze(1)
# Retrieve lora scale.
lora_scale = cross_attention_kwargs.get("scale", 1.0) if cross_attention_kwargs is not None else 1.0
# 1. Input
if self.is_input_continuous:
batch, _, height, width = hidden_states.shape
residual = hidden_states
hidden_states = self.norm(hidden_states)
if not self.use_linear_projection:
hidden_states = (
self.proj_in(hidden_states, scale=lora_scale)
if not USE_PEFT_BACKEND
else self.proj_in(hidden_states)
)
inner_dim = hidden_states.shape[1]
hidden_states = hidden_states.permute(0, 2, 3, 1).reshape(batch, height * width, inner_dim)
else:
inner_dim = hidden_states.shape[1]
hidden_states = hidden_states.permute(0, 2, 3, 1).reshape(batch, height * width, inner_dim)
hidden_states = (
self.proj_in(hidden_states, scale=lora_scale)
if not USE_PEFT_BACKEND
else self.proj_in(hidden_states)
)
elif self.is_input_vectorized:
hidden_states = self.latent_image_embedding(hidden_states)
elif self.is_input_patches:
height, width = hidden_states.shape[-2] // self.patch_size, hidden_states.shape[-1] // self.patch_size
hidden_states = self.pos_embed(hidden_states)
if self.adaln_single is not None:
if self.use_additional_conditions and added_cond_kwargs is None:
raise ValueError(
"`added_cond_kwargs` cannot be None when using additional conditions for `adaln_single`."
)
batch_size = hidden_states.shape[0]
timestep, embedded_timestep = self.adaln_single(
timestep, added_cond_kwargs, batch_size=batch_size, hidden_dtype=hidden_states.dtype
)
# 2. Blocks
if self.caption_projection is not None:
batch_size = hidden_states.shape[0]
encoder_hidden_states = self.caption_projection(encoder_hidden_states)
encoder_hidden_states = encoder_hidden_states.view(batch_size, -1, hidden_states.shape[-1])
for block in self.transformer_blocks:
if self.training and self.gradient_checkpointing:
hidden_states, spatial_attn_inputs, spatial_attn_idx = torch.utils.checkpoint.checkpoint(
block,
hidden_states,
spatial_attn_inputs,
spatial_attn_idx,
attention_mask,
encoder_hidden_states,
encoder_attention_mask,
timestep,
cross_attention_kwargs,
class_labels,
use_reentrant=False,
)
else:
hidden_states, spatial_attn_inputs, spatial_attn_idx = block(
hidden_states,
spatial_attn_inputs,
spatial_attn_idx,
attention_mask=attention_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
timestep=timestep,
cross_attention_kwargs=cross_attention_kwargs,
class_labels=class_labels,
)
# 3. Output
if self.is_input_continuous:
if not self.use_linear_projection:
hidden_states = hidden_states.reshape(batch, height, width, inner_dim).permute(0, 3, 1, 2).contiguous()
hidden_states = (
self.proj_out(hidden_states, scale=lora_scale)
if not USE_PEFT_BACKEND
else self.proj_out(hidden_states)
)
else:
hidden_states = (
self.proj_out(hidden_states, scale=lora_scale)
if not USE_PEFT_BACKEND
else self.proj_out(hidden_states)
)
hidden_states = hidden_states.reshape(batch, height, width, inner_dim).permute(0, 3, 1, 2).contiguous()
output = hidden_states + residual
elif self.is_input_vectorized:
hidden_states = self.norm_out(hidden_states)
logits = self.out(hidden_states)
# (batch, self.num_vector_embeds - 1, self.num_latent_pixels)
logits = logits.permute(0, 2, 1)
# log(p(x_0))
output = F.log_softmax(logits.double(), dim=1).float()
if self.is_input_patches:
if self.config.norm_type != "ada_norm_single":
conditioning = self.transformer_blocks[0].norm1.emb(
timestep, class_labels, hidden_dtype=hidden_states.dtype
)
shift, scale = self.proj_out_1(F.silu(conditioning)).chunk(2, dim=1)
hidden_states = self.norm_out(hidden_states) * (1 + scale[:, None]) + shift[:, None]
hidden_states = self.proj_out_2(hidden_states)
elif self.config.norm_type == "ada_norm_single":
shift, scale = (self.scale_shift_table[None] + embedded_timestep[:, None]).chunk(2, dim=1)
hidden_states = self.norm_out(hidden_states)
# Modulation
hidden_states = hidden_states * (1 + scale) + shift
hidden_states = self.proj_out(hidden_states)
hidden_states = hidden_states.squeeze(1)
# unpatchify
if self.adaln_single is None:
height = width = int(hidden_states.shape[1] ** 0.5)
hidden_states = hidden_states.reshape(
shape=(-1, height, width, self.patch_size, self.patch_size, self.out_channels)
)
hidden_states = torch.einsum("nhwpqc->nchpwq", hidden_states)
output = hidden_states.reshape(
shape=(-1, self.out_channels, height * self.patch_size, width * self.patch_size)
)
if not return_dict:
return (output,), spatial_attn_inputs, spatial_attn_idx
return Transformer2DModelOutput(sample=output), spatial_attn_inputs, spatial_attn_idx
ootd/pipelines_ootd/unet_garm_2d_blocks.py 0 → 100644
View file @ 54a066bf
# Copyright 2023 The HuggingFace Team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# Modified by Yuhao Xu for OOTDiffusion (https://github.com/levihsu/OOTDiffusion)
from typing import Any, Dict, Optional, Tuple, Union
import numpy as np
import torch
import torch.nn.functional as F
from torch import nn
from .transformer_garm_2d import Transformer2DModel
from diffusers.utils import is_torch_version, logging
from diffusers.utils.torch_utils import apply_freeu
from diffusers.models.activations import get_activation
from diffusers.models.attention_processor import Attention, AttnAddedKVProcessor, AttnAddedKVProcessor2_0
from diffusers.models.dual_transformer_2d import DualTransformer2DModel
from diffusers.models.normalization import AdaGroupNorm
from diffusers.models.resnet import Downsample2D, FirDownsample2D, FirUpsample2D, KDownsample2D, KUpsample2D, ResnetBlock2D, Upsample2D
# from diffusers.models.transformer_2d import Transformer2DModel
logger = logging.get_logger(__name__) # pylint: disable=invalid-name
def get_down_block(
down_block_type: str,
num_layers: int,
in_channels: int,
out_channels: int,
temb_channels: int,
add_downsample: bool,
resnet_eps: float,
resnet_act_fn: str,
transformer_layers_per_block: int = 1,
num_attention_heads: Optional[int] = None,
resnet_groups: Optional[int] = None,
cross_attention_dim: Optional[int] = None,
downsample_padding: Optional[int] = None,
dual_cross_attention: bool = False,
use_linear_projection: bool = False,
only_cross_attention: bool = False,
upcast_attention: bool = False,
resnet_time_scale_shift: str = "default",
attention_type: str = "default",
resnet_skip_time_act: bool = False,
resnet_out_scale_factor: float = 1.0,
cross_attention_norm: Optional[str] = None,
attention_head_dim: Optional[int] = None,
downsample_type: Optional[str] = None,
dropout: float = 0.0,
):
# If attn head dim is not defined, we default it to the number of heads
if attention_head_dim is None:
logger.warn(
f"It is recommended to provide `attention_head_dim` when calling `get_down_block`. Defaulting `attention_head_dim` to {num_attention_heads}."
)
attention_head_dim = num_attention_heads
down_block_type = down_block_type[7:] if down_block_type.startswith("UNetRes") else down_block_type
if down_block_type == "DownBlock2D":
return DownBlock2D(
num_layers=num_layers,
in_channels=in_channels,
out_channels=out_channels,
temb_channels=temb_channels,
dropout=dropout,
add_downsample=add_downsample,
resnet_eps=resnet_eps,
resnet_act_fn=resnet_act_fn,
resnet_groups=resnet_groups,
downsample_padding=downsample_padding,
resnet_time_scale_shift=resnet_time_scale_shift,
)
elif down_block_type == "ResnetDownsampleBlock2D":
return ResnetDownsampleBlock2D(
num_layers=num_layers,
in_channels=in_channels,
out_channels=out_channels,
temb_channels=temb_channels,
dropout=dropout,
add_downsample=add_downsample,
resnet_eps=resnet_eps,
resnet_act_fn=resnet_act_fn,
resnet_groups=resnet_groups,
resnet_time_scale_shift=resnet_time_scale_shift,
skip_time_act=resnet_skip_time_act,
output_scale_factor=resnet_out_scale_factor,
)
elif down_block_type == "AttnDownBlock2D":
if add_downsample is False:
downsample_type = None
else:
downsample_type = downsample_type or "conv" # default to 'conv'
return AttnDownBlock2D(
num_layers=num_layers,
in_channels=in_channels,
out_channels=out_channels,
temb_channels=temb_channels,
dropout=dropout,
resnet_eps=resnet_eps,
resnet_act_fn=resnet_act_fn,
resnet_groups=resnet_groups,
downsample_padding=downsample_padding,
attention_head_dim=attention_head_dim,
resnet_time_scale_shift=resnet_time_scale_shift,
downsample_type=downsample_type,
)
elif down_block_type == "CrossAttnDownBlock2D":
if cross_attention_dim is None:
raise ValueError("cross_attention_dim must be specified for CrossAttnDownBlock2D")
return CrossAttnDownBlock2D(
num_layers=num_layers,
transformer_layers_per_block=transformer_layers_per_block,
in_channels=in_channels,
out_channels=out_channels,
temb_channels=temb_channels,
dropout=dropout,
add_downsample=add_downsample,
resnet_eps=resnet_eps,
resnet_act_fn=resnet_act_fn,
resnet_groups=resnet_groups,
downsample_padding=downsample_padding,
cross_attention_dim=cross_attention_dim,
num_attention_heads=num_attention_heads,
dual_cross_attention=dual_cross_attention,
use_linear_projection=use_linear_projection,
only_cross_attention=only_cross_attention,
upcast_attention=upcast_attention,
resnet_time_scale_shift=resnet_time_scale_shift,
attention_type=attention_type,
)
elif down_block_type == "SimpleCrossAttnDownBlock2D":
if cross_attention_dim is None:
raise ValueError("cross_attention_dim must be specified for SimpleCrossAttnDownBlock2D")
return SimpleCrossAttnDownBlock2D(
num_layers=num_layers,
in_channels=in_channels,
out_channels=out_channels,
temb_channels=temb_channels,
dropout=dropout,
add_downsample=add_downsample,
resnet_eps=resnet_eps,
resnet_act_fn=resnet_act_fn,
resnet_groups=resnet_groups,
cross_attention_dim=cross_attention_dim,
attention_head_dim=attention_head_dim,
resnet_time_scale_shift=resnet_time_scale_shift,
skip_time_act=resnet_skip_time_act,
output_scale_factor=resnet_out_scale_factor,
only_cross_attention=only_cross_attention,
cross_attention_norm=cross_attention_norm,
)
elif down_block_type == "SkipDownBlock2D":
return SkipDownBlock2D(
num_layers=num_layers,
in_channels=in_channels,
out_channels=out_channels,
temb_channels=temb_channels,
dropout=dropout,
add_downsample=add_downsample,
resnet_eps=resnet_eps,
resnet_act_fn=resnet_act_fn,
downsample_padding=downsample_padding,
resnet_time_scale_shift=resnet_time_scale_shift,
)
elif down_block_type == "AttnSkipDownBlock2D":
return AttnSkipDownBlock2D(
num_layers=num_layers,
in_channels=in_channels,
out_channels=out_channels,
temb_channels=temb_channels,
dropout=dropout,
add_downsample=add_downsample,
resnet_eps=resnet_eps,
resnet_act_fn=resnet_act_fn,
attention_head_dim=attention_head_dim,
resnet_time_scale_shift=resnet_time_scale_shift,
)
elif down_block_type == "DownEncoderBlock2D":
return DownEncoderBlock2D(
num_layers=num_layers,
in_channels=in_channels,
out_channels=out_channels,
dropout=dropout,
add_downsample=add_downsample,
resnet_eps=resnet_eps,
resnet_act_fn=resnet_act_fn,
resnet_groups=resnet_groups,
downsample_padding=downsample_padding,
resnet_time_scale_shift=resnet_time_scale_shift,
)
elif down_block_type == "AttnDownEncoderBlock2D":
return AttnDownEncoderBlock2D(
num_layers=num_layers,
in_channels=in_channels,
out_channels=out_channels,
dropout=dropout,
add_downsample=add_downsample,
resnet_eps=resnet_eps,
resnet_act_fn=resnet_act_fn,
resnet_groups=resnet_groups,
downsample_padding=downsample_padding,
attention_head_dim=attention_head_dim,
resnet_time_scale_shift=resnet_time_scale_shift,
)
elif down_block_type == "KDownBlock2D":
return KDownBlock2D(
num_layers=num_layers,
in_channels=in_channels,
out_channels=out_channels,
temb_channels=temb_channels,
dropout=dropout,
add_downsample=add_downsample,
resnet_eps=resnet_eps,
resnet_act_fn=resnet_act_fn,
)
elif down_block_type == "KCrossAttnDownBlock2D":
return KCrossAttnDownBlock2D(
num_layers=num_layers,
in_channels=in_channels,
out_channels=out_channels,
temb_channels=temb_channels,
dropout=dropout,
add_downsample=add_downsample,
resnet_eps=resnet_eps,
resnet_act_fn=resnet_act_fn,
cross_attention_dim=cross_attention_dim,
attention_head_dim=attention_head_dim,
add_self_attention=True if not add_downsample else False,
)
raise ValueError(f"{down_block_type} does not exist.")
def get_up_block(
up_block_type: str,
num_layers: int,
in_channels: int,
out_channels: int,
prev_output_channel: int,
temb_channels: int,
add_upsample: bool,
resnet_eps: float,
resnet_act_fn: str,
resolution_idx: Optional[int] = None,
transformer_layers_per_block: int = 1,
num_attention_heads: Optional[int] = None,
resnet_groups: Optional[int] = None,
cross_attention_dim: Optional[int] = None,
dual_cross_attention: bool = False,
use_linear_projection: bool = False,
only_cross_attention: bool = False,
upcast_attention: bool = False,
resnet_time_scale_shift: str = "default",
attention_type: str = "default",
resnet_skip_time_act: bool = False,
resnet_out_scale_factor: float = 1.0,
cross_attention_norm: Optional[str] = None,
attention_head_dim: Optional[int] = None,
upsample_type: Optional[str] = None,
dropout: float = 0.0,
) -> nn.Module:
# If attn head dim is not defined, we default it to the number of heads
if attention_head_dim is None:
logger.warn(
f"It is recommended to provide `attention_head_dim` when calling `get_up_block`. Defaulting `attention_head_dim` to {num_attention_heads}."
)
attention_head_dim = num_attention_heads
up_block_type = up_block_type[7:] if up_block_type.startswith("UNetRes") else up_block_type
if up_block_type == "UpBlock2D":
return UpBlock2D(
num_layers=num_layers,
in_channels=in_channels,
out_channels=out_channels,
prev_output_channel=prev_output_channel,
temb_channels=temb_channels,
resolution_idx=resolution_idx,
dropout=dropout,
add_upsample=add_upsample,
resnet_eps=resnet_eps,
resnet_act_fn=resnet_act_fn,
resnet_groups=resnet_groups,
resnet_time_scale_shift=resnet_time_scale_shift,
)
elif up_block_type == "ResnetUpsampleBlock2D":
return ResnetUpsampleBlock2D(
num_layers=num_layers,
in_channels=in_channels,
out_channels=out_channels,
prev_output_channel=prev_output_channel,
temb_channels=temb_channels,
resolution_idx=resolution_idx,
dropout=dropout,
add_upsample=add_upsample,
resnet_eps=resnet_eps,
resnet_act_fn=resnet_act_fn,
resnet_groups=resnet_groups,
resnet_time_scale_shift=resnet_time_scale_shift,
skip_time_act=resnet_skip_time_act,
output_scale_factor=resnet_out_scale_factor,
)
elif up_block_type == "CrossAttnUpBlock2D":
if cross_attention_dim is None:
raise ValueError("cross_attention_dim must be specified for CrossAttnUpBlock2D")
return CrossAttnUpBlock2D(
num_layers=num_layers,
transformer_layers_per_block=transformer_layers_per_block,
in_channels=in_channels,
out_channels=out_channels,
prev_output_channel=prev_output_channel,
temb_channels=temb_channels,
resolution_idx=resolution_idx,
dropout=dropout,
add_upsample=add_upsample,
resnet_eps=resnet_eps,
resnet_act_fn=resnet_act_fn,
resnet_groups=resnet_groups,
cross_attention_dim=cross_attention_dim,
num_attention_heads=num_attention_heads,
dual_cross_attention=dual_cross_attention,
use_linear_projection=use_linear_projection,
only_cross_attention=only_cross_attention,
upcast_attention=upcast_attention,
resnet_time_scale_shift=resnet_time_scale_shift,
attention_type=attention_type,
)
elif up_block_type == "SimpleCrossAttnUpBlock2D":
if cross_attention_dim is None:
raise ValueError("cross_attention_dim must be specified for SimpleCrossAttnUpBlock2D")
return SimpleCrossAttnUpBlock2D(
num_layers=num_layers,
in_channels=in_channels,
out_channels=out_channels,
prev_output_channel=prev_output_channel,
temb_channels=temb_channels,
resolution_idx=resolution_idx,
dropout=dropout,
add_upsample=add_upsample,
resnet_eps=resnet_eps,
resnet_act_fn=resnet_act_fn,
resnet_groups=resnet_groups,
cross_attention_dim=cross_attention_dim,
attention_head_dim=attention_head_dim,
resnet_time_scale_shift=resnet_time_scale_shift,
skip_time_act=resnet_skip_time_act,
output_scale_factor=resnet_out_scale_factor,
only_cross_attention=only_cross_attention,
cross_attention_norm=cross_attention_norm,
)
elif up_block_type == "AttnUpBlock2D":
if add_upsample is False:
upsample_type = None
else:
upsample_type = upsample_type or "conv" # default to 'conv'
return AttnUpBlock2D(
num_layers=num_layers,
in_channels=in_channels,
out_channels=out_channels,
prev_output_channel=prev_output_channel,
temb_channels=temb_channels,
resolution_idx=resolution_idx,
dropout=dropout,
resnet_eps=resnet_eps,
resnet_act_fn=resnet_act_fn,
resnet_groups=resnet_groups,
attention_head_dim=attention_head_dim,
resnet_time_scale_shift=resnet_time_scale_shift,
upsample_type=upsample_type,
)
elif up_block_type == "SkipUpBlock2D":
return SkipUpBlock2D(
num_layers=num_layers,
in_channels=in_channels,
out_channels=out_channels,
prev_output_channel=prev_output_channel,
temb_channels=temb_channels,
resolution_idx=resolution_idx,
dropout=dropout,
add_upsample=add_upsample,
resnet_eps=resnet_eps,
resnet_act_fn=resnet_act_fn,
resnet_time_scale_shift=resnet_time_scale_shift,
)
elif up_block_type == "AttnSkipUpBlock2D":
return AttnSkipUpBlock2D(
num_layers=num_layers,
in_channels=in_channels,
out_channels=out_channels,
prev_output_channel=prev_output_channel,
temb_channels=temb_channels,
resolution_idx=resolution_idx,
dropout=dropout,
add_upsample=add_upsample,
resnet_eps=resnet_eps,
resnet_act_fn=resnet_act_fn,
attention_head_dim=attention_head_dim,
resnet_time_scale_shift=resnet_time_scale_shift,
)
elif up_block_type == "UpDecoderBlock2D":
return UpDecoderBlock2D(
num_layers=num_layers,
in_channels=in_channels,
out_channels=out_channels,
resolution_idx=resolution_idx,
dropout=dropout,
add_upsample=add_upsample,
resnet_eps=resnet_eps,
resnet_act_fn=resnet_act_fn,
resnet_groups=resnet_groups,
resnet_time_scale_shift=resnet_time_scale_shift,
temb_channels=temb_channels,
)
elif up_block_type == "AttnUpDecoderBlock2D":
return AttnUpDecoderBlock2D(
num_layers=num_layers,
in_channels=in_channels,
out_channels=out_channels,
resolution_idx=resolution_idx,
dropout=dropout,
add_upsample=add_upsample,
resnet_eps=resnet_eps,
resnet_act_fn=resnet_act_fn,
resnet_groups=resnet_groups,
attention_head_dim=attention_head_dim,
resnet_time_scale_shift=resnet_time_scale_shift,
temb_channels=temb_channels,
)
elif up_block_type == "KUpBlock2D":
return KUpBlock2D(
num_layers=num_layers,
in_channels=in_channels,
out_channels=out_channels,
temb_channels=temb_channels,
resolution_idx=resolution_idx,
dropout=dropout,
add_upsample=add_upsample,
resnet_eps=resnet_eps,
resnet_act_fn=resnet_act_fn,
)
elif up_block_type == "KCrossAttnUpBlock2D":
return KCrossAttnUpBlock2D(
num_layers=num_layers,
in_channels=in_channels,
out_channels=out_channels,
temb_channels=temb_channels,
resolution_idx=resolution_idx,
dropout=dropout,
add_upsample=add_upsample,
resnet_eps=resnet_eps,
resnet_act_fn=resnet_act_fn,
cross_attention_dim=cross_attention_dim,
attention_head_dim=attention_head_dim,
)
raise ValueError(f"{up_block_type} does not exist.")
class AutoencoderTinyBlock(nn.Module):
"""
Tiny Autoencoder block used in [`AutoencoderTiny`]. It is a mini residual module consisting of plain conv + ReLU
blocks.
Args:
in_channels (`int`): The number of input channels.
out_channels (`int`): The number of output channels.
act_fn (`str`):
` The activation function to use. Supported values are `"swish"`, `"mish"`, `"gelu"`, and `"relu"`.
Returns:
`torch.FloatTensor`: A tensor with the same shape as the input tensor, but with the number of channels equal to
`out_channels`.
"""
def __init__(self, in_channels: int, out_channels: int, act_fn: str):
super().__init__()
act_fn = get_activation(act_fn)
self.conv = nn.Sequential(
nn.Conv2d(in_channels, out_channels, kernel_size=3, padding=1),
act_fn,
nn.Conv2d(out_channels, out_channels, kernel_size=3, padding=1),
act_fn,
nn.Conv2d(out_channels, out_channels, kernel_size=3, padding=1),
)
self.skip = (
nn.Conv2d(in_channels, out_channels, kernel_size=1, bias=False)
if in_channels != out_channels
else nn.Identity()
)
self.fuse = nn.ReLU()
def forward(self, x: torch.FloatTensor) -> torch.FloatTensor:
return self.fuse(self.conv(x) + self.skip(x))
class UNetMidBlock2D(nn.Module):
"""
A 2D UNet mid-block [`UNetMidBlock2D`] with multiple residual blocks and optional attention blocks.
Args:
in_channels (`int`): The number of input channels.
temb_channels (`int`): The number of temporal embedding channels.
dropout (`float`, *optional*, defaults to 0.0): The dropout rate.
num_layers (`int`, *optional*, defaults to 1): The number of residual blocks.
resnet_eps (`float`, *optional*, 1e-6 ): The epsilon value for the resnet blocks.
resnet_time_scale_shift (`str`, *optional*, defaults to `default`):
The type of normalization to apply to the time embeddings. This can help to improve the performance of the
model on tasks with long-range temporal dependencies.
resnet_act_fn (`str`, *optional*, defaults to `swish`): The activation function for the resnet blocks.
resnet_groups (`int`, *optional*, defaults to 32):
The number of groups to use in the group normalization layers of the resnet blocks.
attn_groups (`Optional[int]`, *optional*, defaults to None): The number of groups for the attention blocks.
resnet_pre_norm (`bool`, *optional*, defaults to `True`):
Whether to use pre-normalization for the resnet blocks.
add_attention (`bool`, *optional*, defaults to `True`): Whether to add attention blocks.
attention_head_dim (`int`, *optional*, defaults to 1):
Dimension of a single attention head. The number of attention heads is determined based on this value and
the number of input channels.
output_scale_factor (`float`, *optional*, defaults to 1.0): The output scale factor.
Returns:
`torch.FloatTensor`: The output of the last residual block, which is a tensor of shape `(batch_size,
in_channels, height, width)`.
"""
def __init__(
self,
in_channels: int,
temb_channels: int,
dropout: float = 0.0,
num_layers: int = 1,
resnet_eps: float = 1e-6,
resnet_time_scale_shift: str = "default", # default, spatial
resnet_act_fn: str = "swish",
resnet_groups: int = 32,
attn_groups: Optional[int] = None,
resnet_pre_norm: bool = True,
add_attention: bool = True,
attention_head_dim: int = 1,
output_scale_factor: float = 1.0,
):
super().__init__()
resnet_groups = resnet_groups if resnet_groups is not None else min(in_channels // 4, 32)
self.add_attention = add_attention
if attn_groups is None:
attn_groups = resnet_groups if resnet_time_scale_shift == "default" else None
# there is always at least one resnet
resnets = [
ResnetBlock2D(
in_channels=in_channels,
out_channels=in_channels,
temb_channels=temb_channels,
eps=resnet_eps,
groups=resnet_groups,
dropout=dropout,
time_embedding_norm=resnet_time_scale_shift,
non_linearity=resnet_act_fn,
output_scale_factor=output_scale_factor,
pre_norm=resnet_pre_norm,
)
]
attentions = []
if attention_head_dim is None:
logger.warn(
f"It is not recommend to pass `attention_head_dim=None`. Defaulting `attention_head_dim` to `in_channels`: {in_channels}."
)
attention_head_dim = in_channels
for _ in range(num_layers):
if self.add_attention:
attentions.append(
Attention(
in_channels,
heads=in_channels // attention_head_dim,
dim_head=attention_head_dim,
rescale_output_factor=output_scale_factor,
eps=resnet_eps,
norm_num_groups=attn_groups,
spatial_norm_dim=temb_channels if resnet_time_scale_shift == "spatial" else None,
residual_connection=True,
bias=True,
upcast_softmax=True,
_from_deprecated_attn_block=True,
)
)
else:
attentions.append(None)
resnets.append(
ResnetBlock2D(
in_channels=in_channels,
out_channels=in_channels,
temb_channels=temb_channels,
eps=resnet_eps,
groups=resnet_groups,
dropout=dropout,
time_embedding_norm=resnet_time_scale_shift,
non_linearity=resnet_act_fn,
output_scale_factor=output_scale_factor,
pre_norm=resnet_pre_norm,
)
)
self.attentions = nn.ModuleList(attentions)
self.resnets = nn.ModuleList(resnets)
def forward(self, hidden_states: torch.FloatTensor, temb: Optional[torch.FloatTensor] = None) -> torch.FloatTensor:
hidden_states = self.resnets[0](hidden_states, temb)
for attn, resnet in zip(self.attentions, self.resnets[1:]):
if attn is not None:
hidden_states = attn(hidden_states, temb=temb)
hidden_states = resnet(hidden_states, temb)
return hidden_states
class UNetMidBlock2DCrossAttn(nn.Module):
def __init__(
self,
in_channels: int,
temb_channels: int,
dropout: float = 0.0,
num_layers: int = 1,
transformer_layers_per_block: Union[int, Tuple[int]] = 1,
resnet_eps: float = 1e-6,
resnet_time_scale_shift: str = "default",
resnet_act_fn: str = "swish",
resnet_groups: int = 32,
resnet_pre_norm: bool = True,
num_attention_heads: int = 1,
output_scale_factor: float = 1.0,
cross_attention_dim: int = 1280,
dual_cross_attention: bool = False,
use_linear_projection: bool = False,
upcast_attention: bool = False,
attention_type: str = "default",
):
super().__init__()
self.has_cross_attention = True
self.num_attention_heads = num_attention_heads
resnet_groups = resnet_groups if resnet_groups is not None else min(in_channels // 4, 32)
# support for variable transformer layers per block
if isinstance(transformer_layers_per_block, int):
transformer_layers_per_block = [transformer_layers_per_block] * num_layers
# there is always at least one resnet
resnets = [
ResnetBlock2D(
in_channels=in_channels,
out_channels=in_channels,
temb_channels=temb_channels,
eps=resnet_eps,
groups=resnet_groups,
dropout=dropout,
time_embedding_norm=resnet_time_scale_shift,
non_linearity=resnet_act_fn,
output_scale_factor=output_scale_factor,
pre_norm=resnet_pre_norm,
)
]
attentions = []
for i in range(num_layers):
if not dual_cross_attention:
attentions.append(
Transformer2DModel(
num_attention_heads,
in_channels // num_attention_heads,
in_channels=in_channels,
num_layers=transformer_layers_per_block[i],
cross_attention_dim=cross_attention_dim,
norm_num_groups=resnet_groups,
use_linear_projection=use_linear_projection,
upcast_attention=upcast_attention,
attention_type=attention_type,
)
)
else:
attentions.append(
DualTransformer2DModel(
num_attention_heads,
in_channels // num_attention_heads,
in_channels=in_channels,
num_layers=1,
cross_attention_dim=cross_attention_dim,
norm_num_groups=resnet_groups,
)
)
resnets.append(
ResnetBlock2D(
in_channels=in_channels,
out_channels=in_channels,
temb_channels=temb_channels,
eps=resnet_eps,
groups=resnet_groups,
dropout=dropout,
time_embedding_norm=resnet_time_scale_shift,
non_linearity=resnet_act_fn,
output_scale_factor=output_scale_factor,
pre_norm=resnet_pre_norm,
)
)
self.attentions = nn.ModuleList(attentions)
self.resnets = nn.ModuleList(resnets)
self.gradient_checkpointing = False
def forward(
self,
hidden_states: torch.FloatTensor,
spatial_attn_inputs = [],
temb: Optional[torch.FloatTensor] = None,
encoder_hidden_states: Optional[torch.FloatTensor] = None,
attention_mask: Optional[torch.FloatTensor] = None,
cross_attention_kwargs: Optional[Dict[str, Any]] = None,
encoder_attention_mask: Optional[torch.FloatTensor] = None,
) -> torch.FloatTensor:
lora_scale = cross_attention_kwargs.get("scale", 1.0) if cross_attention_kwargs is not None else 1.0
hidden_states = self.resnets[0](hidden_states, temb, scale=lora_scale)
for attn, resnet in zip(self.attentions, self.resnets[1:]):
if self.training and self.gradient_checkpointing:
def create_custom_forward(module, return_dict=None):
def custom_forward(*inputs):
if return_dict is not None:
return module(*inputs, return_dict=return_dict)
else:
return module(*inputs)
return custom_forward
ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
hidden_states, spatial_attn_inputs = attn(
hidden_states,
spatial_attn_inputs=spatial_attn_inputs,
encoder_hidden_states=encoder_hidden_states,
cross_attention_kwargs=cross_attention_kwargs,
attention_mask=attention_mask,
encoder_attention_mask=encoder_attention_mask,
return_dict=False,
)
hidden_states = hidden_states[0]
hidden_states = torch.utils.checkpoint.checkpoint(
create_custom_forward(resnet),
hidden_states,
temb,
**ckpt_kwargs,
)
else:
hidden_states, spatial_attn_inputs = attn(
hidden_states,
spatial_attn_inputs=spatial_attn_inputs,
encoder_hidden_states=encoder_hidden_states,
cross_attention_kwargs=cross_attention_kwargs,
attention_mask=attention_mask,
encoder_attention_mask=encoder_attention_mask,
return_dict=False,
)
hidden_states = hidden_states[0]
hidden_states = resnet(hidden_states, temb, scale=lora_scale)
return hidden_states, spatial_attn_inputs
class UNetMidBlock2DSimpleCrossAttn(nn.Module):
def __init__(
self,
in_channels: int,
temb_channels: int,
dropout: float = 0.0,
num_layers: int = 1,
resnet_eps: float = 1e-6,
resnet_time_scale_shift: str = "default",
resnet_act_fn: str = "swish",
resnet_groups: int = 32,
resnet_pre_norm: bool = True,
attention_head_dim: int = 1,
output_scale_factor: float = 1.0,
cross_attention_dim: int = 1280,
skip_time_act: bool = False,
only_cross_attention: bool = False,
cross_attention_norm: Optional[str] = None,
):
super().__init__()
self.has_cross_attention = True
self.attention_head_dim = attention_head_dim
resnet_groups = resnet_groups if resnet_groups is not None else min(in_channels // 4, 32)
self.num_heads = in_channels // self.attention_head_dim
# there is always at least one resnet
resnets = [
ResnetBlock2D(
in_channels=in_channels,
out_channels=in_channels,
temb_channels=temb_channels,
eps=resnet_eps,
groups=resnet_groups,
dropout=dropout,
time_embedding_norm=resnet_time_scale_shift,
non_linearity=resnet_act_fn,
output_scale_factor=output_scale_factor,
pre_norm=resnet_pre_norm,
skip_time_act=skip_time_act,
)
]
attentions = []
for _ in range(num_layers):
processor = (
AttnAddedKVProcessor2_0() if hasattr(F, "scaled_dot_product_attention") else AttnAddedKVProcessor()
)
attentions.append(
Attention(
query_dim=in_channels,
cross_attention_dim=in_channels,
heads=self.num_heads,
dim_head=self.attention_head_dim,
added_kv_proj_dim=cross_attention_dim,
norm_num_groups=resnet_groups,
bias=True,
upcast_softmax=True,
only_cross_attention=only_cross_attention,
cross_attention_norm=cross_attention_norm,
processor=processor,
)
)
resnets.append(
ResnetBlock2D(
in_channels=in_channels,
out_channels=in_channels,
temb_channels=temb_channels,
eps=resnet_eps,
groups=resnet_groups,
dropout=dropout,
time_embedding_norm=resnet_time_scale_shift,
non_linearity=resnet_act_fn,
output_scale_factor=output_scale_factor,
pre_norm=resnet_pre_norm,
skip_time_act=skip_time_act,
)
)
self.attentions = nn.ModuleList(attentions)
self.resnets = nn.ModuleList(resnets)
def forward(
self,
hidden_states: torch.FloatTensor,
temb: Optional[torch.FloatTensor] = None,
encoder_hidden_states: Optional[torch.FloatTensor] = None,
attention_mask: Optional[torch.FloatTensor] = None,
cross_attention_kwargs: Optional[Dict[str, Any]] = None,
encoder_attention_mask: Optional[torch.FloatTensor] = None,
) -> torch.FloatTensor:
cross_attention_kwargs = cross_attention_kwargs if cross_attention_kwargs is not None else {}
lora_scale = cross_attention_kwargs.get("scale", 1.0)
if attention_mask is None:
# if encoder_hidden_states is defined: we are doing cross-attn, so we should use cross-attn mask.
mask = None if encoder_hidden_states is None else encoder_attention_mask
else:
# when attention_mask is defined: we don't even check for encoder_attention_mask.
# this is to maintain compatibility with UnCLIP, which uses 'attention_mask' param for cross-attn masks.
# TODO: UnCLIP should express cross-attn mask via encoder_attention_mask param instead of via attention_mask.
# then we can simplify this whole if/else block to:
# mask = attention_mask if encoder_hidden_states is None else encoder_attention_mask
mask = attention_mask
hidden_states = self.resnets[0](hidden_states, temb, scale=lora_scale)
for attn, resnet in zip(self.attentions, self.resnets[1:]):
# attn
hidden_states = attn(
hidden_states,
encoder_hidden_states=encoder_hidden_states,
attention_mask=mask,
**cross_attention_kwargs,
)
# resnet
hidden_states = resnet(hidden_states, temb, scale=lora_scale)
return hidden_states
class AttnDownBlock2D(nn.Module):
def __init__(
self,
in_channels: int,
out_channels: int,
temb_channels: int,
dropout: float = 0.0,
num_layers: int = 1,
resnet_eps: float = 1e-6,
resnet_time_scale_shift: str = "default",
resnet_act_fn: str = "swish",
resnet_groups: int = 32,
resnet_pre_norm: bool = True,
attention_head_dim: int = 1,
output_scale_factor: float = 1.0,
downsample_padding: int = 1,
downsample_type: str = "conv",
):
super().__init__()
resnets = []
attentions = []
self.downsample_type = downsample_type
if attention_head_dim is None:
logger.warn(
f"It is not recommend to pass `attention_head_dim=None`. Defaulting `attention_head_dim` to `in_channels`: {out_channels}."
)
attention_head_dim = out_channels
for i in range(num_layers):
in_channels = in_channels if i == 0 else out_channels
resnets.append(
ResnetBlock2D(
in_channels=in_channels,
out_channels=out_channels,
temb_channels=temb_channels,
eps=resnet_eps,
groups=resnet_groups,
dropout=dropout,
time_embedding_norm=resnet_time_scale_shift,
non_linearity=resnet_act_fn,
output_scale_factor=output_scale_factor,
pre_norm=resnet_pre_norm,
)
)
attentions.append(
Attention(
out_channels,
heads=out_channels // attention_head_dim,
dim_head=attention_head_dim,
rescale_output_factor=output_scale_factor,
eps=resnet_eps,
norm_num_groups=resnet_groups,
residual_connection=True,
bias=True,
upcast_softmax=True,
_from_deprecated_attn_block=True,
)
)
self.attentions = nn.ModuleList(attentions)
self.resnets = nn.ModuleList(resnets)
if downsample_type == "conv":
self.downsamplers = nn.ModuleList(
[
Downsample2D(
out_channels, use_conv=True, out_channels=out_channels, padding=downsample_padding, name="op"
)
]
)
elif downsample_type == "resnet":
self.downsamplers = nn.ModuleList(
[
ResnetBlock2D(
in_channels=out_channels,
out_channels=out_channels,
temb_channels=temb_channels,
eps=resnet_eps,
groups=resnet_groups,
dropout=dropout,
time_embedding_norm=resnet_time_scale_shift,
non_linearity=resnet_act_fn,
output_scale_factor=output_scale_factor,
pre_norm=resnet_pre_norm,
down=True,
)
]
)
else:
self.downsamplers = None
def forward(
self,
hidden_states: torch.FloatTensor,
temb: Optional[torch.FloatTensor] = None,
upsample_size: Optional[int] = None,
cross_attention_kwargs: Optional[Dict[str, Any]] = None,
) -> Tuple[torch.FloatTensor, Tuple[torch.FloatTensor, ...]]:
cross_attention_kwargs = cross_attention_kwargs if cross_attention_kwargs is not None else {}
lora_scale = cross_attention_kwargs.get("scale", 1.0)
output_states = ()
for resnet, attn in zip(self.resnets, self.attentions):
cross_attention_kwargs.update({"scale": lora_scale})
hidden_states = resnet(hidden_states, temb, scale=lora_scale)
hidden_states = attn(hidden_states, **cross_attention_kwargs)
output_states = output_states + (hidden_states,)
if self.downsamplers is not None:
for downsampler in self.downsamplers:
if self.downsample_type == "resnet":
hidden_states = downsampler(hidden_states, temb=temb, scale=lora_scale)
else:
hidden_states = downsampler(hidden_states, scale=lora_scale)
output_states += (hidden_states,)
return hidden_states, output_states
class CrossAttnDownBlock2D(nn.Module):
def __init__(
self,
in_channels: int,
out_channels: int,
temb_channels: int,
dropout: float = 0.0,
num_layers: int = 1,
transformer_layers_per_block: Union[int, Tuple[int]] = 1,
resnet_eps: float = 1e-6,
resnet_time_scale_shift: str = "default",
resnet_act_fn: str = "swish",
resnet_groups: int = 32,
resnet_pre_norm: bool = True,
num_attention_heads: int = 1,
cross_attention_dim: int = 1280,
output_scale_factor: float = 1.0,
downsample_padding: int = 1,
add_downsample: bool = True,
dual_cross_attention: bool = False,
use_linear_projection: bool = False,
only_cross_attention: bool = False,
upcast_attention: bool = False,
attention_type: str = "default",
):
super().__init__()
resnets = []
attentions = []
self.has_cross_attention = True
self.num_attention_heads = num_attention_heads
if isinstance(transformer_layers_per_block, int):
transformer_layers_per_block = [transformer_layers_per_block] * num_layers
for i in range(num_layers):
in_channels = in_channels if i == 0 else out_channels
resnets.append(
ResnetBlock2D(
in_channels=in_channels,
out_channels=out_channels,
temb_channels=temb_channels,
eps=resnet_eps,
groups=resnet_groups,
dropout=dropout,
time_embedding_norm=resnet_time_scale_shift,
non_linearity=resnet_act_fn,
output_scale_factor=output_scale_factor,
pre_norm=resnet_pre_norm,
)
)
if not dual_cross_attention:
attentions.append(
Transformer2DModel(
num_attention_heads,
out_channels // num_attention_heads,
in_channels=out_channels,
num_layers=transformer_layers_per_block[i],
cross_attention_dim=cross_attention_dim,
norm_num_groups=resnet_groups,
use_linear_projection=use_linear_projection,
only_cross_attention=only_cross_attention,
upcast_attention=upcast_attention,
attention_type=attention_type,
)
)
else:
attentions.append(
DualTransformer2DModel(
num_attention_heads,
out_channels // num_attention_heads,
in_channels=out_channels,
num_layers=1,
cross_attention_dim=cross_attention_dim,
norm_num_groups=resnet_groups,
)
)
self.attentions = nn.ModuleList(attentions)
self.resnets = nn.ModuleList(resnets)
if add_downsample:
self.downsamplers = nn.ModuleList(
[
Downsample2D(
out_channels, use_conv=True, out_channels=out_channels, padding=downsample_padding, name="op"
)
]
)
else:
self.downsamplers = None
self.gradient_checkpointing = False
def forward(
self,
hidden_states: torch.FloatTensor,
spatial_attn_inputs = [],
temb: Optional[torch.FloatTensor] = None,
encoder_hidden_states: Optional[torch.FloatTensor] = None,
attention_mask: Optional[torch.FloatTensor] = None,
cross_attention_kwargs: Optional[Dict[str, Any]] = None,
encoder_attention_mask: Optional[torch.FloatTensor] = None,
additional_residuals: Optional[torch.FloatTensor] = None,
) -> Tuple[torch.FloatTensor, Tuple[torch.FloatTensor, ...]]:
output_states = ()
lora_scale = cross_attention_kwargs.get("scale", 1.0) if cross_attention_kwargs is not None else 1.0
blocks = list(zip(self.resnets, self.attentions))
for i, (resnet, attn) in enumerate(blocks):
if self.training and self.gradient_checkpointing:
def create_custom_forward(module, return_dict=None):
def custom_forward(*inputs):
if return_dict is not None:
return module(*inputs, return_dict=return_dict)
else:
return module(*inputs)
return custom_forward
ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
hidden_states = torch.utils.checkpoint.checkpoint(
create_custom_forward(resnet),
hidden_states,
temb,
**ckpt_kwargs,
)
hidden_states, spatial_attn_inputs = attn(
hidden_states,
spatial_attn_inputs=spatial_attn_inputs,
encoder_hidden_states=encoder_hidden_states,
cross_attention_kwargs=cross_attention_kwargs,
attention_mask=attention_mask,
encoder_attention_mask=encoder_attention_mask,
return_dict=False,
)
hidden_states = hidden_states[0]
else:
hidden_states = resnet(hidden_states, temb, scale=lora_scale)
hidden_states, spatial_attn_inputs = attn(
hidden_states,
spatial_attn_inputs=spatial_attn_inputs,
encoder_hidden_states=encoder_hidden_states,
cross_attention_kwargs=cross_attention_kwargs,
attention_mask=attention_mask,
encoder_attention_mask=encoder_attention_mask,
return_dict=False,
)
hidden_states = hidden_states[0]
# apply additional residuals to the output of the last pair of resnet and attention blocks
if i == len(blocks) - 1 and additional_residuals is not None:
hidden_states = hidden_states + additional_residuals
output_states = output_states + (hidden_states,)
if self.downsamplers is not None:
for downsampler in self.downsamplers:
hidden_states = downsampler(hidden_states, scale=lora_scale)
output_states = output_states + (hidden_states,)
return hidden_states, output_states, spatial_attn_inputs
class DownBlock2D(nn.Module):
def __init__(
self,
in_channels: int,
out_channels: int,
temb_channels: int,
dropout: float = 0.0,
num_layers: int = 1,
resnet_eps: float = 1e-6,
resnet_time_scale_shift: str = "default",
resnet_act_fn: str = "swish",
resnet_groups: int = 32,
resnet_pre_norm: bool = True,
output_scale_factor: float = 1.0,
add_downsample: bool = True,
downsample_padding: int = 1,
):
super().__init__()
resnets = []
for i in range(num_layers):
in_channels = in_channels if i == 0 else out_channels
resnets.append(
ResnetBlock2D(
in_channels=in_channels,
out_channels=out_channels,
temb_channels=temb_channels,
eps=resnet_eps,
groups=resnet_groups,
dropout=dropout,
time_embedding_norm=resnet_time_scale_shift,
non_linearity=resnet_act_fn,
output_scale_factor=output_scale_factor,
pre_norm=resnet_pre_norm,
)
)
self.resnets = nn.ModuleList(resnets)
if add_downsample:
self.downsamplers = nn.ModuleList(
[
Downsample2D(
out_channels, use_conv=True, out_channels=out_channels, padding=downsample_padding, name="op"
)
]
)
else:
self.downsamplers = None
self.gradient_checkpointing = False
def forward(
self, hidden_states: torch.FloatTensor, temb: Optional[torch.FloatTensor] = None, scale: float = 1.0
) -> Tuple[torch.FloatTensor, Tuple[torch.FloatTensor, ...]]:
output_states = ()
for resnet in self.resnets:
if self.training and self.gradient_checkpointing:
def create_custom_forward(module):
def custom_forward(*inputs):
return module(*inputs)
return custom_forward
if is_torch_version(">=", "1.11.0"):
hidden_states = torch.utils.checkpoint.checkpoint(
create_custom_forward(resnet), hidden_states, temb, use_reentrant=False
)
else:
hidden_states = torch.utils.checkpoint.checkpoint(
create_custom_forward(resnet), hidden_states, temb
)
else:
hidden_states = resnet(hidden_states, temb, scale=scale)
output_states = output_states + (hidden_states,)
if self.downsamplers is not None:
for downsampler in self.downsamplers:
hidden_states = downsampler(hidden_states, scale=scale)
output_states = output_states + (hidden_states,)
return hidden_states, output_states
class DownEncoderBlock2D(nn.Module):
def __init__(
self,
in_channels: int,
out_channels: int,
dropout: float = 0.0,
num_layers: int = 1,
resnet_eps: float = 1e-6,
resnet_time_scale_shift: str = "default",
resnet_act_fn: str = "swish",
resnet_groups: int = 32,
resnet_pre_norm: bool = True,
output_scale_factor: float = 1.0,
add_downsample: bool = True,
downsample_padding: int = 1,
):
super().__init__()
resnets = []
for i in range(num_layers):
in_channels = in_channels if i == 0 else out_channels
resnets.append(
ResnetBlock2D(
in_channels=in_channels,
out_channels=out_channels,
temb_channels=None,
eps=resnet_eps,
groups=resnet_groups,
dropout=dropout,
time_embedding_norm=resnet_time_scale_shift,
non_linearity=resnet_act_fn,
output_scale_factor=output_scale_factor,
pre_norm=resnet_pre_norm,
)
)
self.resnets = nn.ModuleList(resnets)
if add_downsample:
self.downsamplers = nn.ModuleList(
[
Downsample2D(
out_channels, use_conv=True, out_channels=out_channels, padding=downsample_padding, name="op"
)
]
)
else:
self.downsamplers = None
def forward(self, hidden_states: torch.FloatTensor, scale: float = 1.0) -> torch.FloatTensor:
for resnet in self.resnets:
hidden_states = resnet(hidden_states, temb=None, scale=scale)
if self.downsamplers is not None:
for downsampler in self.downsamplers:
hidden_states = downsampler(hidden_states, scale)
return hidden_states
class AttnDownEncoderBlock2D(nn.Module):
def __init__(
self,
in_channels: int,
out_channels: int,
dropout: float = 0.0,
num_layers: int = 1,
resnet_eps: float = 1e-6,
resnet_time_scale_shift: str = "default",
resnet_act_fn: str = "swish",
resnet_groups: int = 32,
resnet_pre_norm: bool = True,
attention_head_dim: int = 1,
output_scale_factor: float = 1.0,
add_downsample: bool = True,
downsample_padding: int = 1,
):
super().__init__()
resnets = []
attentions = []
if attention_head_dim is None:
logger.warn(
f"It is not recommend to pass `attention_head_dim=None`. Defaulting `attention_head_dim` to `in_channels`: {out_channels}."
)
attention_head_dim = out_channels
for i in range(num_layers):
in_channels = in_channels if i == 0 else out_channels
resnets.append(
ResnetBlock2D(
in_channels=in_channels,
out_channels=out_channels,
temb_channels=None,
eps=resnet_eps,
groups=resnet_groups,
dropout=dropout,
time_embedding_norm=resnet_time_scale_shift,
non_linearity=resnet_act_fn,
output_scale_factor=output_scale_factor,
pre_norm=resnet_pre_norm,
)
)
attentions.append(
Attention(
out_channels,
heads=out_channels // attention_head_dim,
dim_head=attention_head_dim,
rescale_output_factor=output_scale_factor,
eps=resnet_eps,
norm_num_groups=resnet_groups,
residual_connection=True,
bias=True,
upcast_softmax=True,
_from_deprecated_attn_block=True,
)
)
self.attentions = nn.ModuleList(attentions)
self.resnets = nn.ModuleList(resnets)
if add_downsample:
self.downsamplers = nn.ModuleList(
[
Downsample2D(
out_channels, use_conv=True, out_channels=out_channels, padding=downsample_padding, name="op"
)
]
)
else:
self.downsamplers = None
def forward(self, hidden_states: torch.FloatTensor, scale: float = 1.0) -> torch.FloatTensor:
for resnet, attn in zip(self.resnets, self.attentions):
hidden_states = resnet(hidden_states, temb=None, scale=scale)
cross_attention_kwargs = {"scale": scale}
hidden_states = attn(hidden_states, **cross_attention_kwargs)
if self.downsamplers is not None:
for downsampler in self.downsamplers:
hidden_states = downsampler(hidden_states, scale)
return hidden_states
class AttnSkipDownBlock2D(nn.Module):
def __init__(
self,
in_channels: int,
out_channels: int,
temb_channels: int,
dropout: float = 0.0,
num_layers: int = 1,
resnet_eps: float = 1e-6,
resnet_time_scale_shift: str = "default",
resnet_act_fn: str = "swish",
resnet_pre_norm: bool = True,
attention_head_dim: int = 1,
output_scale_factor: float = np.sqrt(2.0),
add_downsample: bool = True,
):
super().__init__()
self.attentions = nn.ModuleList([])
self.resnets = nn.ModuleList([])
if attention_head_dim is None:
logger.warn(
f"It is not recommend to pass `attention_head_dim=None`. Defaulting `attention_head_dim` to `in_channels`: {out_channels}."
)
attention_head_dim = out_channels
for i in range(num_layers):
in_channels = in_channels if i == 0 else out_channels
self.resnets.append(
ResnetBlock2D(
in_channels=in_channels,
out_channels=out_channels,
temb_channels=temb_channels,
eps=resnet_eps,
groups=min(in_channels // 4, 32),
groups_out=min(out_channels // 4, 32),
dropout=dropout,
time_embedding_norm=resnet_time_scale_shift,
non_linearity=resnet_act_fn,
output_scale_factor=output_scale_factor,
pre_norm=resnet_pre_norm,
)
)
self.attentions.append(
Attention(
out_channels,
heads=out_channels // attention_head_dim,
dim_head=attention_head_dim,
rescale_output_factor=output_scale_factor,
eps=resnet_eps,
norm_num_groups=32,
residual_connection=True,
bias=True,
upcast_softmax=True,
_from_deprecated_attn_block=True,
)
)
if add_downsample:
self.resnet_down = ResnetBlock2D(
in_channels=out_channels,
out_channels=out_channels,
temb_channels=temb_channels,
eps=resnet_eps,
groups=min(out_channels // 4, 32),
dropout=dropout,
time_embedding_norm=resnet_time_scale_shift,
non_linearity=resnet_act_fn,
output_scale_factor=output_scale_factor,
pre_norm=resnet_pre_norm,
use_in_shortcut=True,
down=True,
kernel="fir",
)
self.downsamplers = nn.ModuleList([FirDownsample2D(out_channels, out_channels=out_channels)])
self.skip_conv = nn.Conv2d(3, out_channels, kernel_size=(1, 1), stride=(1, 1))
else:
self.resnet_down = None
self.downsamplers = None
self.skip_conv = None
def forward(
self,
hidden_states: torch.FloatTensor,
temb: Optional[torch.FloatTensor] = None,
skip_sample: Optional[torch.FloatTensor] = None,
scale: float = 1.0,
) -> Tuple[torch.FloatTensor, Tuple[torch.FloatTensor, ...], torch.FloatTensor]:
output_states = ()
for resnet, attn in zip(self.resnets, self.attentions):
hidden_states = resnet(hidden_states, temb, scale=scale)
cross_attention_kwargs = {"scale": scale}
hidden_states = attn(hidden_states, **cross_attention_kwargs)
output_states += (hidden_states,)
if self.downsamplers is not None:
hidden_states = self.resnet_down(hidden_states, temb, scale=scale)
for downsampler in self.downsamplers:
skip_sample = downsampler(skip_sample)
hidden_states = self.skip_conv(skip_sample) + hidden_states
output_states += (hidden_states,)
return hidden_states, output_states, skip_sample
class SkipDownBlock2D(nn.Module):
def __init__(
self,
in_channels: int,
out_channels: int,
temb_channels: int,
dropout: float = 0.0,
num_layers: int = 1,
resnet_eps: float = 1e-6,
resnet_time_scale_shift: str = "default",
resnet_act_fn: str = "swish",
resnet_pre_norm: bool = True,
output_scale_factor: float = np.sqrt(2.0),
add_downsample: bool = True,
downsample_padding: int = 1,
):
super().__init__()
self.resnets = nn.ModuleList([])
for i in range(num_layers):
in_channels = in_channels if i == 0 else out_channels
self.resnets.append(
ResnetBlock2D(
in_channels=in_channels,
out_channels=out_channels,
temb_channels=temb_channels,
eps=resnet_eps,
groups=min(in_channels // 4, 32),
groups_out=min(out_channels // 4, 32),
dropout=dropout,
time_embedding_norm=resnet_time_scale_shift,
non_linearity=resnet_act_fn,
output_scale_factor=output_scale_factor,
pre_norm=resnet_pre_norm,
)
)
if add_downsample:
self.resnet_down = ResnetBlock2D(
in_channels=out_channels,
out_channels=out_channels,
temb_channels=temb_channels,
eps=resnet_eps,
groups=min(out_channels // 4, 32),
dropout=dropout,
time_embedding_norm=resnet_time_scale_shift,
non_linearity=resnet_act_fn,
output_scale_factor=output_scale_factor,
pre_norm=resnet_pre_norm,
use_in_shortcut=True,
down=True,
kernel="fir",
)
self.downsamplers = nn.ModuleList([FirDownsample2D(out_channels, out_channels=out_channels)])
self.skip_conv = nn.Conv2d(3, out_channels, kernel_size=(1, 1), stride=(1, 1))
else:
self.resnet_down = None
self.downsamplers = None
self.skip_conv = None
def forward(
self,
hidden_states: torch.FloatTensor,
temb: Optional[torch.FloatTensor] = None,
skip_sample: Optional[torch.FloatTensor] = None,
scale: float = 1.0,
) -> Tuple[torch.FloatTensor, Tuple[torch.FloatTensor, ...], torch.FloatTensor]:
output_states = ()
for resnet in self.resnets:
hidden_states = resnet(hidden_states, temb, scale)
output_states += (hidden_states,)
if self.downsamplers is not None:
hidden_states = self.resnet_down(hidden_states, temb, scale)
for downsampler in self.downsamplers:
skip_sample = downsampler(skip_sample)
hidden_states = self.skip_conv(skip_sample) + hidden_states
output_states += (hidden_states,)
return hidden_states, output_states, skip_sample
class ResnetDownsampleBlock2D(nn.Module):
def __init__(
self,
in_channels: int,
out_channels: int,
temb_channels: int,
dropout: float = 0.0,
num_layers: int = 1,
resnet_eps: float = 1e-6,
resnet_time_scale_shift: str = "default",
resnet_act_fn: str = "swish",
resnet_groups: int = 32,
resnet_pre_norm: bool = True,
output_scale_factor: float = 1.0,
add_downsample: bool = True,
skip_time_act: bool = False,
):
super().__init__()
resnets = []
for i in range(num_layers):
in_channels = in_channels if i == 0 else out_channels
resnets.append(
ResnetBlock2D(
in_channels=in_channels,
out_channels=out_channels,
temb_channels=temb_channels,
eps=resnet_eps,
groups=resnet_groups,
dropout=dropout,
time_embedding_norm=resnet_time_scale_shift,
non_linearity=resnet_act_fn,
output_scale_factor=output_scale_factor,
pre_norm=resnet_pre_norm,
skip_time_act=skip_time_act,
)
)
self.resnets = nn.ModuleList(resnets)
if add_downsample:
self.downsamplers = nn.ModuleList(
[
ResnetBlock2D(
in_channels=out_channels,
out_channels=out_channels,
temb_channels=temb_channels,
eps=resnet_eps,
groups=resnet_groups,
dropout=dropout,
time_embedding_norm=resnet_time_scale_shift,
non_linearity=resnet_act_fn,
output_scale_factor=output_scale_factor,
pre_norm=resnet_pre_norm,
skip_time_act=skip_time_act,
down=True,
)
]
)
else:
self.downsamplers = None
self.gradient_checkpointing = False
def forward(
self, hidden_states: torch.FloatTensor, temb: Optional[torch.FloatTensor] = None, scale: float = 1.0
) -> Tuple[torch.FloatTensor, Tuple[torch.FloatTensor, ...]]:
output_states = ()
for resnet in self.resnets:
if self.training and self.gradient_checkpointing:
def create_custom_forward(module):
def custom_forward(*inputs):
return module(*inputs)
return custom_forward
if is_torch_version(">=", "1.11.0"):
hidden_states = torch.utils.checkpoint.checkpoint(
create_custom_forward(resnet), hidden_states, temb, use_reentrant=False
)
else:
hidden_states = torch.utils.checkpoint.checkpoint(
create_custom_forward(resnet), hidden_states, temb
)
else:
hidden_states = resnet(hidden_states, temb, scale)
output_states = output_states + (hidden_states,)
if self.downsamplers is not None:
for downsampler in self.downsamplers:
hidden_states = downsampler(hidden_states, temb, scale)
output_states = output_states + (hidden_states,)
return hidden_states, output_states
class SimpleCrossAttnDownBlock2D(nn.Module):
def __init__(
self,
in_channels: int,
out_channels: int,
temb_channels: int,
dropout: float = 0.0,
num_layers: int = 1,
resnet_eps: float = 1e-6,
resnet_time_scale_shift: str = "default",
resnet_act_fn: str = "swish",
resnet_groups: int = 32,
resnet_pre_norm: bool = True,
attention_head_dim: int = 1,
cross_attention_dim: int = 1280,
output_scale_factor: float = 1.0,
add_downsample: bool = True,
skip_time_act: bool = False,
only_cross_attention: bool = False,
cross_attention_norm: Optional[str] = None,
):
super().__init__()
self.has_cross_attention = True
resnets = []
attentions = []
self.attention_head_dim = attention_head_dim
self.num_heads = out_channels // self.attention_head_dim
for i in range(num_layers):
in_channels = in_channels if i == 0 else out_channels
resnets.append(
ResnetBlock2D(
in_channels=in_channels,
out_channels=out_channels,
temb_channels=temb_channels,
eps=resnet_eps,
groups=resnet_groups,
dropout=dropout,
time_embedding_norm=resnet_time_scale_shift,
non_linearity=resnet_act_fn,
output_scale_factor=output_scale_factor,
pre_norm=resnet_pre_norm,
skip_time_act=skip_time_act,
)
)
processor = (
AttnAddedKVProcessor2_0() if hasattr(F, "scaled_dot_product_attention") else AttnAddedKVProcessor()
)
attentions.append(
Attention(
query_dim=out_channels,
cross_attention_dim=out_channels,
heads=self.num_heads,
dim_head=attention_head_dim,
added_kv_proj_dim=cross_attention_dim,
norm_num_groups=resnet_groups,
bias=True,
upcast_softmax=True,
only_cross_attention=only_cross_attention,
cross_attention_norm=cross_attention_norm,
processor=processor,
)
)
self.attentions = nn.ModuleList(attentions)
self.resnets = nn.ModuleList(resnets)
if add_downsample:
self.downsamplers = nn.ModuleList(
[
ResnetBlock2D(
in_channels=out_channels,
out_channels=out_channels,
temb_channels=temb_channels,
eps=resnet_eps,
groups=resnet_groups,
dropout=dropout,
time_embedding_norm=resnet_time_scale_shift,
non_linearity=resnet_act_fn,
output_scale_factor=output_scale_factor,
pre_norm=resnet_pre_norm,
skip_time_act=skip_time_act,
down=True,
)
]
)
else:
self.downsamplers = None
self.gradient_checkpointing = False
def forward(
self,
hidden_states: torch.FloatTensor,
temb: Optional[torch.FloatTensor] = None,
encoder_hidden_states: Optional[torch.FloatTensor] = None,
attention_mask: Optional[torch.FloatTensor] = None,
cross_attention_kwargs: Optional[Dict[str, Any]] = None,
encoder_attention_mask: Optional[torch.FloatTensor] = None,
) -> Tuple[torch.FloatTensor, Tuple[torch.FloatTensor, ...]]:
output_states = ()
cross_attention_kwargs = cross_attention_kwargs if cross_attention_kwargs is not None else {}
lora_scale = cross_attention_kwargs.get("scale", 1.0)
if attention_mask is None:
# if encoder_hidden_states is defined: we are doing cross-attn, so we should use cross-attn mask.
mask = None if encoder_hidden_states is None else encoder_attention_mask
else:
# when attention_mask is defined: we don't even check for encoder_attention_mask.
# this is to maintain compatibility with UnCLIP, which uses 'attention_mask' param for cross-attn masks.
# TODO: UnCLIP should express cross-attn mask via encoder_attention_mask param instead of via attention_mask.
# then we can simplify this whole if/else block to:
# mask = attention_mask if encoder_hidden_states is None else encoder_attention_mask
mask = attention_mask
for resnet, attn in zip(self.resnets, self.attentions):
if self.training and self.gradient_checkpointing:
def create_custom_forward(module, return_dict=None):
def custom_forward(*inputs):
if return_dict is not None:
return module(*inputs, return_dict=return_dict)
else:
return module(*inputs)
return custom_forward
hidden_states = torch.utils.checkpoint.checkpoint(create_custom_forward(resnet), hidden_states, temb)
hidden_states = attn(
hidden_states,
encoder_hidden_states=encoder_hidden_states,
attention_mask=mask,
**cross_attention_kwargs,
)
else:
hidden_states = resnet(hidden_states, temb, scale=lora_scale)
hidden_states = attn(
hidden_states,
encoder_hidden_states=encoder_hidden_states,
attention_mask=mask,
**cross_attention_kwargs,
)
output_states = output_states + (hidden_states,)
if self.downsamplers is not None:
for downsampler in self.downsamplers:
hidden_states = downsampler(hidden_states, temb, scale=lora_scale)
output_states = output_states + (hidden_states,)
return hidden_states, output_states
class KDownBlock2D(nn.Module):
def __init__(
self,
in_channels: int,
out_channels: int,
temb_channels: int,
dropout: float = 0.0,
num_layers: int = 4,
resnet_eps: float = 1e-5,
resnet_act_fn: str = "gelu",
resnet_group_size: int = 32,
add_downsample: bool = False,
):
super().__init__()
resnets = []
for i in range(num_layers):
in_channels = in_channels if i == 0 else out_channels
groups = in_channels // resnet_group_size
groups_out = out_channels // resnet_group_size
resnets.append(
ResnetBlock2D(
in_channels=in_channels,
out_channels=out_channels,
dropout=dropout,
temb_channels=temb_channels,
groups=groups,
groups_out=groups_out,
eps=resnet_eps,
non_linearity=resnet_act_fn,
time_embedding_norm="ada_group",
conv_shortcut_bias=False,
)
)
self.resnets = nn.ModuleList(resnets)
if add_downsample:
# YiYi's comments- might be able to use FirDownsample2D, look into details later
self.downsamplers = nn.ModuleList([KDownsample2D()])
else:
self.downsamplers = None
self.gradient_checkpointing = False
def forward(
self, hidden_states: torch.FloatTensor, temb: Optional[torch.FloatTensor] = None, scale: float = 1.0
) -> Tuple[torch.FloatTensor, Tuple[torch.FloatTensor, ...]]:
output_states = ()
for resnet in self.resnets:
if self.training and self.gradient_checkpointing:
def create_custom_forward(module):
def custom_forward(*inputs):
return module(*inputs)
return custom_forward
if is_torch_version(">=", "1.11.0"):
hidden_states = torch.utils.checkpoint.checkpoint(
create_custom_forward(resnet), hidden_states, temb, use_reentrant=False
)
else:
hidden_states = torch.utils.checkpoint.checkpoint(
create_custom_forward(resnet), hidden_states, temb
)
else:
hidden_states = resnet(hidden_states, temb, scale)
output_states += (hidden_states,)
if self.downsamplers is not None:
for downsampler in self.downsamplers:
hidden_states = downsampler(hidden_states)
return hidden_states, output_states
class KCrossAttnDownBlock2D(nn.Module):
def __init__(
self,
in_channels: int,
out_channels: int,
temb_channels: int,
cross_attention_dim: int,
dropout: float = 0.0,
num_layers: int = 4,
resnet_group_size: int = 32,
add_downsample: bool = True,
attention_head_dim: int = 64,
add_self_attention: bool = False,
resnet_eps: float = 1e-5,
resnet_act_fn: str = "gelu",
):
super().__init__()
resnets = []
attentions = []
self.has_cross_attention = True
for i in range(num_layers):
in_channels = in_channels if i == 0 else out_channels
groups = in_channels // resnet_group_size
groups_out = out_channels // resnet_group_size
resnets.append(
ResnetBlock2D(
in_channels=in_channels,
out_channels=out_channels,
dropout=dropout,
temb_channels=temb_channels,
groups=groups,
groups_out=groups_out,
eps=resnet_eps,
non_linearity=resnet_act_fn,
time_embedding_norm="ada_group",
conv_shortcut_bias=False,
)
)
attentions.append(
KAttentionBlock(
out_channels,
out_channels // attention_head_dim,
attention_head_dim,
cross_attention_dim=cross_attention_dim,
temb_channels=temb_channels,
attention_bias=True,
add_self_attention=add_self_attention,
cross_attention_norm="layer_norm",
group_size=resnet_group_size,
)
)
self.resnets = nn.ModuleList(resnets)
self.attentions = nn.ModuleList(attentions)
if add_downsample:
self.downsamplers = nn.ModuleList([KDownsample2D()])
else:
self.downsamplers = None
self.gradient_checkpointing = False
def forward(
self,
hidden_states: torch.FloatTensor,
temb: Optional[torch.FloatTensor] = None,
encoder_hidden_states: Optional[torch.FloatTensor] = None,
attention_mask: Optional[torch.FloatTensor] = None,
cross_attention_kwargs: Optional[Dict[str, Any]] = None,
encoder_attention_mask: Optional[torch.FloatTensor] = None,
) -> Tuple[torch.FloatTensor, Tuple[torch.FloatTensor, ...]]:
output_states = ()
lora_scale = cross_attention_kwargs.get("scale", 1.0) if cross_attention_kwargs is not None else 1.0
for resnet, attn in zip(self.resnets, self.attentions):
if self.training and self.gradient_checkpointing:
def create_custom_forward(module, return_dict=None):
def custom_forward(*inputs):
if return_dict is not None:
return module(*inputs, return_dict=return_dict)
else:
return module(*inputs)
return custom_forward
ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
hidden_states = torch.utils.checkpoint.checkpoint(
create_custom_forward(resnet),
hidden_states,
temb,
**ckpt_kwargs,
)
hidden_states = attn(
hidden_states,
encoder_hidden_states=encoder_hidden_states,
emb=temb,
attention_mask=attention_mask,
cross_attention_kwargs=cross_attention_kwargs,
encoder_attention_mask=encoder_attention_mask,
)
else:
hidden_states = resnet(hidden_states, temb, scale=lora_scale)
hidden_states = attn(
hidden_states,
encoder_hidden_states=encoder_hidden_states,
emb=temb,
attention_mask=attention_mask,
cross_attention_kwargs=cross_attention_kwargs,
encoder_attention_mask=encoder_attention_mask,
)
if self.downsamplers is None:
output_states += (None,)
else:
output_states += (hidden_states,)
if self.downsamplers is not None:
for downsampler in self.downsamplers:
hidden_states = downsampler(hidden_states)
return hidden_states, output_states
class AttnUpBlock2D(nn.Module):
def __init__(
self,
in_channels: int,
prev_output_channel: int,
out_channels: int,
temb_channels: int,
resolution_idx: int = None,
dropout: float = 0.0,
num_layers: int = 1,
resnet_eps: float = 1e-6,
resnet_time_scale_shift: str = "default",
resnet_act_fn: str = "swish",
resnet_groups: int = 32,
resnet_pre_norm: bool = True,
attention_head_dim: int = 1,
output_scale_factor: float = 1.0,
upsample_type: str = "conv",
):
super().__init__()
resnets = []
attentions = []
self.upsample_type = upsample_type
if attention_head_dim is None:
logger.warn(
f"It is not recommend to pass `attention_head_dim=None`. Defaulting `attention_head_dim` to `in_channels`: {out_channels}."
)
attention_head_dim = out_channels
for i in range(num_layers):
res_skip_channels = in_channels if (i == num_layers - 1) else out_channels
resnet_in_channels = prev_output_channel if i == 0 else out_channels
resnets.append(
ResnetBlock2D(
in_channels=resnet_in_channels + res_skip_channels,
out_channels=out_channels,
temb_channels=temb_channels,
eps=resnet_eps,
groups=resnet_groups,
dropout=dropout,
time_embedding_norm=resnet_time_scale_shift,
non_linearity=resnet_act_fn,
output_scale_factor=output_scale_factor,
pre_norm=resnet_pre_norm,
)
)
attentions.append(
Attention(
out_channels,
heads=out_channels // attention_head_dim,
dim_head=attention_head_dim,
rescale_output_factor=output_scale_factor,
eps=resnet_eps,
norm_num_groups=resnet_groups,
residual_connection=True,
bias=True,
upcast_softmax=True,
_from_deprecated_attn_block=True,
)
)
self.attentions = nn.ModuleList(attentions)
self.resnets = nn.ModuleList(resnets)
if upsample_type == "conv":
self.upsamplers = nn.ModuleList([Upsample2D(out_channels, use_conv=True, out_channels=out_channels)])
elif upsample_type == "resnet":
self.upsamplers = nn.ModuleList(
[
ResnetBlock2D(
in_channels=out_channels,
out_channels=out_channels,
temb_channels=temb_channels,
eps=resnet_eps,
groups=resnet_groups,
dropout=dropout,
time_embedding_norm=resnet_time_scale_shift,
non_linearity=resnet_act_fn,
output_scale_factor=output_scale_factor,
pre_norm=resnet_pre_norm,
up=True,
)
]
)
else:
self.upsamplers = None
self.resolution_idx = resolution_idx
def forward(
self,
hidden_states: torch.FloatTensor,
res_hidden_states_tuple: Tuple[torch.FloatTensor, ...],
temb: Optional[torch.FloatTensor] = None,
upsample_size: Optional[int] = None,
scale: float = 1.0,
) -> torch.FloatTensor:
for resnet, attn in zip(self.resnets, self.attentions):
# pop res hidden states
res_hidden_states = res_hidden_states_tuple[-1]
res_hidden_states_tuple = res_hidden_states_tuple[:-1]
hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)
hidden_states = resnet(hidden_states, temb, scale=scale)
cross_attention_kwargs = {"scale": scale}
hidden_states = attn(hidden_states, **cross_attention_kwargs)
if self.upsamplers is not None:
for upsampler in self.upsamplers:
if self.upsample_type == "resnet":
hidden_states = upsampler(hidden_states, temb=temb, scale=scale)
else:
hidden_states = upsampler(hidden_states, scale=scale)
return hidden_states
class CrossAttnUpBlock2D(nn.Module):
def __init__(
self,
in_channels: int,
out_channels: int,
prev_output_channel: int,
temb_channels: int,
resolution_idx: Optional[int] = None,
dropout: float = 0.0,
num_layers: int = 1,
transformer_layers_per_block: Union[int, Tuple[int]] = 1,
resnet_eps: float = 1e-6,
resnet_time_scale_shift: str = "default",
resnet_act_fn: str = "swish",
resnet_groups: int = 32,
resnet_pre_norm: bool = True,
num_attention_heads: int = 1,
cross_attention_dim: int = 1280,
output_scale_factor: float = 1.0,
add_upsample: bool = True,
dual_cross_attention: bool = False,
use_linear_projection: bool = False,
only_cross_attention: bool = False,
upcast_attention: bool = False,
attention_type: str = "default",
):
super().__init__()
resnets = []
attentions = []
self.has_cross_attention = True
self.num_attention_heads = num_attention_heads
if isinstance(transformer_layers_per_block, int):
transformer_layers_per_block = [transformer_layers_per_block] * num_layers
for i in range(num_layers):
res_skip_channels = in_channels if (i == num_layers - 1) else out_channels
resnet_in_channels = prev_output_channel if i == 0 else out_channels
resnets.append(
ResnetBlock2D(
in_channels=resnet_in_channels + res_skip_channels,
out_channels=out_channels,
temb_channels=temb_channels,
eps=resnet_eps,
groups=resnet_groups,
dropout=dropout,
time_embedding_norm=resnet_time_scale_shift,
non_linearity=resnet_act_fn,
output_scale_factor=output_scale_factor,
pre_norm=resnet_pre_norm,
)
)
if not dual_cross_attention:
attentions.append(
Transformer2DModel(
num_attention_heads,
out_channels // num_attention_heads,
in_channels=out_channels,
num_layers=transformer_layers_per_block[i],
cross_attention_dim=cross_attention_dim,
norm_num_groups=resnet_groups,
use_linear_projection=use_linear_projection,
only_cross_attention=only_cross_attention,
upcast_attention=upcast_attention,
attention_type=attention_type,
)
)
else:
attentions.append(
DualTransformer2DModel(
num_attention_heads,
out_channels // num_attention_heads,
in_channels=out_channels,
num_layers=1,
cross_attention_dim=cross_attention_dim,
norm_num_groups=resnet_groups,
)
)
self.attentions = nn.ModuleList(attentions)
self.resnets = nn.ModuleList(resnets)
if add_upsample:
self.upsamplers = nn.ModuleList([Upsample2D(out_channels, use_conv=True, out_channels=out_channels)])
else:
self.upsamplers = None
self.gradient_checkpointing = False
self.resolution_idx = resolution_idx
def forward(
self,
hidden_states: torch.FloatTensor,
res_hidden_states_tuple: Tuple[torch.FloatTensor, ...],
spatial_attn_inputs = [],
temb: Optional[torch.FloatTensor] = None,
encoder_hidden_states: Optional[torch.FloatTensor] = None,
cross_attention_kwargs: Optional[Dict[str, Any]] = None,
upsample_size: Optional[int] = None,
attention_mask: Optional[torch.FloatTensor] = None,
encoder_attention_mask: Optional[torch.FloatTensor] = None,
) -> torch.FloatTensor:
lora_scale = cross_attention_kwargs.get("scale", 1.0) if cross_attention_kwargs is not None else 1.0
is_freeu_enabled = (
getattr(self, "s1", None)
and getattr(self, "s2", None)
and getattr(self, "b1", None)
and getattr(self, "b2", None)
)
for resnet, attn in zip(self.resnets, self.attentions):
# pop res hidden states
res_hidden_states = res_hidden_states_tuple[-1]
res_hidden_states_tuple = res_hidden_states_tuple[:-1]
# FreeU: Only operate on the first two stages
if is_freeu_enabled:
hidden_states, res_hidden_states = apply_freeu(
self.resolution_idx,
hidden_states,
res_hidden_states,
s1=self.s1,
s2=self.s2,
b1=self.b1,
b2=self.b2,
)
hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)
if self.training and self.gradient_checkpointing:
def create_custom_forward(module, return_dict=None):
def custom_forward(*inputs):
if return_dict is not None:
return module(*inputs, return_dict=return_dict)
else:
return module(*inputs)
return custom_forward
ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
hidden_states = torch.utils.checkpoint.checkpoint(
create_custom_forward(resnet),
hidden_states,
temb,
**ckpt_kwargs,
)
hidden_states, spatial_attn_inputs = attn(
hidden_states,
spatial_attn_inputs=spatial_attn_inputs,
encoder_hidden_states=encoder_hidden_states,
cross_attention_kwargs=cross_attention_kwargs,
attention_mask=attention_mask,
encoder_attention_mask=encoder_attention_mask,
return_dict=False,
)
hidden_states = hidden_states[0]
else:
hidden_states = resnet(hidden_states, temb, scale=lora_scale)
hidden_states, spatial_attn_inputs = attn(
hidden_states,
spatial_attn_inputs=spatial_attn_inputs,
encoder_hidden_states=encoder_hidden_states,
cross_attention_kwargs=cross_attention_kwargs,
attention_mask=attention_mask,
encoder_attention_mask=encoder_attention_mask,
return_dict=False,
)
hidden_states = hidden_states[0]
if self.upsamplers is not None:
for upsampler in self.upsamplers:
hidden_states = upsampler(hidden_states, upsample_size, scale=lora_scale)
return hidden_states, spatial_attn_inputs
class UpBlock2D(nn.Module):
def __init__(
self,
in_channels: int,
prev_output_channel: int,
out_channels: int,
temb_channels: int,
resolution_idx: Optional[int] = None,
dropout: float = 0.0,
num_layers: int = 1,
resnet_eps: float = 1e-6,
resnet_time_scale_shift: str = "default",
resnet_act_fn: str = "swish",
resnet_groups: int = 32,
resnet_pre_norm: bool = True,
output_scale_factor: float = 1.0,
add_upsample: bool = True,
):
super().__init__()
resnets = []
for i in range(num_layers):
res_skip_channels = in_channels if (i == num_layers - 1) else out_channels
resnet_in_channels = prev_output_channel if i == 0 else out_channels
resnets.append(
ResnetBlock2D(
in_channels=resnet_in_channels + res_skip_channels,
out_channels=out_channels,
temb_channels=temb_channels,
eps=resnet_eps,
groups=resnet_groups,
dropout=dropout,
time_embedding_norm=resnet_time_scale_shift,
non_linearity=resnet_act_fn,
output_scale_factor=output_scale_factor,
pre_norm=resnet_pre_norm,
)
)
self.resnets = nn.ModuleList(resnets)
if add_upsample:
self.upsamplers = nn.ModuleList([Upsample2D(out_channels, use_conv=True, out_channels=out_channels)])
else:
self.upsamplers = None
self.gradient_checkpointing = False
self.resolution_idx = resolution_idx
def forward(
self,
hidden_states: torch.FloatTensor,
res_hidden_states_tuple: Tuple[torch.FloatTensor, ...],
temb: Optional[torch.FloatTensor] = None,
upsample_size: Optional[int] = None,
scale: float = 1.0,
) -> torch.FloatTensor:
is_freeu_enabled = (
getattr(self, "s1", None)
and getattr(self, "s2", None)
and getattr(self, "b1", None)
and getattr(self, "b2", None)
)
for resnet in self.resnets:
# pop res hidden states
res_hidden_states = res_hidden_states_tuple[-1]
res_hidden_states_tuple = res_hidden_states_tuple[:-1]
# FreeU: Only operate on the first two stages
if is_freeu_enabled:
hidden_states, res_hidden_states = apply_freeu(
self.resolution_idx,
hidden_states,
res_hidden_states,
s1=self.s1,
s2=self.s2,
b1=self.b1,
b2=self.b2,
)
hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)
if self.training and self.gradient_checkpointing:
def create_custom_forward(module):
def custom_forward(*inputs):
return module(*inputs)
return custom_forward
if is_torch_version(">=", "1.11.0"):
hidden_states = torch.utils.checkpoint.checkpoint(
create_custom_forward(resnet), hidden_states, temb, use_reentrant=False
)
else:
hidden_states = torch.utils.checkpoint.checkpoint(
create_custom_forward(resnet), hidden_states, temb
)
else:
hidden_states = resnet(hidden_states, temb, scale=scale)
if self.upsamplers is not None:
for upsampler in self.upsamplers:
hidden_states = upsampler(hidden_states, upsample_size, scale=scale)
return hidden_states
class UpDecoderBlock2D(nn.Module):
def __init__(
self,
in_channels: int,
out_channels: int,
resolution_idx: Optional[int] = None,
dropout: float = 0.0,
num_layers: int = 1,
resnet_eps: float = 1e-6,
resnet_time_scale_shift: str = "default", # default, spatial
resnet_act_fn: str = "swish",
resnet_groups: int = 32,
resnet_pre_norm: bool = True,
output_scale_factor: float = 1.0,
add_upsample: bool = True,
temb_channels: Optional[int] = None,
):
super().__init__()
resnets = []
for i in range(num_layers):
input_channels = in_channels if i == 0 else out_channels
resnets.append(
ResnetBlock2D(
in_channels=input_channels,
out_channels=out_channels,
temb_channels=temb_channels,
eps=resnet_eps,
groups=resnet_groups,
dropout=dropout,
time_embedding_norm=resnet_time_scale_shift,
non_linearity=resnet_act_fn,
output_scale_factor=output_scale_factor,
pre_norm=resnet_pre_norm,
)
)
self.resnets = nn.ModuleList(resnets)
if add_upsample:
self.upsamplers = nn.ModuleList([Upsample2D(out_channels, use_conv=True, out_channels=out_channels)])
else:
self.upsamplers = None
self.resolution_idx = resolution_idx
def forward(
self, hidden_states: torch.FloatTensor, temb: Optional[torch.FloatTensor] = None, scale: float = 1.0
) -> torch.FloatTensor:
for resnet in self.resnets:
hidden_states = resnet(hidden_states, temb=temb, scale=scale)
if self.upsamplers is not None:
for upsampler in self.upsamplers:
hidden_states = upsampler(hidden_states)
return hidden_states
class AttnUpDecoderBlock2D(nn.Module):
def __init__(
self,
in_channels: int,
out_channels: int,
resolution_idx: Optional[int] = None,
dropout: float = 0.0,
num_layers: int = 1,
resnet_eps: float = 1e-6,
resnet_time_scale_shift: str = "default",
resnet_act_fn: str = "swish",
resnet_groups: int = 32,
resnet_pre_norm: bool = True,
attention_head_dim: int = 1,
output_scale_factor: float = 1.0,
add_upsample: bool = True,
temb_channels: Optional[int] = None,
):
super().__init__()
resnets = []
attentions = []
if attention_head_dim is None:
logger.warn(
f"It is not recommend to pass `attention_head_dim=None`. Defaulting `attention_head_dim` to `out_channels`: {out_channels}."
)
attention_head_dim = out_channels
for i in range(num_layers):
input_channels = in_channels if i == 0 else out_channels
resnets.append(
ResnetBlock2D(
in_channels=input_channels,
out_channels=out_channels,
temb_channels=temb_channels,
eps=resnet_eps,
groups=resnet_groups,
dropout=dropout,
time_embedding_norm=resnet_time_scale_shift,
non_linearity=resnet_act_fn,
output_scale_factor=output_scale_factor,
pre_norm=resnet_pre_norm,
)
)
attentions.append(
Attention(
out_channels,
heads=out_channels // attention_head_dim,
dim_head=attention_head_dim,
rescale_output_factor=output_scale_factor,
eps=resnet_eps,
norm_num_groups=resnet_groups if resnet_time_scale_shift != "spatial" else None,
spatial_norm_dim=temb_channels if resnet_time_scale_shift == "spatial" else None,
residual_connection=True,
bias=True,
upcast_softmax=True,
_from_deprecated_attn_block=True,
)
)
self.attentions = nn.ModuleList(attentions)
self.resnets = nn.ModuleList(resnets)
if add_upsample:
self.upsamplers = nn.ModuleList([Upsample2D(out_channels, use_conv=True, out_channels=out_channels)])
else:
self.upsamplers = None
self.resolution_idx = resolution_idx
def forward(
self, hidden_states: torch.FloatTensor, temb: Optional[torch.FloatTensor] = None, scale: float = 1.0
) -> torch.FloatTensor:
for resnet, attn in zip(self.resnets, self.attentions):
hidden_states = resnet(hidden_states, temb=temb, scale=scale)
cross_attention_kwargs = {"scale": scale}
hidden_states = attn(hidden_states, temb=temb, **cross_attention_kwargs)
if self.upsamplers is not None:
for upsampler in self.upsamplers:
hidden_states = upsampler(hidden_states, scale=scale)
return hidden_states
class AttnSkipUpBlock2D(nn.Module):
def __init__(
self,
in_channels: int,
prev_output_channel: int,
out_channels: int,
temb_channels: int,
resolution_idx: Optional[int] = None,
dropout: float = 0.0,
num_layers: int = 1,
resnet_eps: float = 1e-6,
resnet_time_scale_shift: str = "default",
resnet_act_fn: str = "swish",
resnet_pre_norm: bool = True,
attention_head_dim: int = 1,
output_scale_factor: float = np.sqrt(2.0),
add_upsample: bool = True,
):
super().__init__()
self.attentions = nn.ModuleList([])
self.resnets = nn.ModuleList([])
for i in range(num_layers):
res_skip_channels = in_channels if (i == num_layers - 1) else out_channels
resnet_in_channels = prev_output_channel if i == 0 else out_channels
self.resnets.append(
ResnetBlock2D(
in_channels=resnet_in_channels + res_skip_channels,
out_channels=out_channels,
temb_channels=temb_channels,
eps=resnet_eps,
groups=min(resnet_in_channels + res_skip_channels // 4, 32),
groups_out=min(out_channels // 4, 32),
dropout=dropout,
time_embedding_norm=resnet_time_scale_shift,
non_linearity=resnet_act_fn,
output_scale_factor=output_scale_factor,
pre_norm=resnet_pre_norm,
)
)
if attention_head_dim is None:
logger.warn(
f"It is not recommend to pass `attention_head_dim=None`. Defaulting `attention_head_dim` to `out_channels`: {out_channels}."
)
attention_head_dim = out_channels
self.attentions.append(
Attention(
out_channels,
heads=out_channels // attention_head_dim,
dim_head=attention_head_dim,
rescale_output_factor=output_scale_factor,
eps=resnet_eps,
norm_num_groups=32,
residual_connection=True,
bias=True,
upcast_softmax=True,
_from_deprecated_attn_block=True,
)
)
self.upsampler = FirUpsample2D(in_channels, out_channels=out_channels)
if add_upsample:
self.resnet_up = ResnetBlock2D(
in_channels=out_channels,
out_channels=out_channels,
temb_channels=temb_channels,
eps=resnet_eps,
groups=min(out_channels // 4, 32),
groups_out=min(out_channels // 4, 32),
dropout=dropout,
time_embedding_norm=resnet_time_scale_shift,
non_linearity=resnet_act_fn,
output_scale_factor=output_scale_factor,
pre_norm=resnet_pre_norm,
use_in_shortcut=True,
up=True,
kernel="fir",
)
self.skip_conv = nn.Conv2d(out_channels, 3, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
self.skip_norm = torch.nn.GroupNorm(
num_groups=min(out_channels // 4, 32), num_channels=out_channels, eps=resnet_eps, affine=True
)
self.act = nn.SiLU()
else:
self.resnet_up = None
self.skip_conv = None
self.skip_norm = None
self.act = None
self.resolution_idx = resolution_idx
def forward(
self,
hidden_states: torch.FloatTensor,
res_hidden_states_tuple: Tuple[torch.FloatTensor, ...],
temb: Optional[torch.FloatTensor] = None,
skip_sample=None,
scale: float = 1.0,
) -> Tuple[torch.FloatTensor, torch.FloatTensor]:
for resnet in self.resnets:
# pop res hidden states
res_hidden_states = res_hidden_states_tuple[-1]
res_hidden_states_tuple = res_hidden_states_tuple[:-1]
hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)
hidden_states = resnet(hidden_states, temb, scale=scale)
cross_attention_kwargs = {"scale": scale}
hidden_states = self.attentions[0](hidden_states, **cross_attention_kwargs)
if skip_sample is not None:
skip_sample = self.upsampler(skip_sample)
else:
skip_sample = 0
if self.resnet_up is not None:
skip_sample_states = self.skip_norm(hidden_states)
skip_sample_states = self.act(skip_sample_states)
skip_sample_states = self.skip_conv(skip_sample_states)
skip_sample = skip_sample + skip_sample_states
hidden_states = self.resnet_up(hidden_states, temb, scale=scale)
return hidden_states, skip_sample
class SkipUpBlock2D(nn.Module):
def __init__(
self,
in_channels: int,
prev_output_channel: int,
out_channels: int,
temb_channels: int,
resolution_idx: Optional[int] = None,
dropout: float = 0.0,
num_layers: int = 1,
resnet_eps: float = 1e-6,
resnet_time_scale_shift: str = "default",
resnet_act_fn: str = "swish",
resnet_pre_norm: bool = True,
output_scale_factor: float = np.sqrt(2.0),
add_upsample: bool = True,
upsample_padding: int = 1,
):
super().__init__()
self.resnets = nn.ModuleList([])
for i in range(num_layers):
res_skip_channels = in_channels if (i == num_layers - 1) else out_channels
resnet_in_channels = prev_output_channel if i == 0 else out_channels
self.resnets.append(
ResnetBlock2D(
in_channels=resnet_in_channels + res_skip_channels,
out_channels=out_channels,
temb_channels=temb_channels,
eps=resnet_eps,
groups=min((resnet_in_channels + res_skip_channels) // 4, 32),
groups_out=min(out_channels // 4, 32),
dropout=dropout,
time_embedding_norm=resnet_time_scale_shift,
non_linearity=resnet_act_fn,
output_scale_factor=output_scale_factor,
pre_norm=resnet_pre_norm,
)
)
self.upsampler = FirUpsample2D(in_channels, out_channels=out_channels)
if add_upsample:
self.resnet_up = ResnetBlock2D(
in_channels=out_channels,
out_channels=out_channels,
temb_channels=temb_channels,
eps=resnet_eps,
groups=min(out_channels // 4, 32),
groups_out=min(out_channels // 4, 32),
dropout=dropout,
time_embedding_norm=resnet_time_scale_shift,
non_linearity=resnet_act_fn,
output_scale_factor=output_scale_factor,
pre_norm=resnet_pre_norm,
use_in_shortcut=True,
up=True,
kernel="fir",
)
self.skip_conv = nn.Conv2d(out_channels, 3, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
self.skip_norm = torch.nn.GroupNorm(
num_groups=min(out_channels // 4, 32), num_channels=out_channels, eps=resnet_eps, affine=True
)
self.act = nn.SiLU()
else:
self.resnet_up = None
self.skip_conv = None
self.skip_norm = None
self.act = None
self.resolution_idx = resolution_idx
def forward(
self,
hidden_states: torch.FloatTensor,
res_hidden_states_tuple: Tuple[torch.FloatTensor, ...],
temb: Optional[torch.FloatTensor] = None,
skip_sample=None,
scale: float = 1.0,
) -> Tuple[torch.FloatTensor, torch.FloatTensor]:
for resnet in self.resnets:
# pop res hidden states
res_hidden_states = res_hidden_states_tuple[-1]
res_hidden_states_tuple = res_hidden_states_tuple[:-1]
hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)
hidden_states = resnet(hidden_states, temb, scale=scale)
if skip_sample is not None:
skip_sample = self.upsampler(skip_sample)
else:
skip_sample = 0
if self.resnet_up is not None:
skip_sample_states = self.skip_norm(hidden_states)
skip_sample_states = self.act(skip_sample_states)
skip_sample_states = self.skip_conv(skip_sample_states)
skip_sample = skip_sample + skip_sample_states
hidden_states = self.resnet_up(hidden_states, temb, scale=scale)
return hidden_states, skip_sample
class ResnetUpsampleBlock2D(nn.Module):
def __init__(
self,
in_channels: int,
prev_output_channel: int,
out_channels: int,
temb_channels: int,
resolution_idx: Optional[int] = None,
dropout: float = 0.0,
num_layers: int = 1,
resnet_eps: float = 1e-6,
resnet_time_scale_shift: str = "default",
resnet_act_fn: str = "swish",
resnet_groups: int = 32,
resnet_pre_norm: bool = True,
output_scale_factor: float = 1.0,
add_upsample: bool = True,
skip_time_act: bool = False,
):
super().__init__()
resnets = []
for i in range(num_layers):
res_skip_channels = in_channels if (i == num_layers - 1) else out_channels
resnet_in_channels = prev_output_channel if i == 0 else out_channels
resnets.append(
ResnetBlock2D(
in_channels=resnet_in_channels + res_skip_channels,
out_channels=out_channels,
temb_channels=temb_channels,
eps=resnet_eps,
groups=resnet_groups,
dropout=dropout,
time_embedding_norm=resnet_time_scale_shift,
non_linearity=resnet_act_fn,
output_scale_factor=output_scale_factor,
pre_norm=resnet_pre_norm,
skip_time_act=skip_time_act,
)
)
self.resnets = nn.ModuleList(resnets)
if add_upsample:
self.upsamplers = nn.ModuleList(
[
ResnetBlock2D(
in_channels=out_channels,
out_channels=out_channels,
temb_channels=temb_channels,
eps=resnet_eps,
groups=resnet_groups,
dropout=dropout,
time_embedding_norm=resnet_time_scale_shift,
non_linearity=resnet_act_fn,
output_scale_factor=output_scale_factor,
pre_norm=resnet_pre_norm,
skip_time_act=skip_time_act,
up=True,
)
]
)
else:
self.upsamplers = None
self.gradient_checkpointing = False
self.resolution_idx = resolution_idx
def forward(
self,
hidden_states: torch.FloatTensor,
res_hidden_states_tuple: Tuple[torch.FloatTensor, ...],
temb: Optional[torch.FloatTensor] = None,
upsample_size: Optional[int] = None,
scale: float = 1.0,
) -> torch.FloatTensor:
for resnet in self.resnets:
# pop res hidden states
res_hidden_states = res_hidden_states_tuple[-1]
res_hidden_states_tuple = res_hidden_states_tuple[:-1]
hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)
if self.training and self.gradient_checkpointing:
def create_custom_forward(module):
def custom_forward(*inputs):
return module(*inputs)
return custom_forward
if is_torch_version(">=", "1.11.0"):
hidden_states = torch.utils.checkpoint.checkpoint(
create_custom_forward(resnet), hidden_states, temb, use_reentrant=False
)
else:
hidden_states = torch.utils.checkpoint.checkpoint(
create_custom_forward(resnet), hidden_states, temb
)
else:
hidden_states = resnet(hidden_states, temb, scale=scale)
if self.upsamplers is not None:
for upsampler in self.upsamplers:
hidden_states = upsampler(hidden_states, temb, scale=scale)
return hidden_states
class SimpleCrossAttnUpBlock2D(nn.Module):
def __init__(
self,
in_channels: int,
out_channels: int,
prev_output_channel: int,
temb_channels: int,
resolution_idx: Optional[int] = None,
dropout: float = 0.0,
num_layers: int = 1,
resnet_eps: float = 1e-6,
resnet_time_scale_shift: str = "default",
resnet_act_fn: str = "swish",
resnet_groups: int = 32,
resnet_pre_norm: bool = True,
attention_head_dim: int = 1,
cross_attention_dim: int = 1280,
output_scale_factor: float = 1.0,
add_upsample: bool = True,
skip_time_act: bool = False,
only_cross_attention: bool = False,
cross_attention_norm: Optional[str] = None,
):
super().__init__()
resnets = []
attentions = []
self.has_cross_attention = True
self.attention_head_dim = attention_head_dim
self.num_heads = out_channels // self.attention_head_dim
for i in range(num_layers):
res_skip_channels = in_channels if (i == num_layers - 1) else out_channels
resnet_in_channels = prev_output_channel if i == 0 else out_channels
resnets.append(
ResnetBlock2D(
in_channels=resnet_in_channels + res_skip_channels,
out_channels=out_channels,
temb_channels=temb_channels,
eps=resnet_eps,
groups=resnet_groups,
dropout=dropout,
time_embedding_norm=resnet_time_scale_shift,
non_linearity=resnet_act_fn,
output_scale_factor=output_scale_factor,
pre_norm=resnet_pre_norm,
skip_time_act=skip_time_act,
)
)
processor = (
AttnAddedKVProcessor2_0() if hasattr(F, "scaled_dot_product_attention") else AttnAddedKVProcessor()
)
attentions.append(
Attention(
query_dim=out_channels,
cross_attention_dim=out_channels,
heads=self.num_heads,
dim_head=self.attention_head_dim,
added_kv_proj_dim=cross_attention_dim,
norm_num_groups=resnet_groups,
bias=True,
upcast_softmax=True,
only_cross_attention=only_cross_attention,
cross_attention_norm=cross_attention_norm,
processor=processor,
)
)
self.attentions = nn.ModuleList(attentions)
self.resnets = nn.ModuleList(resnets)
if add_upsample:
self.upsamplers = nn.ModuleList(
[
ResnetBlock2D(
in_channels=out_channels,
out_channels=out_channels,
temb_channels=temb_channels,
eps=resnet_eps,
groups=resnet_groups,
dropout=dropout,
time_embedding_norm=resnet_time_scale_shift,
non_linearity=resnet_act_fn,
output_scale_factor=output_scale_factor,
pre_norm=resnet_pre_norm,
skip_time_act=skip_time_act,
up=True,
)
]
)
else:
self.upsamplers = None
self.gradient_checkpointing = False
self.resolution_idx = resolution_idx
def forward(
self,
hidden_states: torch.FloatTensor,
res_hidden_states_tuple: Tuple[torch.FloatTensor, ...],
temb: Optional[torch.FloatTensor] = None,
encoder_hidden_states: Optional[torch.FloatTensor] = None,
upsample_size: Optional[int] = None,
attention_mask: Optional[torch.FloatTensor] = None,
cross_attention_kwargs: Optional[Dict[str, Any]] = None,
encoder_attention_mask: Optional[torch.FloatTensor] = None,
) -> torch.FloatTensor:
cross_attention_kwargs = cross_attention_kwargs if cross_attention_kwargs is not None else {}
lora_scale = cross_attention_kwargs.get("scale", 1.0)
if attention_mask is None:
# if encoder_hidden_states is defined: we are doing cross-attn, so we should use cross-attn mask.
mask = None if encoder_hidden_states is None else encoder_attention_mask
else:
# when attention_mask is defined: we don't even check for encoder_attention_mask.
# this is to maintain compatibility with UnCLIP, which uses 'attention_mask' param for cross-attn masks.
# TODO: UnCLIP should express cross-attn mask via encoder_attention_mask param instead of via attention_mask.
# then we can simplify this whole if/else block to:
# mask = attention_mask if encoder_hidden_states is None else encoder_attention_mask
mask = attention_mask
for resnet, attn in zip(self.resnets, self.attentions):
# resnet
# pop res hidden states
res_hidden_states = res_hidden_states_tuple[-1]
res_hidden_states_tuple = res_hidden_states_tuple[:-1]
hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)
if self.training and self.gradient_checkpointing:
def create_custom_forward(module, return_dict=None):
def custom_forward(*inputs):
if return_dict is not None:
return module(*inputs, return_dict=return_dict)
else:
return module(*inputs)
return custom_forward
hidden_states = torch.utils.checkpoint.checkpoint(create_custom_forward(resnet), hidden_states, temb)
hidden_states = attn(
hidden_states,
encoder_hidden_states=encoder_hidden_states,
attention_mask=mask,
**cross_attention_kwargs,
)
else:
hidden_states = resnet(hidden_states, temb, scale=lora_scale)
hidden_states = attn(
hidden_states,
encoder_hidden_states=encoder_hidden_states,
attention_mask=mask,
**cross_attention_kwargs,
)
if self.upsamplers is not None:
for upsampler in self.upsamplers:
hidden_states = upsampler(hidden_states, temb, scale=lora_scale)
return hidden_states
class KUpBlock2D(nn.Module):
def __init__(
self,
in_channels: int,
out_channels: int,
temb_channels: int,
resolution_idx: int,
dropout: float = 0.0,
num_layers: int = 5,
resnet_eps: float = 1e-5,
resnet_act_fn: str = "gelu",
resnet_group_size: Optional[int] = 32,
add_upsample: bool = True,
):
super().__init__()
resnets = []
k_in_channels = 2 * out_channels
k_out_channels = in_channels
num_layers = num_layers - 1
for i in range(num_layers):
in_channels = k_in_channels if i == 0 else out_channels
groups = in_channels // resnet_group_size
groups_out = out_channels // resnet_group_size
resnets.append(
ResnetBlock2D(
in_channels=in_channels,
out_channels=k_out_channels if (i == num_layers - 1) else out_channels,
temb_channels=temb_channels,
eps=resnet_eps,
groups=groups,
groups_out=groups_out,
dropout=dropout,
non_linearity=resnet_act_fn,
time_embedding_norm="ada_group",
conv_shortcut_bias=False,
)
)
self.resnets = nn.ModuleList(resnets)
if add_upsample:
self.upsamplers = nn.ModuleList([KUpsample2D()])
else:
self.upsamplers = None
self.gradient_checkpointing = False
self.resolution_idx = resolution_idx
def forward(
self,
hidden_states: torch.FloatTensor,
res_hidden_states_tuple: Tuple[torch.FloatTensor, ...],
temb: Optional[torch.FloatTensor] = None,
upsample_size: Optional[int] = None,
scale: float = 1.0,
) -> torch.FloatTensor:
res_hidden_states_tuple = res_hidden_states_tuple[-1]
if res_hidden_states_tuple is not None:
hidden_states = torch.cat([hidden_states, res_hidden_states_tuple], dim=1)
for resnet in self.resnets:
if self.training and self.gradient_checkpointing:
def create_custom_forward(module):
def custom_forward(*inputs):
return module(*inputs)
return custom_forward
if is_torch_version(">=", "1.11.0"):
hidden_states = torch.utils.checkpoint.checkpoint(
create_custom_forward(resnet), hidden_states, temb, use_reentrant=False
)
else:
hidden_states = torch.utils.checkpoint.checkpoint(
create_custom_forward(resnet), hidden_states, temb
)
else:
hidden_states = resnet(hidden_states, temb, scale=scale)
if self.upsamplers is not None:
for upsampler in self.upsamplers:
hidden_states = upsampler(hidden_states)
return hidden_states
class KCrossAttnUpBlock2D(nn.Module):
def __init__(
self,
in_channels: int,
out_channels: int,
temb_channels: int,
resolution_idx: int,
dropout: float = 0.0,
num_layers: int = 4,
resnet_eps: float = 1e-5,
resnet_act_fn: str = "gelu",
resnet_group_size: int = 32,
attention_head_dim: int = 1, # attention dim_head
cross_attention_dim: int = 768,
add_upsample: bool = True,
upcast_attention: bool = False,
):
super().__init__()
resnets = []
attentions = []
is_first_block = in_channels == out_channels == temb_channels
is_middle_block = in_channels != out_channels
add_self_attention = True if is_first_block else False
self.has_cross_attention = True
self.attention_head_dim = attention_head_dim
# in_channels, and out_channels for the block (k-unet)
k_in_channels = out_channels if is_first_block else 2 * out_channels
k_out_channels = in_channels
num_layers = num_layers - 1
for i in range(num_layers):
in_channels = k_in_channels if i == 0 else out_channels
groups = in_channels // resnet_group_size
groups_out = out_channels // resnet_group_size
if is_middle_block and (i == num_layers - 1):
conv_2d_out_channels = k_out_channels
else:
conv_2d_out_channels = None
resnets.append(
ResnetBlock2D(
in_channels=in_channels,
out_channels=out_channels,
conv_2d_out_channels=conv_2d_out_channels,
temb_channels=temb_channels,
eps=resnet_eps,
groups=groups,
groups_out=groups_out,
dropout=dropout,
non_linearity=resnet_act_fn,
time_embedding_norm="ada_group",
conv_shortcut_bias=False,
)
)
attentions.append(
KAttentionBlock(
k_out_channels if (i == num_layers - 1) else out_channels,
k_out_channels // attention_head_dim
if (i == num_layers - 1)
else out_channels // attention_head_dim,
attention_head_dim,
cross_attention_dim=cross_attention_dim,
temb_channels=temb_channels,
attention_bias=True,
add_self_attention=add_self_attention,
cross_attention_norm="layer_norm",
upcast_attention=upcast_attention,
)
)
self.resnets = nn.ModuleList(resnets)
self.attentions = nn.ModuleList(attentions)
if add_upsample:
self.upsamplers = nn.ModuleList([KUpsample2D()])
else:
self.upsamplers = None
self.gradient_checkpointing = False
self.resolution_idx = resolution_idx
def forward(
self,
hidden_states: torch.FloatTensor,
res_hidden_states_tuple: Tuple[torch.FloatTensor, ...],
temb: Optional[torch.FloatTensor] = None,
encoder_hidden_states: Optional[torch.FloatTensor] = None,
cross_attention_kwargs: Optional[Dict[str, Any]] = None,
upsample_size: Optional[int] = None,
attention_mask: Optional[torch.FloatTensor] = None,
encoder_attention_mask: Optional[torch.FloatTensor] = None,
) -> torch.FloatTensor:
res_hidden_states_tuple = res_hidden_states_tuple[-1]
if res_hidden_states_tuple is not None:
hidden_states = torch.cat([hidden_states, res_hidden_states_tuple], dim=1)
lora_scale = cross_attention_kwargs.get("scale", 1.0) if cross_attention_kwargs is not None else 1.0
for resnet, attn in zip(self.resnets, self.attentions):
if self.training and self.gradient_checkpointing:
def create_custom_forward(module, return_dict=None):
def custom_forward(*inputs):
if return_dict is not None:
return module(*inputs, return_dict=return_dict)
else:
return module(*inputs)
return custom_forward
ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
hidden_states = torch.utils.checkpoint.checkpoint(
create_custom_forward(resnet),
hidden_states,
temb,
**ckpt_kwargs,
)
hidden_states = attn(
hidden_states,
encoder_hidden_states=encoder_hidden_states,
emb=temb,
attention_mask=attention_mask,
cross_attention_kwargs=cross_attention_kwargs,
encoder_attention_mask=encoder_attention_mask,
)
else:
hidden_states = resnet(hidden_states, temb, scale=lora_scale)
hidden_states = attn(
hidden_states,
encoder_hidden_states=encoder_hidden_states,
emb=temb,
attention_mask=attention_mask,
cross_attention_kwargs=cross_attention_kwargs,
encoder_attention_mask=encoder_attention_mask,
)
if self.upsamplers is not None:
for upsampler in self.upsamplers:
hidden_states = upsampler(hidden_states)
return hidden_states
# can potentially later be renamed to `No-feed-forward` attention
class KAttentionBlock(nn.Module):
r"""
A basic Transformer block.
Parameters:
dim (`int`): The number of channels in the input and output.
num_attention_heads (`int`): The number of heads to use for multi-head attention.
attention_head_dim (`int`): The number of channels in each head.
dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
cross_attention_dim (`int`, *optional*): The size of the encoder_hidden_states vector for cross attention.
attention_bias (`bool`, *optional*, defaults to `False`):
Configure if the attention layers should contain a bias parameter.
upcast_attention (`bool`, *optional*, defaults to `False`):
Set to `True` to upcast the attention computation to `float32`.
temb_channels (`int`, *optional*, defaults to 768):
The number of channels in the token embedding.
add_self_attention (`bool`, *optional*, defaults to `False`):
Set to `True` to add self-attention to the block.
cross_attention_norm (`str`, *optional*, defaults to `None`):
The type of normalization to use for the cross attention. Can be `None`, `layer_norm`, or `group_norm`.
group_size (`int`, *optional*, defaults to 32):
The number of groups to separate the channels into for group normalization.
"""
def __init__(
self,
dim: int,
num_attention_heads: int,
attention_head_dim: int,
dropout: float = 0.0,
cross_attention_dim: Optional[int] = None,
attention_bias: bool = False,
upcast_attention: bool = False,
temb_channels: int = 768, # for ada_group_norm
add_self_attention: bool = False,
cross_attention_norm: Optional[str] = None,
group_size: int = 32,
):
super().__init__()
self.add_self_attention = add_self_attention
# 1. Self-Attn
if add_self_attention:
self.norm1 = AdaGroupNorm(temb_channels, dim, max(1, dim // group_size))
self.attn1 = Attention(
query_dim=dim,
heads=num_attention_heads,
dim_head=attention_head_dim,
dropout=dropout,
bias=attention_bias,
cross_attention_dim=None,
cross_attention_norm=None,
)
# 2. Cross-Attn
self.norm2 = AdaGroupNorm(temb_channels, dim, max(1, dim // group_size))
self.attn2 = Attention(
query_dim=dim,
cross_attention_dim=cross_attention_dim,
heads=num_attention_heads,
dim_head=attention_head_dim,
dropout=dropout,
bias=attention_bias,
upcast_attention=upcast_attention,
cross_attention_norm=cross_attention_norm,
)
def _to_3d(self, hidden_states: torch.FloatTensor, height: int, weight: int) -> torch.FloatTensor:
return hidden_states.permute(0, 2, 3, 1).reshape(hidden_states.shape[0], height * weight, -1)
def _to_4d(self, hidden_states: torch.FloatTensor, height: int, weight: int) -> torch.FloatTensor:
return hidden_states.permute(0, 2, 1).reshape(hidden_states.shape[0], -1, height, weight)
def forward(
self,
hidden_states: torch.FloatTensor,
encoder_hidden_states: Optional[torch.FloatTensor] = None,
# TODO: mark emb as non-optional (self.norm2 requires it).
# requires assessing impact of change to positional param interface.
emb: Optional[torch.FloatTensor] = None,
attention_mask: Optional[torch.FloatTensor] = None,
cross_attention_kwargs: Optional[Dict[str, Any]] = None,
encoder_attention_mask: Optional[torch.FloatTensor] = None,
) -> torch.FloatTensor:
cross_attention_kwargs = cross_attention_kwargs if cross_attention_kwargs is not None else {}
# 1. Self-Attention
if self.add_self_attention:
norm_hidden_states = self.norm1(hidden_states, emb)
height, weight = norm_hidden_states.shape[2:]
norm_hidden_states = self._to_3d(norm_hidden_states, height, weight)
attn_output = self.attn1(
norm_hidden_states,
encoder_hidden_states=None,
attention_mask=attention_mask,
**cross_attention_kwargs,
)
attn_output = self._to_4d(attn_output, height, weight)
hidden_states = attn_output + hidden_states
# 2. Cross-Attention/None
norm_hidden_states = self.norm2(hidden_states, emb)
height, weight = norm_hidden_states.shape[2:]
norm_hidden_states = self._to_3d(norm_hidden_states, height, weight)
attn_output = self.attn2(
norm_hidden_states,
encoder_hidden_states=encoder_hidden_states,
attention_mask=attention_mask if encoder_hidden_states is None else encoder_attention_mask,
**cross_attention_kwargs,
)
attn_output = self._to_4d(attn_output, height, weight)
hidden_states = attn_output + hidden_states
return hidden_states
ootd/pipelines_ootd/unet_garm_2d_condition.py 0 → 100644
View file @ 54a066bf
# Copyright 2023 The HuggingFace Team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# Modified by Yuhao Xu for OOTDiffusion (https://github.com/levihsu/OOTDiffusion)
from dataclasses import dataclass
from typing import Any, Dict, List, Optional, Tuple, Union
import torch
import torch.nn as nn
import torch.utils.checkpoint
from .unet_garm_2d_blocks import (
UNetMidBlock2D,
UNetMidBlock2DCrossAttn,
UNetMidBlock2DSimpleCrossAttn,
get_down_block,
get_up_block,
)
from diffusers.configuration_utils import ConfigMixin, register_to_config
from diffusers.loaders import UNet2DConditionLoadersMixin
from diffusers.utils import USE_PEFT_BACKEND, BaseOutput, deprecate, logging, scale_lora_layers, unscale_lora_layers
from diffusers.models.activations import get_activation
from diffusers.models.attention_processor import (
ADDED_KV_ATTENTION_PROCESSORS,
CROSS_ATTENTION_PROCESSORS,
AttentionProcessor,
AttnAddedKVProcessor,
AttnProcessor,
)
from diffusers.models.embeddings import (
GaussianFourierProjection,
ImageHintTimeEmbedding,
ImageProjection,
ImageTimeEmbedding,
PositionNet,
TextImageProjection,
TextImageTimeEmbedding,
TextTimeEmbedding,
TimestepEmbedding,
Timesteps,
)
from diffusers.models.modeling_utils import ModelMixin
# from diffusers.models.unet_2d_blocks import (
# UNetMidBlock2D,
# UNetMidBlock2DCrossAttn,
# UNetMidBlock2DSimpleCrossAttn,
# get_down_block,
# get_up_block,
# )
logger = logging.get_logger(__name__) # pylint: disable=invalid-name
@dataclass
class UNet2DConditionOutput(BaseOutput):
"""
The output of [`UNet2DConditionModel`].
Args:
sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`):
The hidden states output conditioned on `encoder_hidden_states` input. Output of last layer of model.
"""
sample: torch.FloatTensor = None
class UNetGarm2DConditionModel(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin):
r"""
A conditional 2D UNet model that takes a noisy sample, conditional state, and a timestep and returns a sample
shaped output.
This model inherits from [`ModelMixin`]. Check the superclass documentation for it's generic methods implemented
for all models (such as downloading or saving).
Parameters:
sample_size (`int` or `Tuple[int, int]`, *optional*, defaults to `None`):
Height and width of input/output sample.
in_channels (`int`, *optional*, defaults to 4): Number of channels in the input sample.
out_channels (`int`, *optional*, defaults to 4): Number of channels in the output.
center_input_sample (`bool`, *optional*, defaults to `False`): Whether to center the input sample.
flip_sin_to_cos (`bool`, *optional*, defaults to `False`):
Whether to flip the sin to cos in the time embedding.
freq_shift (`int`, *optional*, defaults to 0): The frequency shift to apply to the time embedding.
down_block_types (`Tuple[str]`, *optional*, defaults to `("CrossAttnDownBlock2D", "CrossAttnDownBlock2D", "CrossAttnDownBlock2D", "DownBlock2D")`):
The tuple of downsample blocks to use.
mid_block_type (`str`, *optional*, defaults to `"UNetMidBlock2DCrossAttn"`):
Block type for middle of UNet, it can be one of `UNetMidBlock2DCrossAttn`, `UNetMidBlock2D`, or
`UNetMidBlock2DSimpleCrossAttn`. If `None`, the mid block layer is skipped.
up_block_types (`Tuple[str]`, *optional*, defaults to `("UpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D")`):
The tuple of upsample blocks to use.
only_cross_attention(`bool` or `Tuple[bool]`, *optional*, default to `False`):
Whether to include self-attention in the basic transformer blocks, see
[`~models.attention.BasicTransformerBlock`].
block_out_channels (`Tuple[int]`, *optional*, defaults to `(320, 640, 1280, 1280)`):
The tuple of output channels for each block.
layers_per_block (`int`, *optional*, defaults to 2): The number of layers per block.
downsample_padding (`int`, *optional*, defaults to 1): The padding to use for the downsampling convolution.
mid_block_scale_factor (`float`, *optional*, defaults to 1.0): The scale factor to use for the mid block.
dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
act_fn (`str`, *optional*, defaults to `"silu"`): The activation function to use.
norm_num_groups (`int`, *optional*, defaults to 32): The number of groups to use for the normalization.
If `None`, normalization and activation layers is skipped in post-processing.
norm_eps (`float`, *optional*, defaults to 1e-5): The epsilon to use for the normalization.
cross_attention_dim (`int` or `Tuple[int]`, *optional*, defaults to 1280):
The dimension of the cross attention features.
transformer_layers_per_block (`int`, `Tuple[int]`, or `Tuple[Tuple]` , *optional*, defaults to 1):
The number of transformer blocks of type [`~models.attention.BasicTransformerBlock`]. Only relevant for
[`~models.unet_2d_blocks.CrossAttnDownBlock2D`], [`~models.unet_2d_blocks.CrossAttnUpBlock2D`],
[`~models.unet_2d_blocks.UNetMidBlock2DCrossAttn`].
reverse_transformer_layers_per_block : (`Tuple[Tuple]`, *optional*, defaults to None):
The number of transformer blocks of type [`~models.attention.BasicTransformerBlock`], in the upsampling
blocks of the U-Net. Only relevant if `transformer_layers_per_block` is of type `Tuple[Tuple]` and for
[`~models.unet_2d_blocks.CrossAttnDownBlock2D`], [`~models.unet_2d_blocks.CrossAttnUpBlock2D`],
[`~models.unet_2d_blocks.UNetMidBlock2DCrossAttn`].
encoder_hid_dim (`int`, *optional*, defaults to None):
If `encoder_hid_dim_type` is defined, `encoder_hidden_states` will be projected from `encoder_hid_dim`
dimension to `cross_attention_dim`.
encoder_hid_dim_type (`str`, *optional*, defaults to `None`):
If given, the `encoder_hidden_states` and potentially other embeddings are down-projected to text
embeddings of dimension `cross_attention` according to `encoder_hid_dim_type`.
attention_head_dim (`int`, *optional*, defaults to 8): The dimension of the attention heads.
num_attention_heads (`int`, *optional*):
The number of attention heads. If not defined, defaults to `attention_head_dim`
resnet_time_scale_shift (`str`, *optional*, defaults to `"default"`): Time scale shift config
for ResNet blocks (see [`~models.resnet.ResnetBlock2D`]). Choose from `default` or `scale_shift`.
class_embed_type (`str`, *optional*, defaults to `None`):
The type of class embedding to use which is ultimately summed with the time embeddings. Choose from `None`,
`"timestep"`, `"identity"`, `"projection"`, or `"simple_projection"`.
addition_embed_type (`str`, *optional*, defaults to `None`):
Configures an optional embedding which will be summed with the time embeddings. Choose from `None` or
"text". "text" will use the `TextTimeEmbedding` layer.
addition_time_embed_dim: (`int`, *optional*, defaults to `None`):
Dimension for the timestep embeddings.
num_class_embeds (`int`, *optional*, defaults to `None`):
Input dimension of the learnable embedding matrix to be projected to `time_embed_dim`, when performing
class conditioning with `class_embed_type` equal to `None`.
time_embedding_type (`str`, *optional*, defaults to `positional`):
The type of position embedding to use for timesteps. Choose from `positional` or `fourier`.
time_embedding_dim (`int`, *optional*, defaults to `None`):
An optional override for the dimension of the projected time embedding.
time_embedding_act_fn (`str`, *optional*, defaults to `None`):
Optional activation function to use only once on the time embeddings before they are passed to the rest of
the UNet. Choose from `silu`, `mish`, `gelu`, and `swish`.
timestep_post_act (`str`, *optional*, defaults to `None`):
The second activation function to use in timestep embedding. Choose from `silu`, `mish` and `gelu`.
time_cond_proj_dim (`int`, *optional*, defaults to `None`):
The dimension of `cond_proj` layer in the timestep embedding.
conv_in_kernel (`int`, *optional*, default to `3`): The kernel size of `conv_in` layer. conv_out_kernel (`int`,
*optional*, default to `3`): The kernel size of `conv_out` layer. projection_class_embeddings_input_dim (`int`,
*optional*): The dimension of the `class_labels` input when
`class_embed_type="projection"`. Required when `class_embed_type="projection"`.
class_embeddings_concat (`bool`, *optional*, defaults to `False`): Whether to concatenate the time
embeddings with the class embeddings.
mid_block_only_cross_attention (`bool`, *optional*, defaults to `None`):
Whether to use cross attention with the mid block when using the `UNetMidBlock2DSimpleCrossAttn`. If
`only_cross_attention` is given as a single boolean and `mid_block_only_cross_attention` is `None`, the
`only_cross_attention` value is used as the value for `mid_block_only_cross_attention`. Default to `False`
otherwise.
"""
_supports_gradient_checkpointing = True
@register_to_config
def __init__(
self,
sample_size: Optional[int] = None,
in_channels: int = 4,
out_channels: int = 4,
center_input_sample: bool = False,
flip_sin_to_cos: bool = True,
freq_shift: int = 0,
down_block_types: Tuple[str] = (
"CrossAttnDownBlock2D",
"CrossAttnDownBlock2D",
"CrossAttnDownBlock2D",
"DownBlock2D",
),
mid_block_type: Optional[str] = "UNetMidBlock2DCrossAttn",
up_block_types: Tuple[str] = ("UpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D"),
only_cross_attention: Union[bool, Tuple[bool]] = False,
block_out_channels: Tuple[int] = (320, 640, 1280, 1280),
layers_per_block: Union[int, Tuple[int]] = 2,
downsample_padding: int = 1,
mid_block_scale_factor: float = 1,
dropout: float = 0.0,
act_fn: str = "silu",
norm_num_groups: Optional[int] = 32,
norm_eps: float = 1e-5,
cross_attention_dim: Union[int, Tuple[int]] = 1280,
transformer_layers_per_block: Union[int, Tuple[int], Tuple[Tuple]] = 1,
reverse_transformer_layers_per_block: Optional[Tuple[Tuple[int]]] = None,
encoder_hid_dim: Optional[int] = None,
encoder_hid_dim_type: Optional[str] = None,
attention_head_dim: Union[int, Tuple[int]] = 8,
num_attention_heads: Optional[Union[int, Tuple[int]]] = None,
dual_cross_attention: bool = False,
use_linear_projection: bool = False,
class_embed_type: Optional[str] = None,
addition_embed_type: Optional[str] = None,
addition_time_embed_dim: Optional[int] = None,
num_class_embeds: Optional[int] = None,
upcast_attention: bool = False,
resnet_time_scale_shift: str = "default",
resnet_skip_time_act: bool = False,
resnet_out_scale_factor: int = 1.0,
time_embedding_type: str = "positional",
time_embedding_dim: Optional[int] = None,
time_embedding_act_fn: Optional[str] = None,
timestep_post_act: Optional[str] = None,
time_cond_proj_dim: Optional[int] = None,
conv_in_kernel: int = 3,
conv_out_kernel: int = 3,
projection_class_embeddings_input_dim: Optional[int] = None,
attention_type: str = "default",
class_embeddings_concat: bool = False,
mid_block_only_cross_attention: Optional[bool] = None,
cross_attention_norm: Optional[str] = None,
addition_embed_type_num_heads=64,
):
super().__init__()
self.sample_size = sample_size
if num_attention_heads is not None:
raise ValueError(
"At the moment it is not possible to define the number of attention heads via `num_attention_heads` because of a naming issue as described in https://github.com/huggingface/diffusers/issues/2011#issuecomment-1547958131. Passing `num_attention_heads` will only be supported in diffusers v0.19."
)
# If `num_attention_heads` is not defined (which is the case for most models)
# it will default to `attention_head_dim`. This looks weird upon first reading it and it is.
# The reason for this behavior is to correct for incorrectly named variables that were introduced
# when this library was created. The incorrect naming was only discovered much later in https://github.com/huggingface/diffusers/issues/2011#issuecomment-1547958131
# Changing `attention_head_dim` to `num_attention_heads` for 40,000+ configurations is too backwards breaking
# which is why we correct for the naming here.
num_attention_heads = num_attention_heads or attention_head_dim
# Check inputs
if len(down_block_types) != len(up_block_types):
raise ValueError(
f"Must provide the same number of `down_block_types` as `up_block_types`. `down_block_types`: {down_block_types}. `up_block_types`: {up_block_types}."
)
if len(block_out_channels) != len(down_block_types):
raise ValueError(
f"Must provide the same number of `block_out_channels` as `down_block_types`. `block_out_channels`: {block_out_channels}. `down_block_types`: {down_block_types}."
)
if not isinstance(only_cross_attention, bool) and len(only_cross_attention) != len(down_block_types):
raise ValueError(
f"Must provide the same number of `only_cross_attention` as `down_block_types`. `only_cross_attention`: {only_cross_attention}. `down_block_types`: {down_block_types}."
)
if not isinstance(num_attention_heads, int) and len(num_attention_heads) != len(down_block_types):
raise ValueError(
f"Must provide the same number of `num_attention_heads` as `down_block_types`. `num_attention_heads`: {num_attention_heads}. `down_block_types`: {down_block_types}."
)
if not isinstance(attention_head_dim, int) and len(attention_head_dim) != len(down_block_types):
raise ValueError(
f"Must provide the same number of `attention_head_dim` as `down_block_types`. `attention_head_dim`: {attention_head_dim}. `down_block_types`: {down_block_types}."
)
if isinstance(cross_attention_dim, list) and len(cross_attention_dim) != len(down_block_types):
raise ValueError(
f"Must provide the same number of `cross_attention_dim` as `down_block_types`. `cross_attention_dim`: {cross_attention_dim}. `down_block_types`: {down_block_types}."
)
if not isinstance(layers_per_block, int) and len(layers_per_block) != len(down_block_types):
raise ValueError(
f"Must provide the same number of `layers_per_block` as `down_block_types`. `layers_per_block`: {layers_per_block}. `down_block_types`: {down_block_types}."
)
if isinstance(transformer_layers_per_block, list) and reverse_transformer_layers_per_block is None:
for layer_number_per_block in transformer_layers_per_block:
if isinstance(layer_number_per_block, list):
raise ValueError("Must provide 'reverse_transformer_layers_per_block` if using asymmetrical UNet.")
# input
conv_in_padding = (conv_in_kernel - 1) // 2
self.conv_in = nn.Conv2d(
in_channels, block_out_channels[0], kernel_size=conv_in_kernel, padding=conv_in_padding
)
# time
if time_embedding_type == "fourier":
time_embed_dim = time_embedding_dim or block_out_channels[0] * 2
if time_embed_dim % 2 != 0:
raise ValueError(f"`time_embed_dim` should be divisible by 2, but is {time_embed_dim}.")
self.time_proj = GaussianFourierProjection(
time_embed_dim // 2, set_W_to_weight=False, log=False, flip_sin_to_cos=flip_sin_to_cos
)
timestep_input_dim = time_embed_dim
elif time_embedding_type == "positional":
time_embed_dim = time_embedding_dim or block_out_channels[0] * 4
self.time_proj = Timesteps(block_out_channels[0], flip_sin_to_cos, freq_shift)
timestep_input_dim = block_out_channels[0]
else:
raise ValueError(
f"{time_embedding_type} does not exist. Please make sure to use one of `fourier` or `positional`."
)
self.time_embedding = TimestepEmbedding(
timestep_input_dim,
time_embed_dim,
act_fn=act_fn,
post_act_fn=timestep_post_act,
cond_proj_dim=time_cond_proj_dim,
)
if encoder_hid_dim_type is None and encoder_hid_dim is not None:
encoder_hid_dim_type = "text_proj"
self.register_to_config(encoder_hid_dim_type=encoder_hid_dim_type)
logger.info("encoder_hid_dim_type defaults to 'text_proj' as `encoder_hid_dim` is defined.")
if encoder_hid_dim is None and encoder_hid_dim_type is not None:
raise ValueError(
f"`encoder_hid_dim` has to be defined when `encoder_hid_dim_type` is set to {encoder_hid_dim_type}."
)
if encoder_hid_dim_type == "text_proj":
self.encoder_hid_proj = nn.Linear(encoder_hid_dim, cross_attention_dim)
elif encoder_hid_dim_type == "text_image_proj":
# image_embed_dim DOESN'T have to be `cross_attention_dim`. To not clutter the __init__ too much
# they are set to `cross_attention_dim` here as this is exactly the required dimension for the currently only use
# case when `addition_embed_type == "text_image_proj"` (Kadinsky 2.1)`
self.encoder_hid_proj = TextImageProjection(
text_embed_dim=encoder_hid_dim,
image_embed_dim=cross_attention_dim,
cross_attention_dim=cross_attention_dim,
)
elif encoder_hid_dim_type == "image_proj":
# Kandinsky 2.2
self.encoder_hid_proj = ImageProjection(
image_embed_dim=encoder_hid_dim,
cross_attention_dim=cross_attention_dim,
)
elif encoder_hid_dim_type is not None:
raise ValueError(
f"encoder_hid_dim_type: {encoder_hid_dim_type} must be None, 'text_proj' or 'text_image_proj'."
)
else:
self.encoder_hid_proj = None
# class embedding
if class_embed_type is None and num_class_embeds is not None:
self.class_embedding = nn.Embedding(num_class_embeds, time_embed_dim)
elif class_embed_type == "timestep":
self.class_embedding = TimestepEmbedding(timestep_input_dim, time_embed_dim, act_fn=act_fn)
elif class_embed_type == "identity":
self.class_embedding = nn.Identity(time_embed_dim, time_embed_dim)
elif class_embed_type == "projection":
if projection_class_embeddings_input_dim is None:
raise ValueError(
"`class_embed_type`: 'projection' requires `projection_class_embeddings_input_dim` be set"
)
# The projection `class_embed_type` is the same as the timestep `class_embed_type` except
# 1. the `class_labels` inputs are not first converted to sinusoidal embeddings
# 2. it projects from an arbitrary input dimension.
#
# Note that `TimestepEmbedding` is quite general, being mainly linear layers and activations.
# When used for embedding actual timesteps, the timesteps are first converted to sinusoidal embeddings.
# As a result, `TimestepEmbedding` can be passed arbitrary vectors.
self.class_embedding = TimestepEmbedding(projection_class_embeddings_input_dim, time_embed_dim)
elif class_embed_type == "simple_projection":
if projection_class_embeddings_input_dim is None:
raise ValueError(
"`class_embed_type`: 'simple_projection' requires `projection_class_embeddings_input_dim` be set"
)
self.class_embedding = nn.Linear(projection_class_embeddings_input_dim, time_embed_dim)
else:
self.class_embedding = None
if addition_embed_type == "text":
if encoder_hid_dim is not None:
text_time_embedding_from_dim = encoder_hid_dim
else:
text_time_embedding_from_dim = cross_attention_dim
self.add_embedding = TextTimeEmbedding(
text_time_embedding_from_dim, time_embed_dim, num_heads=addition_embed_type_num_heads
)
elif addition_embed_type == "text_image":
# text_embed_dim and image_embed_dim DON'T have to be `cross_attention_dim`. To not clutter the __init__ too much
# they are set to `cross_attention_dim` here as this is exactly the required dimension for the currently only use
# case when `addition_embed_type == "text_image"` (Kadinsky 2.1)`
self.add_embedding = TextImageTimeEmbedding(
text_embed_dim=cross_attention_dim, image_embed_dim=cross_attention_dim, time_embed_dim=time_embed_dim
)
elif addition_embed_type == "text_time":
self.add_time_proj = Timesteps(addition_time_embed_dim, flip_sin_to_cos, freq_shift)
self.add_embedding = TimestepEmbedding(projection_class_embeddings_input_dim, time_embed_dim)
elif addition_embed_type == "image":
# Kandinsky 2.2
self.add_embedding = ImageTimeEmbedding(image_embed_dim=encoder_hid_dim, time_embed_dim=time_embed_dim)
elif addition_embed_type == "image_hint":
# Kandinsky 2.2 ControlNet
self.add_embedding = ImageHintTimeEmbedding(image_embed_dim=encoder_hid_dim, time_embed_dim=time_embed_dim)
elif addition_embed_type is not None:
raise ValueError(f"addition_embed_type: {addition_embed_type} must be None, 'text' or 'text_image'.")
if time_embedding_act_fn is None:
self.time_embed_act = None
else:
self.time_embed_act = get_activation(time_embedding_act_fn)
self.down_blocks = nn.ModuleList([])
self.up_blocks = nn.ModuleList([])
if isinstance(only_cross_attention, bool):
if mid_block_only_cross_attention is None:
mid_block_only_cross_attention = only_cross_attention
only_cross_attention = [only_cross_attention] * len(down_block_types)
if mid_block_only_cross_attention is None:
mid_block_only_cross_attention = False
if isinstance(num_attention_heads, int):
num_attention_heads = (num_attention_heads,) * len(down_block_types)
if isinstance(attention_head_dim, int):
attention_head_dim = (attention_head_dim,) * len(down_block_types)
if isinstance(cross_attention_dim, int):
cross_attention_dim = (cross_attention_dim,) * len(down_block_types)
if isinstance(layers_per_block, int):
layers_per_block = [layers_per_block] * len(down_block_types)
if isinstance(transformer_layers_per_block, int):
transformer_layers_per_block = [transformer_layers_per_block] * len(down_block_types)
if class_embeddings_concat:
# The time embeddings are concatenated with the class embeddings. The dimension of the
# time embeddings passed to the down, middle, and up blocks is twice the dimension of the
# regular time embeddings
blocks_time_embed_dim = time_embed_dim * 2
else:
blocks_time_embed_dim = time_embed_dim
# down
output_channel = block_out_channels[0]
for i, down_block_type in enumerate(down_block_types):
input_channel = output_channel
output_channel = block_out_channels[i]
is_final_block = i == len(block_out_channels) - 1
down_block = get_down_block(
down_block_type,
num_layers=layers_per_block[i],
transformer_layers_per_block=transformer_layers_per_block[i],
in_channels=input_channel,
out_channels=output_channel,
temb_channels=blocks_time_embed_dim,
add_downsample=not is_final_block,
resnet_eps=norm_eps,
resnet_act_fn=act_fn,
resnet_groups=norm_num_groups,
cross_attention_dim=cross_attention_dim[i],
num_attention_heads=num_attention_heads[i],
downsample_padding=downsample_padding,
dual_cross_attention=dual_cross_attention,
use_linear_projection=use_linear_projection,
only_cross_attention=only_cross_attention[i],
upcast_attention=upcast_attention,
resnet_time_scale_shift=resnet_time_scale_shift,
attention_type=attention_type,
resnet_skip_time_act=resnet_skip_time_act,
resnet_out_scale_factor=resnet_out_scale_factor,
cross_attention_norm=cross_attention_norm,
attention_head_dim=attention_head_dim[i] if attention_head_dim[i] is not None else output_channel,
dropout=dropout,
)
self.down_blocks.append(down_block)
# mid
if mid_block_type == "UNetMidBlock2DCrossAttn":
self.mid_block = UNetMidBlock2DCrossAttn(
transformer_layers_per_block=transformer_layers_per_block[-1],
in_channels=block_out_channels[-1],
temb_channels=blocks_time_embed_dim,
dropout=dropout,
resnet_eps=norm_eps,
resnet_act_fn=act_fn,
output_scale_factor=mid_block_scale_factor,
resnet_time_scale_shift=resnet_time_scale_shift,
cross_attention_dim=cross_attention_dim[-1],
num_attention_heads=num_attention_heads[-1],
resnet_groups=norm_num_groups,
dual_cross_attention=dual_cross_attention,
use_linear_projection=use_linear_projection,
upcast_attention=upcast_attention,
attention_type=attention_type,
)
elif mid_block_type == "UNetMidBlock2DSimpleCrossAttn":
self.mid_block = UNetMidBlock2DSimpleCrossAttn(
in_channels=block_out_channels[-1],
temb_channels=blocks_time_embed_dim,
dropout=dropout,
resnet_eps=norm_eps,
resnet_act_fn=act_fn,
output_scale_factor=mid_block_scale_factor,
cross_attention_dim=cross_attention_dim[-1],
attention_head_dim=attention_head_dim[-1],
resnet_groups=norm_num_groups,
resnet_time_scale_shift=resnet_time_scale_shift,
skip_time_act=resnet_skip_time_act,
only_cross_attention=mid_block_only_cross_attention,
cross_attention_norm=cross_attention_norm,
)
elif mid_block_type == "UNetMidBlock2D":
self.mid_block = UNetMidBlock2D(
in_channels=block_out_channels[-1],
temb_channels=blocks_time_embed_dim,
dropout=dropout,
num_layers=0,
resnet_eps=norm_eps,
resnet_act_fn=act_fn,
output_scale_factor=mid_block_scale_factor,
resnet_groups=norm_num_groups,
resnet_time_scale_shift=resnet_time_scale_shift,
add_attention=False,
)
elif mid_block_type is None:
self.mid_block = None
else:
raise ValueError(f"unknown mid_block_type : {mid_block_type}")
# count how many layers upsample the images
self.num_upsamplers = 0
# up
reversed_block_out_channels = list(reversed(block_out_channels))
reversed_num_attention_heads = list(reversed(num_attention_heads))
reversed_layers_per_block = list(reversed(layers_per_block))
reversed_cross_attention_dim = list(reversed(cross_attention_dim))
reversed_transformer_layers_per_block = (
list(reversed(transformer_layers_per_block))
if reverse_transformer_layers_per_block is None
else reverse_transformer_layers_per_block
)
only_cross_attention = list(reversed(only_cross_attention))
output_channel = reversed_block_out_channels[0]
for i, up_block_type in enumerate(up_block_types):
is_final_block = i == len(block_out_channels) - 1
prev_output_channel = output_channel
output_channel = reversed_block_out_channels[i]
input_channel = reversed_block_out_channels[min(i + 1, len(block_out_channels) - 1)]
# add upsample block for all BUT final layer
if not is_final_block:
add_upsample = True
self.num_upsamplers += 1
else:
add_upsample = False
up_block = get_up_block(
up_block_type,
num_layers=reversed_layers_per_block[i] + 1,
transformer_layers_per_block=reversed_transformer_layers_per_block[i],
in_channels=input_channel,
out_channels=output_channel,
prev_output_channel=prev_output_channel,
temb_channels=blocks_time_embed_dim,
add_upsample=add_upsample,
resnet_eps=norm_eps,
resnet_act_fn=act_fn,
resolution_idx=i,
resnet_groups=norm_num_groups,
cross_attention_dim=reversed_cross_attention_dim[i],
num_attention_heads=reversed_num_attention_heads[i],
dual_cross_attention=dual_cross_attention,
use_linear_projection=use_linear_projection,
only_cross_attention=only_cross_attention[i],
upcast_attention=upcast_attention,
resnet_time_scale_shift=resnet_time_scale_shift,
attention_type=attention_type,
resnet_skip_time_act=resnet_skip_time_act,
resnet_out_scale_factor=resnet_out_scale_factor,
cross_attention_norm=cross_attention_norm,
attention_head_dim=attention_head_dim[i] if attention_head_dim[i] is not None else output_channel,
dropout=dropout,
)
self.up_blocks.append(up_block)
prev_output_channel = output_channel
# out
if norm_num_groups is not None:
self.conv_norm_out = nn.GroupNorm(
num_channels=block_out_channels[0], num_groups=norm_num_groups, eps=norm_eps
)
self.conv_act = get_activation(act_fn)
else:
self.conv_norm_out = None
self.conv_act = None
conv_out_padding = (conv_out_kernel - 1) // 2
self.conv_out = nn.Conv2d(
block_out_channels[0], out_channels, kernel_size=conv_out_kernel, padding=conv_out_padding
)
if attention_type in ["gated", "gated-text-image"]:
positive_len = 768
if isinstance(cross_attention_dim, int):
positive_len = cross_attention_dim
elif isinstance(cross_attention_dim, tuple) or isinstance(cross_attention_dim, list):
positive_len = cross_attention_dim[0]
feature_type = "text-only" if attention_type == "gated" else "text-image"
self.position_net = PositionNet(
positive_len=positive_len, out_dim=cross_attention_dim, feature_type=feature_type
)
@property
def attn_processors(self) -> Dict[str, AttentionProcessor]:
r"""
Returns:
`dict` of attention processors: A dictionary containing all attention processors used in the model with
indexed by its weight name.
"""
# set recursively
processors = {}
def fn_recursive_add_processors(name: str, module: torch.nn.Module, processors: Dict[str, AttentionProcessor]):
if hasattr(module, "get_processor"):
processors[f"{name}.processor"] = module.get_processor(return_deprecated_lora=True)
for sub_name, child in module.named_children():
fn_recursive_add_processors(f"{name}.{sub_name}", child, processors)
return processors
for name, module in self.named_children():
fn_recursive_add_processors(name, module, processors)
return processors
def set_attn_processor(
self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]], _remove_lora=False
):
r"""
Sets the attention processor to use to compute attention.
Parameters:
processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`):
The instantiated processor class or a dictionary of processor classes that will be set as the processor
for **all** `Attention` layers.
If `processor` is a dict, the key needs to define the path to the corresponding cross attention
processor. This is strongly recommended when setting trainable attention processors.
"""
count = len(self.attn_processors.keys())
if isinstance(processor, dict) and len(processor) != count:
raise ValueError(
f"A dict of processors was passed, but the number of processors {len(processor)} does not match the"
f" number of attention layers: {count}. Please make sure to pass {count} processor classes."
)
def fn_recursive_attn_processor(name: str, module: torch.nn.Module, processor):
if hasattr(module, "set_processor"):
if not isinstance(processor, dict):
module.set_processor(processor, _remove_lora=_remove_lora)
else:
module.set_processor(processor.pop(f"{name}.processor"), _remove_lora=_remove_lora)
for sub_name, child in module.named_children():
fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)
for name, module in self.named_children():
fn_recursive_attn_processor(name, module, processor)
def set_default_attn_processor(self):
"""
Disables custom attention processors and sets the default attention implementation.
"""
if all(proc.__class__ in ADDED_KV_ATTENTION_PROCESSORS for proc in self.attn_processors.values()):
processor = AttnAddedKVProcessor()
elif all(proc.__class__ in CROSS_ATTENTION_PROCESSORS for proc in self.attn_processors.values()):
processor = AttnProcessor()
else:
raise ValueError(
f"Cannot call `set_default_attn_processor` when attention processors are of type {next(iter(self.attn_processors.values()))}"
)
self.set_attn_processor(processor, _remove_lora=True)
def set_attention_slice(self, slice_size):
r"""
Enable sliced attention computation.
When this option is enabled, the attention module splits the input tensor in slices to compute attention in
several steps. This is useful for saving some memory in exchange for a small decrease in speed.
Args:
slice_size (`str` or `int` or `list(int)`, *optional*, defaults to `"auto"`):
When `"auto"`, input to the attention heads is halved, so attention is computed in two steps. If
`"max"`, maximum amount of memory is saved by running only one slice at a time. If a number is
provided, uses as many slices as `attention_head_dim // slice_size`. In this case, `attention_head_dim`
must be a multiple of `slice_size`.
"""
sliceable_head_dims = []
def fn_recursive_retrieve_sliceable_dims(module: torch.nn.Module):
if hasattr(module, "set_attention_slice"):
sliceable_head_dims.append(module.sliceable_head_dim)
for child in module.children():
fn_recursive_retrieve_sliceable_dims(child)
# retrieve number of attention layers
for module in self.children():
fn_recursive_retrieve_sliceable_dims(module)
num_sliceable_layers = len(sliceable_head_dims)
if slice_size == "auto":
# half the attention head size is usually a good trade-off between
# speed and memory
slice_size = [dim // 2 for dim in sliceable_head_dims]
elif slice_size == "max":
# make smallest slice possible
slice_size = num_sliceable_layers * [1]
slice_size = num_sliceable_layers * [slice_size] if not isinstance(slice_size, list) else slice_size
if len(slice_size) != len(sliceable_head_dims):
raise ValueError(
f"You have provided {len(slice_size)}, but {self.config} has {len(sliceable_head_dims)} different"
f" attention layers. Make sure to match `len(slice_size)` to be {len(sliceable_head_dims)}."
)
for i in range(len(slice_size)):
size = slice_size[i]
dim = sliceable_head_dims[i]
if size is not None and size > dim:
raise ValueError(f"size {size} has to be smaller or equal to {dim}.")
# Recursively walk through all the children.
# Any children which exposes the set_attention_slice method
# gets the message
def fn_recursive_set_attention_slice(module: torch.nn.Module, slice_size: List[int]):
if hasattr(module, "set_attention_slice"):
module.set_attention_slice(slice_size.pop())
for child in module.children():
fn_recursive_set_attention_slice(child, slice_size)
reversed_slice_size = list(reversed(slice_size))
for module in self.children():
fn_recursive_set_attention_slice(module, reversed_slice_size)
def _set_gradient_checkpointing(self, module, value=False):
if hasattr(module, "gradient_checkpointing"):
module.gradient_checkpointing = value
def enable_freeu(self, s1, s2, b1, b2):
r"""Enables the FreeU mechanism from https://arxiv.org/abs/2309.11497.
The suffixes after the scaling factors represent the stage blocks where they are being applied.
Please refer to the [official repository](https://github.com/ChenyangSi/FreeU) for combinations of values that
are known to work well for different pipelines such as Stable Diffusion v1, v2, and Stable Diffusion XL.
Args:
s1 (`float`):
Scaling factor for stage 1 to attenuate the contributions of the skip features. This is done to
mitigate the "oversmoothing effect" in the enhanced denoising process.
s2 (`float`):
Scaling factor for stage 2 to attenuate the contributions of the skip features. This is done to
mitigate the "oversmoothing effect" in the enhanced denoising process.
b1 (`float`): Scaling factor for stage 1 to amplify the contributions of backbone features.
b2 (`float`): Scaling factor for stage 2 to amplify the contributions of backbone features.
"""
for i, upsample_block in enumerate(self.up_blocks):
setattr(upsample_block, "s1", s1)
setattr(upsample_block, "s2", s2)
setattr(upsample_block, "b1", b1)
setattr(upsample_block, "b2", b2)
def disable_freeu(self):
"""Disables the FreeU mechanism."""
freeu_keys = {"s1", "s2", "b1", "b2"}
for i, upsample_block in enumerate(self.up_blocks):
for k in freeu_keys:
if hasattr(upsample_block, k) or getattr(upsample_block, k, None) is not None:
setattr(upsample_block, k, None)
def forward(
self,
sample: torch.FloatTensor,
timestep: Union[torch.Tensor, float, int],
encoder_hidden_states: torch.Tensor,
class_labels: Optional[torch.Tensor] = None,
timestep_cond: Optional[torch.Tensor] = None,
attention_mask: Optional[torch.Tensor] = None,
cross_attention_kwargs: Optional[Dict[str, Any]] = None,
added_cond_kwargs: Optional[Dict[str, torch.Tensor]] = None,
down_block_additional_residuals: Optional[Tuple[torch.Tensor]] = None,
mid_block_additional_residual: Optional[torch.Tensor] = None,
down_intrablock_additional_residuals: Optional[Tuple[torch.Tensor]] = None,
encoder_attention_mask: Optional[torch.Tensor] = None,
return_dict: bool = True,
) -> Union[UNet2DConditionOutput, Tuple]:
r"""
The [`UNet2DConditionModel`] forward method.
Args:
sample (`torch.FloatTensor`):
The noisy input tensor with the following shape `(batch, channel, height, width)`.
timestep (`torch.FloatTensor` or `float` or `int`): The number of timesteps to denoise an input.
encoder_hidden_states (`torch.FloatTensor`):
The encoder hidden states with shape `(batch, sequence_length, feature_dim)`.
class_labels (`torch.Tensor`, *optional*, defaults to `None`):
Optional class labels for conditioning. Their embeddings will be summed with the timestep embeddings.
timestep_cond: (`torch.Tensor`, *optional*, defaults to `None`):
Conditional embeddings for timestep. If provided, the embeddings will be summed with the samples passed
through the `self.time_embedding` layer to obtain the timestep embeddings.
attention_mask (`torch.Tensor`, *optional*, defaults to `None`):
An attention mask of shape `(batch, key_tokens)` is applied to `encoder_hidden_states`. If `1` the mask
is kept, otherwise if `0` it is discarded. Mask will be converted into a bias, which adds large
negative values to the attention scores corresponding to "discard" tokens.
cross_attention_kwargs (`dict`, *optional*):
A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
`self.processor` in
[diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
added_cond_kwargs: (`dict`, *optional*):
A kwargs dictionary containing additional embeddings that if specified are added to the embeddings that
are passed along to the UNet blocks.
down_block_additional_residuals: (`tuple` of `torch.Tensor`, *optional*):
A tuple of tensors that if specified are added to the residuals of down unet blocks.
mid_block_additional_residual: (`torch.Tensor`, *optional*):
A tensor that if specified is added to the residual of the middle unet block.
encoder_attention_mask (`torch.Tensor`):
A cross-attention mask of shape `(batch, sequence_length)` is applied to `encoder_hidden_states`. If
`True` the mask is kept, otherwise if `False` it is discarded. Mask will be converted into a bias,
which adds large negative values to the attention scores corresponding to "discard" tokens.
return_dict (`bool`, *optional*, defaults to `True`):
Whether or not to return a [`~models.unet_2d_condition.UNet2DConditionOutput`] instead of a plain
tuple.
cross_attention_kwargs (`dict`, *optional*):
A kwargs dictionary that if specified is passed along to the [`AttnProcessor`].
added_cond_kwargs: (`dict`, *optional*):
A kwargs dictionary containin additional embeddings that if specified are added to the embeddings that
are passed along to the UNet blocks.
down_block_additional_residuals (`tuple` of `torch.Tensor`, *optional*):
additional residuals to be added to UNet long skip connections from down blocks to up blocks for
example from ControlNet side model(s)
mid_block_additional_residual (`torch.Tensor`, *optional*):
additional residual to be added to UNet mid block output, for example from ControlNet side model
down_intrablock_additional_residuals (`tuple` of `torch.Tensor`, *optional*):
additional residuals to be added within UNet down blocks, for example from T2I-Adapter side model(s)
Returns:
[`~models.unet_2d_condition.UNet2DConditionOutput`] or `tuple`:
If `return_dict` is True, an [`~models.unet_2d_condition.UNet2DConditionOutput`] is returned, otherwise
a `tuple` is returned where the first element is the sample tensor.
"""
# By default samples have to be AT least a multiple of the overall upsampling factor.
# The overall upsampling factor is equal to 2 ** (# num of upsampling layers).
# However, the upsampling interpolation output size can be forced to fit any upsampling size
# on the fly if necessary.
default_overall_up_factor = 2**self.num_upsamplers
# upsample size should be forwarded when sample is not a multiple of `default_overall_up_factor`
forward_upsample_size = False
upsample_size = None
for dim in sample.shape[-2:]:
if dim % default_overall_up_factor != 0:
# Forward upsample size to force interpolation output size.
forward_upsample_size = True
break
# ensure attention_mask is a bias, and give it a singleton query_tokens dimension
# expects mask of shape:
# [batch, key_tokens]
# adds singleton query_tokens dimension:
# [batch, 1, key_tokens]
# this helps to broadcast it as a bias over attention scores, which will be in one of the following shapes:
# [batch, heads, query_tokens, key_tokens] (e.g. torch sdp attn)
# [batch * heads, query_tokens, key_tokens] (e.g. xformers or classic attn)
if attention_mask is not None:
# assume that mask is expressed as:
# (1 = keep, 0 = discard)
# convert mask into a bias that can be added to attention scores:
# (keep = +0, discard = -10000.0)
attention_mask = (1 - attention_mask.to(sample.dtype)) * -10000.0
attention_mask = attention_mask.unsqueeze(1)
# convert encoder_attention_mask to a bias the same way we do for attention_mask
if encoder_attention_mask is not None:
encoder_attention_mask = (1 - encoder_attention_mask.to(sample.dtype)) * -10000.0
encoder_attention_mask = encoder_attention_mask.unsqueeze(1)
# 0. center input if necessary
if self.config.center_input_sample:
sample = 2 * sample - 1.0
# 1. time
timesteps = timestep
if not torch.is_tensor(timesteps):
# TODO: this requires sync between CPU and GPU. So try to pass timesteps as tensors if you can
# This would be a good case for the `match` statement (Python 3.10+)
is_mps = sample.device.type == "mps"
if isinstance(timestep, float):
dtype = torch.float32 if is_mps else torch.float64
else:
dtype = torch.int32 if is_mps else torch.int64
timesteps = torch.tensor([timesteps], dtype=dtype, device=sample.device)
elif len(timesteps.shape) == 0:
timesteps = timesteps[None].to(sample.device)
# broadcast to batch dimension in a way that's compatible with ONNX/Core ML
timesteps = timesteps.expand(sample.shape[0])
t_emb = self.time_proj(timesteps)
# `Timesteps` does not contain any weights and will always return f32 tensors
# but time_embedding might actually be running in fp16. so we need to cast here.
# there might be better ways to encapsulate this.
t_emb = t_emb.to(dtype=sample.dtype)
emb = self.time_embedding(t_emb, timestep_cond)
aug_emb = None
if self.class_embedding is not None:
if class_labels is None:
raise ValueError("class_labels should be provided when num_class_embeds > 0")
if self.config.class_embed_type == "timestep":
class_labels = self.time_proj(class_labels)
# `Timesteps` does not contain any weights and will always return f32 tensors
# there might be better ways to encapsulate this.
class_labels = class_labels.to(dtype=sample.dtype)
class_emb = self.class_embedding(class_labels).to(dtype=sample.dtype)
if self.config.class_embeddings_concat:
emb = torch.cat([emb, class_emb], dim=-1)
else:
emb = emb + class_emb
if self.config.addition_embed_type == "text":
aug_emb = self.add_embedding(encoder_hidden_states)
elif self.config.addition_embed_type == "text_image":
# Kandinsky 2.1 - style
if "image_embeds" not in added_cond_kwargs:
raise ValueError(
f"{self.__class__} has the config param `addition_embed_type` set to 'text_image' which requires the keyword argument `image_embeds` to be passed in `added_cond_kwargs`"
)
image_embs = added_cond_kwargs.get("image_embeds")
text_embs = added_cond_kwargs.get("text_embeds", encoder_hidden_states)
aug_emb = self.add_embedding(text_embs, image_embs)
elif self.config.addition_embed_type == "text_time":
# SDXL - style
if "text_embeds" not in added_cond_kwargs:
raise ValueError(
f"{self.__class__} has the config param `addition_embed_type` set to 'text_time' which requires the keyword argument `text_embeds` to be passed in `added_cond_kwargs`"
)
text_embeds = added_cond_kwargs.get("text_embeds")
if "time_ids" not in added_cond_kwargs:
raise ValueError(
f"{self.__class__} has the config param `addition_embed_type` set to 'text_time' which requires the keyword argument `time_ids` to be passed in `added_cond_kwargs`"
)
time_ids = added_cond_kwargs.get("time_ids")
time_embeds = self.add_time_proj(time_ids.flatten())
time_embeds = time_embeds.reshape((text_embeds.shape[0], -1))
add_embeds = torch.concat([text_embeds, time_embeds], dim=-1)
add_embeds = add_embeds.to(emb.dtype)
aug_emb = self.add_embedding(add_embeds)
elif self.config.addition_embed_type == "image":
# Kandinsky 2.2 - style
if "image_embeds" not in added_cond_kwargs:
raise ValueError(
f"{self.__class__} has the config param `addition_embed_type` set to 'image' which requires the keyword argument `image_embeds` to be passed in `added_cond_kwargs`"
)
image_embs = added_cond_kwargs.get("image_embeds")
aug_emb = self.add_embedding(image_embs)
elif self.config.addition_embed_type == "image_hint":
# Kandinsky 2.2 - style
if "image_embeds" not in added_cond_kwargs or "hint" not in added_cond_kwargs:
raise ValueError(
f"{self.__class__} has the config param `addition_embed_type` set to 'image_hint' which requires the keyword arguments `image_embeds` and `hint` to be passed in `added_cond_kwargs`"
)
image_embs = added_cond_kwargs.get("image_embeds")
hint = added_cond_kwargs.get("hint")
aug_emb, hint = self.add_embedding(image_embs, hint)
sample = torch.cat([sample, hint], dim=1)
emb = emb + aug_emb if aug_emb is not None else emb
if self.time_embed_act is not None:
emb = self.time_embed_act(emb)
if self.encoder_hid_proj is not None and self.config.encoder_hid_dim_type == "text_proj":
encoder_hidden_states = self.encoder_hid_proj(encoder_hidden_states)
elif self.encoder_hid_proj is not None and self.config.encoder_hid_dim_type == "text_image_proj":
# Kadinsky 2.1 - style
if "image_embeds" not in added_cond_kwargs:
raise ValueError(
f"{self.__class__} has the config param `encoder_hid_dim_type` set to 'text_image_proj' which requires the keyword argument `image_embeds` to be passed in `added_conditions`"
)
image_embeds = added_cond_kwargs.get("image_embeds")
encoder_hidden_states = self.encoder_hid_proj(encoder_hidden_states, image_embeds)
elif self.encoder_hid_proj is not None and self.config.encoder_hid_dim_type == "image_proj":
# Kandinsky 2.2 - style
if "image_embeds" not in added_cond_kwargs:
raise ValueError(
f"{self.__class__} has the config param `encoder_hid_dim_type` set to 'image_proj' which requires the keyword argument `image_embeds` to be passed in `added_conditions`"
)
image_embeds = added_cond_kwargs.get("image_embeds")
encoder_hidden_states = self.encoder_hid_proj(image_embeds)
# 2. pre-process
sample = self.conv_in(sample)
# 2.5 GLIGEN position net
if cross_attention_kwargs is not None and cross_attention_kwargs.get("gligen", None) is not None:
cross_attention_kwargs = cross_attention_kwargs.copy()
gligen_args = cross_attention_kwargs.pop("gligen")
cross_attention_kwargs["gligen"] = {"objs": self.position_net(**gligen_args)}
# For Vton
spatial_attn_inputs = []
# 3. down
lora_scale = cross_attention_kwargs.get("scale", 1.0) if cross_attention_kwargs is not None else 1.0
if USE_PEFT_BACKEND:
# weight the lora layers by setting `lora_scale` for each PEFT layer
scale_lora_layers(self, lora_scale)
is_controlnet = mid_block_additional_residual is not None and down_block_additional_residuals is not None
# using new arg down_intrablock_additional_residuals for T2I-Adapters, to distinguish from controlnets
is_adapter = down_intrablock_additional_residuals is not None
# maintain backward compatibility for legacy usage, where
# T2I-Adapter and ControlNet both use down_block_additional_residuals arg
# but can only use one or the other
if not is_adapter and mid_block_additional_residual is None and down_block_additional_residuals is not None:
deprecate(
"T2I should not use down_block_additional_residuals",
"1.3.0",
"Passing intrablock residual connections with `down_block_additional_residuals` is deprecated \
and will be removed in diffusers 1.3.0. `down_block_additional_residuals` should only be used \
for ControlNet. Please make sure use `down_intrablock_additional_residuals` instead. ",
standard_warn=False,
)
down_intrablock_additional_residuals = down_block_additional_residuals
is_adapter = True
down_block_res_samples = (sample,)
for downsample_block in self.down_blocks:
if hasattr(downsample_block, "has_cross_attention") and downsample_block.has_cross_attention:
# For t2i-adapter CrossAttnDownBlock2D
additional_residuals = {}
if is_adapter and len(down_intrablock_additional_residuals) > 0:
additional_residuals["additional_residuals"] = down_intrablock_additional_residuals.pop(0)
sample, res_samples, spatial_attn_inputs = downsample_block(
hidden_states=sample,
spatial_attn_inputs=spatial_attn_inputs,
temb=emb,
encoder_hidden_states=encoder_hidden_states,
attention_mask=attention_mask,
cross_attention_kwargs=cross_attention_kwargs,
encoder_attention_mask=encoder_attention_mask,
**additional_residuals,
)
else:
sample, res_samples = downsample_block(
hidden_states=sample,
temb=emb,
scale=lora_scale,
)
if is_adapter and len(down_intrablock_additional_residuals) > 0:
sample += down_intrablock_additional_residuals.pop(0)
down_block_res_samples += res_samples
# if is_controlnet:
# new_down_block_res_samples = ()
# for down_block_res_sample, down_block_additional_residual in zip(
# down_block_res_samples, down_block_additional_residuals
# ):
# down_block_res_sample = down_block_res_sample + down_block_additional_residual
# new_down_block_res_samples = new_down_block_res_samples + (down_block_res_sample,)
# down_block_res_samples = new_down_block_res_samples
# 4. mid
if self.mid_block is not None:
if hasattr(self.mid_block, "has_cross_attention") and self.mid_block.has_cross_attention:
sample, spatial_attn_inputs = self.mid_block(
sample,
spatial_attn_inputs=spatial_attn_inputs,
temb=emb,
encoder_hidden_states=encoder_hidden_states,
attention_mask=attention_mask,
cross_attention_kwargs=cross_attention_kwargs,
encoder_attention_mask=encoder_attention_mask,
)
else:
sample = self.mid_block(sample, emb)
# To support T2I-Adapter-XL
if (
is_adapter
and len(down_intrablock_additional_residuals) > 0
and sample.shape == down_intrablock_additional_residuals[0].shape
):
sample += down_intrablock_additional_residuals.pop(0)
if is_controlnet:
sample = sample + mid_block_additional_residual
# 5. up
for i, upsample_block in enumerate(self.up_blocks):
is_final_block = i == len(self.up_blocks) - 1
res_samples = down_block_res_samples[-len(upsample_block.resnets) :]
down_block_res_samples = down_block_res_samples[: -len(upsample_block.resnets)]
# if we have not reached the final block and need to forward the
# upsample size, we do it here
if not is_final_block and forward_upsample_size:
upsample_size = down_block_res_samples[-1].shape[2:]
if hasattr(upsample_block, "has_cross_attention") and upsample_block.has_cross_attention:
sample, spatial_attn_inputs = upsample_block(
hidden_states=sample,
spatial_attn_inputs=spatial_attn_inputs,
temb=emb,
res_hidden_states_tuple=res_samples,
encoder_hidden_states=encoder_hidden_states,
cross_attention_kwargs=cross_attention_kwargs,
upsample_size=upsample_size,
attention_mask=attention_mask,
encoder_attention_mask=encoder_attention_mask,
)
else:
sample = upsample_block(
hidden_states=sample,
temb=emb,
res_hidden_states_tuple=res_samples,
upsample_size=upsample_size,
scale=lora_scale,
)
# 6. post-process
if self.conv_norm_out:
sample = self.conv_norm_out(sample)
sample = self.conv_act(sample)
sample = self.conv_out(sample)
if USE_PEFT_BACKEND:
# remove `lora_scale` from each PEFT layer
unscale_lora_layers(self, lora_scale)
if not return_dict:
return (sample,), spatial_attn_inputs
return UNet2DConditionOutput(sample=sample), spatial_attn_inputs
ootd/pipelines_ootd/unet_vton_2d_blocks.py 0 → 100644
View file @ 54a066bf
# Copyright 2023 The HuggingFace Team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# Modified by Yuhao Xu for OOTDiffusion (https://github.com/levihsu/OOTDiffusion)
from typing import Any, Dict, Optional, Tuple, Union
import numpy as np
import torch
import torch.nn.functional as F
from torch import nn
from .transformer_vton_2d import Transformer2DModel
from diffusers.utils import is_torch_version, logging
from diffusers.utils.torch_utils import apply_freeu
from diffusers.models.activations import get_activation
from diffusers.models.attention_processor import Attention, AttnAddedKVProcessor, AttnAddedKVProcessor2_0
from diffusers.models.dual_transformer_2d import DualTransformer2DModel
from diffusers.models.normalization import AdaGroupNorm
from diffusers.models.resnet import Downsample2D, FirDownsample2D, FirUpsample2D, KDownsample2D, KUpsample2D, ResnetBlock2D, Upsample2D
# from diffusers.models.transformer_2d import Transformer2DModel
logger = logging.get_logger(__name__) # pylint: disable=invalid-name
def get_down_block(
down_block_type: str,
num_layers: int,
in_channels: int,
out_channels: int,
temb_channels: int,
add_downsample: bool,
resnet_eps: float,
resnet_act_fn: str,
transformer_layers_per_block: int = 1,
num_attention_heads: Optional[int] = None,
resnet_groups: Optional[int] = None,
cross_attention_dim: Optional[int] = None,
downsample_padding: Optional[int] = None,
dual_cross_attention: bool = False,
use_linear_projection: bool = False,
only_cross_attention: bool = False,
upcast_attention: bool = False,
resnet_time_scale_shift: str = "default",
attention_type: str = "default",
resnet_skip_time_act: bool = False,
resnet_out_scale_factor: float = 1.0,
cross_attention_norm: Optional[str] = None,
attention_head_dim: Optional[int] = None,
downsample_type: Optional[str] = None,
dropout: float = 0.0,
):
# If attn head dim is not defined, we default it to the number of heads
if attention_head_dim is None:
logger.warn(
f"It is recommended to provide `attention_head_dim` when calling `get_down_block`. Defaulting `attention_head_dim` to {num_attention_heads}."
)
attention_head_dim = num_attention_heads
down_block_type = down_block_type[7:] if down_block_type.startswith("UNetRes") else down_block_type
if down_block_type == "DownBlock2D":
return DownBlock2D(
num_layers=num_layers,
in_channels=in_channels,
out_channels=out_channels,
temb_channels=temb_channels,
dropout=dropout,
add_downsample=add_downsample,
resnet_eps=resnet_eps,
resnet_act_fn=resnet_act_fn,
resnet_groups=resnet_groups,
downsample_padding=downsample_padding,
resnet_time_scale_shift=resnet_time_scale_shift,
)
elif down_block_type == "ResnetDownsampleBlock2D":
return ResnetDownsampleBlock2D(
num_layers=num_layers,
in_channels=in_channels,
out_channels=out_channels,
temb_channels=temb_channels,
dropout=dropout,
add_downsample=add_downsample,
resnet_eps=resnet_eps,
resnet_act_fn=resnet_act_fn,
resnet_groups=resnet_groups,
resnet_time_scale_shift=resnet_time_scale_shift,
skip_time_act=resnet_skip_time_act,
output_scale_factor=resnet_out_scale_factor,
)
elif down_block_type == "AttnDownBlock2D":
if add_downsample is False:
downsample_type = None
else:
downsample_type = downsample_type or "conv" # default to 'conv'
return AttnDownBlock2D(
num_layers=num_layers,
in_channels=in_channels,
out_channels=out_channels,
temb_channels=temb_channels,
dropout=dropout,
resnet_eps=resnet_eps,
resnet_act_fn=resnet_act_fn,
resnet_groups=resnet_groups,
downsample_padding=downsample_padding,
attention_head_dim=attention_head_dim,
resnet_time_scale_shift=resnet_time_scale_shift,
downsample_type=downsample_type,
)
elif down_block_type == "CrossAttnDownBlock2D":
if cross_attention_dim is None:
raise ValueError("cross_attention_dim must be specified for CrossAttnDownBlock2D")
return CrossAttnDownBlock2D(
num_layers=num_layers,
transformer_layers_per_block=transformer_layers_per_block,
in_channels=in_channels,
out_channels=out_channels,
temb_channels=temb_channels,
dropout=dropout,
add_downsample=add_downsample,
resnet_eps=resnet_eps,
resnet_act_fn=resnet_act_fn,
resnet_groups=resnet_groups,
downsample_padding=downsample_padding,
cross_attention_dim=cross_attention_dim,
num_attention_heads=num_attention_heads,
dual_cross_attention=dual_cross_attention,
use_linear_projection=use_linear_projection,
only_cross_attention=only_cross_attention,
upcast_attention=upcast_attention,
resnet_time_scale_shift=resnet_time_scale_shift,
attention_type=attention_type,
)
elif down_block_type == "SimpleCrossAttnDownBlock2D":
if cross_attention_dim is None:
raise ValueError("cross_attention_dim must be specified for SimpleCrossAttnDownBlock2D")
return SimpleCrossAttnDownBlock2D(
num_layers=num_layers,
in_channels=in_channels,
out_channels=out_channels,
temb_channels=temb_channels,
dropout=dropout,
add_downsample=add_downsample,
resnet_eps=resnet_eps,
resnet_act_fn=resnet_act_fn,
resnet_groups=resnet_groups,
cross_attention_dim=cross_attention_dim,
attention_head_dim=attention_head_dim,
resnet_time_scale_shift=resnet_time_scale_shift,
skip_time_act=resnet_skip_time_act,
output_scale_factor=resnet_out_scale_factor,
only_cross_attention=only_cross_attention,
cross_attention_norm=cross_attention_norm,
)
elif down_block_type == "SkipDownBlock2D":
return SkipDownBlock2D(
num_layers=num_layers,
in_channels=in_channels,
out_channels=out_channels,
temb_channels=temb_channels,
dropout=dropout,
add_downsample=add_downsample,
resnet_eps=resnet_eps,
resnet_act_fn=resnet_act_fn,
downsample_padding=downsample_padding,
resnet_time_scale_shift=resnet_time_scale_shift,
)
elif down_block_type == "AttnSkipDownBlock2D":
return AttnSkipDownBlock2D(
num_layers=num_layers,
in_channels=in_channels,
out_channels=out_channels,
temb_channels=temb_channels,
dropout=dropout,
add_downsample=add_downsample,
resnet_eps=resnet_eps,
resnet_act_fn=resnet_act_fn,
attention_head_dim=attention_head_dim,
resnet_time_scale_shift=resnet_time_scale_shift,
)
elif down_block_type == "DownEncoderBlock2D":
return DownEncoderBlock2D(
num_layers=num_layers,
in_channels=in_channels,
out_channels=out_channels,
dropout=dropout,
add_downsample=add_downsample,
resnet_eps=resnet_eps,
resnet_act_fn=resnet_act_fn,
resnet_groups=resnet_groups,
downsample_padding=downsample_padding,
resnet_time_scale_shift=resnet_time_scale_shift,
)
elif down_block_type == "AttnDownEncoderBlock2D":
return AttnDownEncoderBlock2D(
num_layers=num_layers,
in_channels=in_channels,
out_channels=out_channels,
dropout=dropout,
add_downsample=add_downsample,
resnet_eps=resnet_eps,
resnet_act_fn=resnet_act_fn,
resnet_groups=resnet_groups,
downsample_padding=downsample_padding,
attention_head_dim=attention_head_dim,
resnet_time_scale_shift=resnet_time_scale_shift,
)
elif down_block_type == "KDownBlock2D":
return KDownBlock2D(
num_layers=num_layers,
in_channels=in_channels,
out_channels=out_channels,
temb_channels=temb_channels,
dropout=dropout,
add_downsample=add_downsample,
resnet_eps=resnet_eps,
resnet_act_fn=resnet_act_fn,
)
elif down_block_type == "KCrossAttnDownBlock2D":
return KCrossAttnDownBlock2D(
num_layers=num_layers,
in_channels=in_channels,
out_channels=out_channels,
temb_channels=temb_channels,
dropout=dropout,
add_downsample=add_downsample,
resnet_eps=resnet_eps,
resnet_act_fn=resnet_act_fn,
cross_attention_dim=cross_attention_dim,
attention_head_dim=attention_head_dim,
add_self_attention=True if not add_downsample else False,
)
raise ValueError(f"{down_block_type} does not exist.")
def get_up_block(
up_block_type: str,
num_layers: int,
in_channels: int,
out_channels: int,
prev_output_channel: int,
temb_channels: int,
add_upsample: bool,
resnet_eps: float,
resnet_act_fn: str,
resolution_idx: Optional[int] = None,
transformer_layers_per_block: int = 1,
num_attention_heads: Optional[int] = None,
resnet_groups: Optional[int] = None,
cross_attention_dim: Optional[int] = None,
dual_cross_attention: bool = False,
use_linear_projection: bool = False,
only_cross_attention: bool = False,
upcast_attention: bool = False,
resnet_time_scale_shift: str = "default",
attention_type: str = "default",
resnet_skip_time_act: bool = False,
resnet_out_scale_factor: float = 1.0,
cross_attention_norm: Optional[str] = None,
attention_head_dim: Optional[int] = None,
upsample_type: Optional[str] = None,
dropout: float = 0.0,
) -> nn.Module:
# If attn head dim is not defined, we default it to the number of heads
if attention_head_dim is None:
logger.warn(
f"It is recommended to provide `attention_head_dim` when calling `get_up_block`. Defaulting `attention_head_dim` to {num_attention_heads}."
)
attention_head_dim = num_attention_heads
up_block_type = up_block_type[7:] if up_block_type.startswith("UNetRes") else up_block_type
if up_block_type == "UpBlock2D":
return UpBlock2D(
num_layers=num_layers,
in_channels=in_channels,
out_channels=out_channels,
prev_output_channel=prev_output_channel,
temb_channels=temb_channels,
resolution_idx=resolution_idx,
dropout=dropout,
add_upsample=add_upsample,
resnet_eps=resnet_eps,
resnet_act_fn=resnet_act_fn,
resnet_groups=resnet_groups,
resnet_time_scale_shift=resnet_time_scale_shift,
)
elif up_block_type == "ResnetUpsampleBlock2D":
return ResnetUpsampleBlock2D(
num_layers=num_layers,
in_channels=in_channels,
out_channels=out_channels,
prev_output_channel=prev_output_channel,
temb_channels=temb_channels,
resolution_idx=resolution_idx,
dropout=dropout,
add_upsample=add_upsample,
resnet_eps=resnet_eps,
resnet_act_fn=resnet_act_fn,
resnet_groups=resnet_groups,
resnet_time_scale_shift=resnet_time_scale_shift,
skip_time_act=resnet_skip_time_act,
output_scale_factor=resnet_out_scale_factor,
)
elif up_block_type == "CrossAttnUpBlock2D":
if cross_attention_dim is None:
raise ValueError("cross_attention_dim must be specified for CrossAttnUpBlock2D")
return CrossAttnUpBlock2D(
num_layers=num_layers,
transformer_layers_per_block=transformer_layers_per_block,
in_channels=in_channels,
out_channels=out_channels,
prev_output_channel=prev_output_channel,
temb_channels=temb_channels,
resolution_idx=resolution_idx,
dropout=dropout,
add_upsample=add_upsample,
resnet_eps=resnet_eps,
resnet_act_fn=resnet_act_fn,
resnet_groups=resnet_groups,
cross_attention_dim=cross_attention_dim,
num_attention_heads=num_attention_heads,
dual_cross_attention=dual_cross_attention,
use_linear_projection=use_linear_projection,
only_cross_attention=only_cross_attention,
upcast_attention=upcast_attention,
resnet_time_scale_shift=resnet_time_scale_shift,
attention_type=attention_type,
)
elif up_block_type == "SimpleCrossAttnUpBlock2D":
if cross_attention_dim is None:
raise ValueError("cross_attention_dim must be specified for SimpleCrossAttnUpBlock2D")
return SimpleCrossAttnUpBlock2D(
num_layers=num_layers,
in_channels=in_channels,
out_channels=out_channels,
prev_output_channel=prev_output_channel,
temb_channels=temb_channels,
resolution_idx=resolution_idx,
dropout=dropout,
add_upsample=add_upsample,
resnet_eps=resnet_eps,
resnet_act_fn=resnet_act_fn,
resnet_groups=resnet_groups,
cross_attention_dim=cross_attention_dim,
attention_head_dim=attention_head_dim,
resnet_time_scale_shift=resnet_time_scale_shift,
skip_time_act=resnet_skip_time_act,
output_scale_factor=resnet_out_scale_factor,
only_cross_attention=only_cross_attention,
cross_attention_norm=cross_attention_norm,
)
elif up_block_type == "AttnUpBlock2D":
if add_upsample is False:
upsample_type = None
else:
upsample_type = upsample_type or "conv" # default to 'conv'
return AttnUpBlock2D(
num_layers=num_layers,
in_channels=in_channels,
out_channels=out_channels,
prev_output_channel=prev_output_channel,
temb_channels=temb_channels,
resolution_idx=resolution_idx,
dropout=dropout,
resnet_eps=resnet_eps,
resnet_act_fn=resnet_act_fn,
resnet_groups=resnet_groups,
attention_head_dim=attention_head_dim,
resnet_time_scale_shift=resnet_time_scale_shift,
upsample_type=upsample_type,
)
elif up_block_type == "SkipUpBlock2D":
return SkipUpBlock2D(
num_layers=num_layers,
in_channels=in_channels,
out_channels=out_channels,
prev_output_channel=prev_output_channel,
temb_channels=temb_channels,
resolution_idx=resolution_idx,
dropout=dropout,
add_upsample=add_upsample,
resnet_eps=resnet_eps,
resnet_act_fn=resnet_act_fn,
resnet_time_scale_shift=resnet_time_scale_shift,
)
elif up_block_type == "AttnSkipUpBlock2D":
return AttnSkipUpBlock2D(
num_layers=num_layers,
in_channels=in_channels,
out_channels=out_channels,
prev_output_channel=prev_output_channel,
temb_channels=temb_channels,
resolution_idx=resolution_idx,
dropout=dropout,
add_upsample=add_upsample,
resnet_eps=resnet_eps,
resnet_act_fn=resnet_act_fn,
attention_head_dim=attention_head_dim,
resnet_time_scale_shift=resnet_time_scale_shift,
)
elif up_block_type == "UpDecoderBlock2D":
return UpDecoderBlock2D(
num_layers=num_layers,
in_channels=in_channels,
out_channels=out_channels,
resolution_idx=resolution_idx,
dropout=dropout,
add_upsample=add_upsample,
resnet_eps=resnet_eps,
resnet_act_fn=resnet_act_fn,
resnet_groups=resnet_groups,
resnet_time_scale_shift=resnet_time_scale_shift,
temb_channels=temb_channels,
)
elif up_block_type == "AttnUpDecoderBlock2D":
return AttnUpDecoderBlock2D(
num_layers=num_layers,
in_channels=in_channels,
out_channels=out_channels,
resolution_idx=resolution_idx,
dropout=dropout,
add_upsample=add_upsample,
resnet_eps=resnet_eps,
resnet_act_fn=resnet_act_fn,
resnet_groups=resnet_groups,
attention_head_dim=attention_head_dim,
resnet_time_scale_shift=resnet_time_scale_shift,
temb_channels=temb_channels,
)
elif up_block_type == "KUpBlock2D":
return KUpBlock2D(
num_layers=num_layers,
in_channels=in_channels,
out_channels=out_channels,
temb_channels=temb_channels,
resolution_idx=resolution_idx,
dropout=dropout,
add_upsample=add_upsample,
resnet_eps=resnet_eps,
resnet_act_fn=resnet_act_fn,
)
elif up_block_type == "KCrossAttnUpBlock2D":
return KCrossAttnUpBlock2D(
num_layers=num_layers,
in_channels=in_channels,
out_channels=out_channels,
temb_channels=temb_channels,
resolution_idx=resolution_idx,
dropout=dropout,
add_upsample=add_upsample,
resnet_eps=resnet_eps,
resnet_act_fn=resnet_act_fn,
cross_attention_dim=cross_attention_dim,
attention_head_dim=attention_head_dim,
)
raise ValueError(f"{up_block_type} does not exist.")
class AutoencoderTinyBlock(nn.Module):
"""
Tiny Autoencoder block used in [`AutoencoderTiny`]. It is a mini residual module consisting of plain conv + ReLU
blocks.
Args:
in_channels (`int`): The number of input channels.
out_channels (`int`): The number of output channels.
act_fn (`str`):
` The activation function to use. Supported values are `"swish"`, `"mish"`, `"gelu"`, and `"relu"`.
Returns:
`torch.FloatTensor`: A tensor with the same shape as the input tensor, but with the number of channels equal to
`out_channels`.
"""
def __init__(self, in_channels: int, out_channels: int, act_fn: str):
super().__init__()
act_fn = get_activation(act_fn)
self.conv = nn.Sequential(
nn.Conv2d(in_channels, out_channels, kernel_size=3, padding=1),
act_fn,
nn.Conv2d(out_channels, out_channels, kernel_size=3, padding=1),
act_fn,
nn.Conv2d(out_channels, out_channels, kernel_size=3, padding=1),
)
self.skip = (
nn.Conv2d(in_channels, out_channels, kernel_size=1, bias=False)
if in_channels != out_channels
else nn.Identity()
)
self.fuse = nn.ReLU()
def forward(self, x: torch.FloatTensor) -> torch.FloatTensor:
return self.fuse(self.conv(x) + self.skip(x))
class UNetMidBlock2D(nn.Module):
"""
A 2D UNet mid-block [`UNetMidBlock2D`] with multiple residual blocks and optional attention blocks.
Args:
in_channels (`int`): The number of input channels.
temb_channels (`int`): The number of temporal embedding channels.
dropout (`float`, *optional*, defaults to 0.0): The dropout rate.
num_layers (`int`, *optional*, defaults to 1): The number of residual blocks.
resnet_eps (`float`, *optional*, 1e-6 ): The epsilon value for the resnet blocks.
resnet_time_scale_shift (`str`, *optional*, defaults to `default`):
The type of normalization to apply to the time embeddings. This can help to improve the performance of the
model on tasks with long-range temporal dependencies.
resnet_act_fn (`str`, *optional*, defaults to `swish`): The activation function for the resnet blocks.
resnet_groups (`int`, *optional*, defaults to 32):
The number of groups to use in the group normalization layers of the resnet blocks.
attn_groups (`Optional[int]`, *optional*, defaults to None): The number of groups for the attention blocks.
resnet_pre_norm (`bool`, *optional*, defaults to `True`):
Whether to use pre-normalization for the resnet blocks.
add_attention (`bool`, *optional*, defaults to `True`): Whether to add attention blocks.
attention_head_dim (`int`, *optional*, defaults to 1):
Dimension of a single attention head. The number of attention heads is determined based on this value and
the number of input channels.
output_scale_factor (`float`, *optional*, defaults to 1.0): The output scale factor.
Returns:
`torch.FloatTensor`: The output of the last residual block, which is a tensor of shape `(batch_size,
in_channels, height, width)`.
"""
def __init__(
self,
in_channels: int,
temb_channels: int,
dropout: float = 0.0,
num_layers: int = 1,
resnet_eps: float = 1e-6,
resnet_time_scale_shift: str = "default", # default, spatial
resnet_act_fn: str = "swish",
resnet_groups: int = 32,
attn_groups: Optional[int] = None,
resnet_pre_norm: bool = True,
add_attention: bool = True,
attention_head_dim: int = 1,
output_scale_factor: float = 1.0,
):
super().__init__()
resnet_groups = resnet_groups if resnet_groups is not None else min(in_channels // 4, 32)
self.add_attention = add_attention
if attn_groups is None:
attn_groups = resnet_groups if resnet_time_scale_shift == "default" else None
# there is always at least one resnet
resnets = [
ResnetBlock2D(
in_channels=in_channels,
out_channels=in_channels,
temb_channels=temb_channels,
eps=resnet_eps,
groups=resnet_groups,
dropout=dropout,
time_embedding_norm=resnet_time_scale_shift,
non_linearity=resnet_act_fn,
output_scale_factor=output_scale_factor,
pre_norm=resnet_pre_norm,
)
]
attentions = []
if attention_head_dim is None:
logger.warn(
f"It is not recommend to pass `attention_head_dim=None`. Defaulting `attention_head_dim` to `in_channels`: {in_channels}."
)
attention_head_dim = in_channels
for _ in range(num_layers):
if self.add_attention:
attentions.append(
Attention(
in_channels,
heads=in_channels // attention_head_dim,
dim_head=attention_head_dim,
rescale_output_factor=output_scale_factor,
eps=resnet_eps,
norm_num_groups=attn_groups,
spatial_norm_dim=temb_channels if resnet_time_scale_shift == "spatial" else None,
residual_connection=True,
bias=True,
upcast_softmax=True,
_from_deprecated_attn_block=True,
)
)
else:
attentions.append(None)
resnets.append(
ResnetBlock2D(
in_channels=in_channels,
out_channels=in_channels,
temb_channels=temb_channels,
eps=resnet_eps,
groups=resnet_groups,
dropout=dropout,
time_embedding_norm=resnet_time_scale_shift,
non_linearity=resnet_act_fn,
output_scale_factor=output_scale_factor,
pre_norm=resnet_pre_norm,
)
)
self.attentions = nn.ModuleList(attentions)
self.resnets = nn.ModuleList(resnets)
def forward(self, hidden_states: torch.FloatTensor, temb: Optional[torch.FloatTensor] = None) -> torch.FloatTensor:
hidden_states = self.resnets[0](hidden_states, temb)
for attn, resnet in zip(self.attentions, self.resnets[1:]):
if attn is not None:
hidden_states = attn(hidden_states, temb=temb)
hidden_states = resnet(hidden_states, temb)
return hidden_states
class UNetMidBlock2DCrossAttn(nn.Module):
def __init__(
self,
in_channels: int,
temb_channels: int,
dropout: float = 0.0,
num_layers: int = 1,
transformer_layers_per_block: Union[int, Tuple[int]] = 1,
resnet_eps: float = 1e-6,
resnet_time_scale_shift: str = "default",
resnet_act_fn: str = "swish",
resnet_groups: int = 32,
resnet_pre_norm: bool = True,
num_attention_heads: int = 1,
output_scale_factor: float = 1.0,
cross_attention_dim: int = 1280,
dual_cross_attention: bool = False,
use_linear_projection: bool = False,
upcast_attention: bool = False,
attention_type: str = "default",
):
super().__init__()
self.has_cross_attention = True
self.num_attention_heads = num_attention_heads
resnet_groups = resnet_groups if resnet_groups is not None else min(in_channels // 4, 32)
# support for variable transformer layers per block
if isinstance(transformer_layers_per_block, int):
transformer_layers_per_block = [transformer_layers_per_block] * num_layers
# there is always at least one resnet
resnets = [
ResnetBlock2D(
in_channels=in_channels,
out_channels=in_channels,
temb_channels=temb_channels,
eps=resnet_eps,
groups=resnet_groups,
dropout=dropout,
time_embedding_norm=resnet_time_scale_shift,
non_linearity=resnet_act_fn,
output_scale_factor=output_scale_factor,
pre_norm=resnet_pre_norm,
)
]
attentions = []
for i in range(num_layers):
if not dual_cross_attention:
attentions.append(
Transformer2DModel(
num_attention_heads,
in_channels // num_attention_heads,
in_channels=in_channels,
num_layers=transformer_layers_per_block[i],
cross_attention_dim=cross_attention_dim,
norm_num_groups=resnet_groups,
use_linear_projection=use_linear_projection,
upcast_attention=upcast_attention,
attention_type=attention_type,
)
)
else:
attentions.append(
DualTransformer2DModel(
num_attention_heads,
in_channels // num_attention_heads,
in_channels=in_channels,
num_layers=1,
cross_attention_dim=cross_attention_dim,
norm_num_groups=resnet_groups,
)
)
resnets.append(
ResnetBlock2D(
in_channels=in_channels,
out_channels=in_channels,
temb_channels=temb_channels,
eps=resnet_eps,
groups=resnet_groups,
dropout=dropout,
time_embedding_norm=resnet_time_scale_shift,
non_linearity=resnet_act_fn,
output_scale_factor=output_scale_factor,
pre_norm=resnet_pre_norm,
)
)
self.attentions = nn.ModuleList(attentions)
self.resnets = nn.ModuleList(resnets)
self.gradient_checkpointing = False
def forward(
self,
hidden_states: torch.FloatTensor,
spatial_attn_inputs = [],
spatial_attn_idx = 0,
temb: Optional[torch.FloatTensor] = None,
encoder_hidden_states: Optional[torch.FloatTensor] = None,
attention_mask: Optional[torch.FloatTensor] = None,
cross_attention_kwargs: Optional[Dict[str, Any]] = None,
encoder_attention_mask: Optional[torch.FloatTensor] = None,
) -> torch.FloatTensor:
lora_scale = cross_attention_kwargs.get("scale", 1.0) if cross_attention_kwargs is not None else 1.0
hidden_states = self.resnets[0](hidden_states, temb, scale=lora_scale)
for attn, resnet in zip(self.attentions, self.resnets[1:]):
if self.training and self.gradient_checkpointing:
def create_custom_forward(module, return_dict=None):
def custom_forward(*inputs):
if return_dict is not None:
return module(*inputs, return_dict=return_dict)
else:
return module(*inputs)
return custom_forward
ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
hidden_states, spatial_attn_inputs, spatial_attn_idx = attn(
hidden_states,
spatial_attn_inputs=spatial_attn_inputs,
spatial_attn_idx=spatial_attn_idx,
encoder_hidden_states=encoder_hidden_states,
cross_attention_kwargs=cross_attention_kwargs,
attention_mask=attention_mask,
encoder_attention_mask=encoder_attention_mask,
return_dict=False,
)
hidden_states = hidden_states[0]
hidden_states = torch.utils.checkpoint.checkpoint(
create_custom_forward(resnet),
hidden_states,
temb,
**ckpt_kwargs,
)
else:
hidden_states, spatial_attn_inputs, spatial_attn_idx = attn(
hidden_states,
spatial_attn_inputs=spatial_attn_inputs,
spatial_attn_idx=spatial_attn_idx,
encoder_hidden_states=encoder_hidden_states,
cross_attention_kwargs=cross_attention_kwargs,
attention_mask=attention_mask,
encoder_attention_mask=encoder_attention_mask,
return_dict=False,
)
hidden_states = hidden_states[0]
hidden_states = resnet(hidden_states, temb, scale=lora_scale)
return hidden_states, spatial_attn_inputs, spatial_attn_idx
class UNetMidBlock2DSimpleCrossAttn(nn.Module):
def __init__(
self,
in_channels: int,
temb_channels: int,
dropout: float = 0.0,
num_layers: int = 1,
resnet_eps: float = 1e-6,
resnet_time_scale_shift: str = "default",
resnet_act_fn: str = "swish",
resnet_groups: int = 32,
resnet_pre_norm: bool = True,
attention_head_dim: int = 1,
output_scale_factor: float = 1.0,
cross_attention_dim: int = 1280,
skip_time_act: bool = False,
only_cross_attention: bool = False,
cross_attention_norm: Optional[str] = None,
):
super().__init__()
self.has_cross_attention = True
self.attention_head_dim = attention_head_dim
resnet_groups = resnet_groups if resnet_groups is not None else min(in_channels // 4, 32)
self.num_heads = in_channels // self.attention_head_dim
# there is always at least one resnet
resnets = [
ResnetBlock2D(
in_channels=in_channels,
out_channels=in_channels,
temb_channels=temb_channels,
eps=resnet_eps,
groups=resnet_groups,
dropout=dropout,
time_embedding_norm=resnet_time_scale_shift,
non_linearity=resnet_act_fn,
output_scale_factor=output_scale_factor,
pre_norm=resnet_pre_norm,
skip_time_act=skip_time_act,
)
]
attentions = []
for _ in range(num_layers):
processor = (
AttnAddedKVProcessor2_0() if hasattr(F, "scaled_dot_product_attention") else AttnAddedKVProcessor()
)
attentions.append(
Attention(
query_dim=in_channels,
cross_attention_dim=in_channels,
heads=self.num_heads,
dim_head=self.attention_head_dim,
added_kv_proj_dim=cross_attention_dim,
norm_num_groups=resnet_groups,
bias=True,
upcast_softmax=True,
only_cross_attention=only_cross_attention,
cross_attention_norm=cross_attention_norm,
processor=processor,
)
)
resnets.append(
ResnetBlock2D(
in_channels=in_channels,
out_channels=in_channels,
temb_channels=temb_channels,
eps=resnet_eps,
groups=resnet_groups,
dropout=dropout,
time_embedding_norm=resnet_time_scale_shift,
non_linearity=resnet_act_fn,
output_scale_factor=output_scale_factor,
pre_norm=resnet_pre_norm,
skip_time_act=skip_time_act,
)
)
self.attentions = nn.ModuleList(attentions)
self.resnets = nn.ModuleList(resnets)
def forward(
self,
hidden_states: torch.FloatTensor,
temb: Optional[torch.FloatTensor] = None,
encoder_hidden_states: Optional[torch.FloatTensor] = None,
attention_mask: Optional[torch.FloatTensor] = None,
cross_attention_kwargs: Optional[Dict[str, Any]] = None,
encoder_attention_mask: Optional[torch.FloatTensor] = None,
) -> torch.FloatTensor:
cross_attention_kwargs = cross_attention_kwargs if cross_attention_kwargs is not None else {}
lora_scale = cross_attention_kwargs.get("scale", 1.0)
if attention_mask is None:
# if encoder_hidden_states is defined: we are doing cross-attn, so we should use cross-attn mask.
mask = None if encoder_hidden_states is None else encoder_attention_mask
else:
# when attention_mask is defined: we don't even check for encoder_attention_mask.
# this is to maintain compatibility with UnCLIP, which uses 'attention_mask' param for cross-attn masks.
# TODO: UnCLIP should express cross-attn mask via encoder_attention_mask param instead of via attention_mask.
# then we can simplify this whole if/else block to:
# mask = attention_mask if encoder_hidden_states is None else encoder_attention_mask
mask = attention_mask
hidden_states = self.resnets[0](hidden_states, temb, scale=lora_scale)
for attn, resnet in zip(self.attentions, self.resnets[1:]):
# attn
hidden_states = attn(
hidden_states,
encoder_hidden_states=encoder_hidden_states,
attention_mask=mask,
**cross_attention_kwargs,
)
# resnet
hidden_states = resnet(hidden_states, temb, scale=lora_scale)
return hidden_states
class AttnDownBlock2D(nn.Module):
def __init__(
self,
in_channels: int,
out_channels: int,
temb_channels: int,
dropout: float = 0.0,
num_layers: int = 1,
resnet_eps: float = 1e-6,
resnet_time_scale_shift: str = "default",
resnet_act_fn: str = "swish",
resnet_groups: int = 32,
resnet_pre_norm: bool = True,
attention_head_dim: int = 1,
output_scale_factor: float = 1.0,
downsample_padding: int = 1,
downsample_type: str = "conv",
):
super().__init__()
resnets = []
attentions = []
self.downsample_type = downsample_type
if attention_head_dim is None:
logger.warn(
f"It is not recommend to pass `attention_head_dim=None`. Defaulting `attention_head_dim` to `in_channels`: {out_channels}."
)
attention_head_dim = out_channels
for i in range(num_layers):
in_channels = in_channels if i == 0 else out_channels
resnets.append(
ResnetBlock2D(
in_channels=in_channels,
out_channels=out_channels,
temb_channels=temb_channels,
eps=resnet_eps,
groups=resnet_groups,
dropout=dropout,
time_embedding_norm=resnet_time_scale_shift,
non_linearity=resnet_act_fn,
output_scale_factor=output_scale_factor,
pre_norm=resnet_pre_norm,
)
)
attentions.append(
Attention(
out_channels,
heads=out_channels // attention_head_dim,
dim_head=attention_head_dim,
rescale_output_factor=output_scale_factor,
eps=resnet_eps,
norm_num_groups=resnet_groups,
residual_connection=True,
bias=True,
upcast_softmax=True,
_from_deprecated_attn_block=True,
)
)
self.attentions = nn.ModuleList(attentions)
self.resnets = nn.ModuleList(resnets)
if downsample_type == "conv":
self.downsamplers = nn.ModuleList(
[
Downsample2D(
out_channels, use_conv=True, out_channels=out_channels, padding=downsample_padding, name="op"
)
]
)
elif downsample_type == "resnet":
self.downsamplers = nn.ModuleList(
[
ResnetBlock2D(
in_channels=out_channels,
out_channels=out_channels,
temb_channels=temb_channels,
eps=resnet_eps,
groups=resnet_groups,
dropout=dropout,
time_embedding_norm=resnet_time_scale_shift,
non_linearity=resnet_act_fn,
output_scale_factor=output_scale_factor,
pre_norm=resnet_pre_norm,
down=True,
)
]
)
else:
self.downsamplers = None
def forward(
self,
hidden_states: torch.FloatTensor,
temb: Optional[torch.FloatTensor] = None,
upsample_size: Optional[int] = None,
cross_attention_kwargs: Optional[Dict[str, Any]] = None,
) -> Tuple[torch.FloatTensor, Tuple[torch.FloatTensor, ...]]:
cross_attention_kwargs = cross_attention_kwargs if cross_attention_kwargs is not None else {}
lora_scale = cross_attention_kwargs.get("scale", 1.0)
output_states = ()
for resnet, attn in zip(self.resnets, self.attentions):
cross_attention_kwargs.update({"scale": lora_scale})
hidden_states = resnet(hidden_states, temb, scale=lora_scale)
hidden_states = attn(hidden_states, **cross_attention_kwargs)
output_states = output_states + (hidden_states,)
if self.downsamplers is not None:
for downsampler in self.downsamplers:
if self.downsample_type == "resnet":
hidden_states = downsampler(hidden_states, temb=temb, scale=lora_scale)
else:
hidden_states = downsampler(hidden_states, scale=lora_scale)
output_states += (hidden_states,)
return hidden_states, output_states
class CrossAttnDownBlock2D(nn.Module):
def __init__(
self,
in_channels: int,
out_channels: int,
temb_channels: int,
dropout: float = 0.0,
num_layers: int = 1,
transformer_layers_per_block: Union[int, Tuple[int]] = 1,
resnet_eps: float = 1e-6,
resnet_time_scale_shift: str = "default",
resnet_act_fn: str = "swish",
resnet_groups: int = 32,
resnet_pre_norm: bool = True,
num_attention_heads: int = 1,
cross_attention_dim: int = 1280,
output_scale_factor: float = 1.0,
downsample_padding: int = 1,
add_downsample: bool = True,
dual_cross_attention: bool = False,
use_linear_projection: bool = False,
only_cross_attention: bool = False,
upcast_attention: bool = False,
attention_type: str = "default",
):
super().__init__()
resnets = []
attentions = []
self.has_cross_attention = True
self.num_attention_heads = num_attention_heads
if isinstance(transformer_layers_per_block, int):
transformer_layers_per_block = [transformer_layers_per_block] * num_layers
for i in range(num_layers):
in_channels = in_channels if i == 0 else out_channels
resnets.append(
ResnetBlock2D(
in_channels=in_channels,
out_channels=out_channels,
temb_channels=temb_channels,
eps=resnet_eps,
groups=resnet_groups,
dropout=dropout,
time_embedding_norm=resnet_time_scale_shift,
non_linearity=resnet_act_fn,
output_scale_factor=output_scale_factor,
pre_norm=resnet_pre_norm,
)
)
if not dual_cross_attention:
attentions.append(
Transformer2DModel(
num_attention_heads,
out_channels // num_attention_heads,
in_channels=out_channels,
num_layers=transformer_layers_per_block[i],
cross_attention_dim=cross_attention_dim,
norm_num_groups=resnet_groups,
use_linear_projection=use_linear_projection,
only_cross_attention=only_cross_attention,
upcast_attention=upcast_attention,
attention_type=attention_type,
)
)
else:
attentions.append(
DualTransformer2DModel(
num_attention_heads,
out_channels // num_attention_heads,
in_channels=out_channels,
num_layers=1,
cross_attention_dim=cross_attention_dim,
norm_num_groups=resnet_groups,
)
)
self.attentions = nn.ModuleList(attentions)
self.resnets = nn.ModuleList(resnets)
if add_downsample:
self.downsamplers = nn.ModuleList(
[
Downsample2D(
out_channels, use_conv=True, out_channels=out_channels, padding=downsample_padding, name="op"
)
]
)
else:
self.downsamplers = None
self.gradient_checkpointing = False
def forward(
self,
hidden_states: torch.FloatTensor,
spatial_attn_inputs = [],
spatial_attn_idx = 0,
temb: Optional[torch.FloatTensor] = None,
encoder_hidden_states: Optional[torch.FloatTensor] = None,
attention_mask: Optional[torch.FloatTensor] = None,
cross_attention_kwargs: Optional[Dict[str, Any]] = None,
encoder_attention_mask: Optional[torch.FloatTensor] = None,
additional_residuals: Optional[torch.FloatTensor] = None,
) -> Tuple[torch.FloatTensor, Tuple[torch.FloatTensor, ...]]:
output_states = ()
lora_scale = cross_attention_kwargs.get("scale", 1.0) if cross_attention_kwargs is not None else 1.0
blocks = list(zip(self.resnets, self.attentions))
for i, (resnet, attn) in enumerate(blocks):
if self.training and self.gradient_checkpointing:
def create_custom_forward(module, return_dict=None):
def custom_forward(*inputs):
if return_dict is not None:
return module(*inputs, return_dict=return_dict)
else:
return module(*inputs)
return custom_forward
ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
hidden_states = torch.utils.checkpoint.checkpoint(
create_custom_forward(resnet),
hidden_states,
temb,
**ckpt_kwargs,
)
hidden_states, spatial_attn_inputs, spatial_attn_idx = attn(
hidden_states,
spatial_attn_inputs=spatial_attn_inputs,
spatial_attn_idx=spatial_attn_idx,
encoder_hidden_states=encoder_hidden_states,
cross_attention_kwargs=cross_attention_kwargs,
attention_mask=attention_mask,
encoder_attention_mask=encoder_attention_mask,
return_dict=False,
)
hidden_states = hidden_states[0]
else:
hidden_states = resnet(hidden_states, temb, scale=lora_scale)
hidden_states, spatial_attn_inputs, spatial_attn_idx = attn(
hidden_states,
spatial_attn_inputs=spatial_attn_inputs,
spatial_attn_idx=spatial_attn_idx,
encoder_hidden_states=encoder_hidden_states,
cross_attention_kwargs=cross_attention_kwargs,
attention_mask=attention_mask,
encoder_attention_mask=encoder_attention_mask,
return_dict=False,
)
hidden_states = hidden_states[0]
# apply additional residuals to the output of the last pair of resnet and attention blocks
if i == len(blocks) - 1 and additional_residuals is not None:
hidden_states = hidden_states + additional_residuals
output_states = output_states + (hidden_states,)
if self.downsamplers is not None:
for downsampler in self.downsamplers:
hidden_states = downsampler(hidden_states, scale=lora_scale)
output_states = output_states + (hidden_states,)
return hidden_states, output_states, spatial_attn_inputs, spatial_attn_idx
class DownBlock2D(nn.Module):
def __init__(
self,
in_channels: int,
out_channels: int,
temb_channels: int,
dropout: float = 0.0,
num_layers: int = 1,
resnet_eps: float = 1e-6,
resnet_time_scale_shift: str = "default",
resnet_act_fn: str = "swish",
resnet_groups: int = 32,
resnet_pre_norm: bool = True,
output_scale_factor: float = 1.0,
add_downsample: bool = True,
downsample_padding: int = 1,
):
super().__init__()
resnets = []
for i in range(num_layers):
in_channels = in_channels if i == 0 else out_channels
resnets.append(
ResnetBlock2D(
in_channels=in_channels,
out_channels=out_channels,
temb_channels=temb_channels,
eps=resnet_eps,
groups=resnet_groups,
dropout=dropout,
time_embedding_norm=resnet_time_scale_shift,
non_linearity=resnet_act_fn,
output_scale_factor=output_scale_factor,
pre_norm=resnet_pre_norm,
)
)
self.resnets = nn.ModuleList(resnets)
if add_downsample:
self.downsamplers = nn.ModuleList(
[
Downsample2D(
out_channels, use_conv=True, out_channels=out_channels, padding=downsample_padding, name="op"
)
]
)
else:
self.downsamplers = None
self.gradient_checkpointing = False
def forward(
self, hidden_states: torch.FloatTensor, temb: Optional[torch.FloatTensor] = None, scale: float = 1.0
) -> Tuple[torch.FloatTensor, Tuple[torch.FloatTensor, ...]]:
output_states = ()
for resnet in self.resnets:
if self.training and self.gradient_checkpointing:
def create_custom_forward(module):
def custom_forward(*inputs):
return module(*inputs)
return custom_forward
if is_torch_version(">=", "1.11.0"):
hidden_states = torch.utils.checkpoint.checkpoint(
create_custom_forward(resnet), hidden_states, temb, use_reentrant=False
)
else:
hidden_states = torch.utils.checkpoint.checkpoint(
create_custom_forward(resnet), hidden_states, temb
)
else:
hidden_states = resnet(hidden_states, temb, scale=scale)
output_states = output_states + (hidden_states,)
if self.downsamplers is not None:
for downsampler in self.downsamplers:
hidden_states = downsampler(hidden_states, scale=scale)
output_states = output_states + (hidden_states,)
return hidden_states, output_states
class DownEncoderBlock2D(nn.Module):
def __init__(
self,
in_channels: int,
out_channels: int,
dropout: float = 0.0,
num_layers: int = 1,
resnet_eps: float = 1e-6,
resnet_time_scale_shift: str = "default",
resnet_act_fn: str = "swish",
resnet_groups: int = 32,
resnet_pre_norm: bool = True,
output_scale_factor: float = 1.0,
add_downsample: bool = True,
downsample_padding: int = 1,
):
super().__init__()
resnets = []
for i in range(num_layers):
in_channels = in_channels if i == 0 else out_channels
resnets.append(
ResnetBlock2D(
in_channels=in_channels,
out_channels=out_channels,
temb_channels=None,
eps=resnet_eps,
groups=resnet_groups,
dropout=dropout,
time_embedding_norm=resnet_time_scale_shift,
non_linearity=resnet_act_fn,
output_scale_factor=output_scale_factor,
pre_norm=resnet_pre_norm,
)
)
self.resnets = nn.ModuleList(resnets)
if add_downsample:
self.downsamplers = nn.ModuleList(
[
Downsample2D(
out_channels, use_conv=True, out_channels=out_channels, padding=downsample_padding, name="op"
)
]
)
else:
self.downsamplers = None
def forward(self, hidden_states: torch.FloatTensor, scale: float = 1.0) -> torch.FloatTensor:
for resnet in self.resnets:
hidden_states = resnet(hidden_states, temb=None, scale=scale)
if self.downsamplers is not None:
for downsampler in self.downsamplers:
hidden_states = downsampler(hidden_states, scale)
return hidden_states
class AttnDownEncoderBlock2D(nn.Module):
def __init__(
self,
in_channels: int,
out_channels: int,
dropout: float = 0.0,
num_layers: int = 1,
resnet_eps: float = 1e-6,
resnet_time_scale_shift: str = "default",
resnet_act_fn: str = "swish",
resnet_groups: int = 32,
resnet_pre_norm: bool = True,
attention_head_dim: int = 1,
output_scale_factor: float = 1.0,
add_downsample: bool = True,
downsample_padding: int = 1,
):
super().__init__()
resnets = []
attentions = []
if attention_head_dim is None:
logger.warn(
f"It is not recommend to pass `attention_head_dim=None`. Defaulting `attention_head_dim` to `in_channels`: {out_channels}."
)
attention_head_dim = out_channels
for i in range(num_layers):
in_channels = in_channels if i == 0 else out_channels
resnets.append(
ResnetBlock2D(
in_channels=in_channels,
out_channels=out_channels,
temb_channels=None,
eps=resnet_eps,
groups=resnet_groups,
dropout=dropout,
time_embedding_norm=resnet_time_scale_shift,
non_linearity=resnet_act_fn,
output_scale_factor=output_scale_factor,
pre_norm=resnet_pre_norm,
)
)
attentions.append(
Attention(
out_channels,
heads=out_channels // attention_head_dim,
dim_head=attention_head_dim,
rescale_output_factor=output_scale_factor,
eps=resnet_eps,
norm_num_groups=resnet_groups,
residual_connection=True,
bias=True,
upcast_softmax=True,
_from_deprecated_attn_block=True,
)
)
self.attentions = nn.ModuleList(attentions)
self.resnets = nn.ModuleList(resnets)
if add_downsample:
self.downsamplers = nn.ModuleList(
[
Downsample2D(
out_channels, use_conv=True, out_channels=out_channels, padding=downsample_padding, name="op"
)
]
)
else:
self.downsamplers = None
def forward(self, hidden_states: torch.FloatTensor, scale: float = 1.0) -> torch.FloatTensor:
for resnet, attn in zip(self.resnets, self.attentions):
hidden_states = resnet(hidden_states, temb=None, scale=scale)
cross_attention_kwargs = {"scale": scale}
hidden_states = attn(hidden_states, **cross_attention_kwargs)
if self.downsamplers is not None:
for downsampler in self.downsamplers:
hidden_states = downsampler(hidden_states, scale)
return hidden_states
class AttnSkipDownBlock2D(nn.Module):
def __init__(
self,
in_channels: int,
out_channels: int,
temb_channels: int,
dropout: float = 0.0,
num_layers: int = 1,
resnet_eps: float = 1e-6,
resnet_time_scale_shift: str = "default",
resnet_act_fn: str = "swish",
resnet_pre_norm: bool = True,
attention_head_dim: int = 1,
output_scale_factor: float = np.sqrt(2.0),
add_downsample: bool = True,
):
super().__init__()
self.attentions = nn.ModuleList([])
self.resnets = nn.ModuleList([])
if attention_head_dim is None:
logger.warn(
f"It is not recommend to pass `attention_head_dim=None`. Defaulting `attention_head_dim` to `in_channels`: {out_channels}."
)
attention_head_dim = out_channels
for i in range(num_layers):
in_channels = in_channels if i == 0 else out_channels
self.resnets.append(
ResnetBlock2D(
in_channels=in_channels,
out_channels=out_channels,
temb_channels=temb_channels,
eps=resnet_eps,
groups=min(in_channels // 4, 32),
groups_out=min(out_channels // 4, 32),
dropout=dropout,
time_embedding_norm=resnet_time_scale_shift,
non_linearity=resnet_act_fn,
output_scale_factor=output_scale_factor,
pre_norm=resnet_pre_norm,
)
)
self.attentions.append(
Attention(
out_channels,
heads=out_channels // attention_head_dim,
dim_head=attention_head_dim,
rescale_output_factor=output_scale_factor,
eps=resnet_eps,
norm_num_groups=32,
residual_connection=True,
bias=True,
upcast_softmax=True,
_from_deprecated_attn_block=True,
)
)
if add_downsample:
self.resnet_down = ResnetBlock2D(
in_channels=out_channels,
out_channels=out_channels,
temb_channels=temb_channels,
eps=resnet_eps,
groups=min(out_channels // 4, 32),
dropout=dropout,
time_embedding_norm=resnet_time_scale_shift,
non_linearity=resnet_act_fn,
output_scale_factor=output_scale_factor,
pre_norm=resnet_pre_norm,
use_in_shortcut=True,
down=True,
kernel="fir",
)
self.downsamplers = nn.ModuleList([FirDownsample2D(out_channels, out_channels=out_channels)])
self.skip_conv = nn.Conv2d(3, out_channels, kernel_size=(1, 1), stride=(1, 1))
else:
self.resnet_down = None
self.downsamplers = None
self.skip_conv = None
def forward(
self,
hidden_states: torch.FloatTensor,
temb: Optional[torch.FloatTensor] = None,
skip_sample: Optional[torch.FloatTensor] = None,
scale: float = 1.0,
) -> Tuple[torch.FloatTensor, Tuple[torch.FloatTensor, ...], torch.FloatTensor]:
output_states = ()
for resnet, attn in zip(self.resnets, self.attentions):
hidden_states = resnet(hidden_states, temb, scale=scale)
cross_attention_kwargs = {"scale": scale}
hidden_states = attn(hidden_states, **cross_attention_kwargs)
output_states += (hidden_states,)
if self.downsamplers is not None:
hidden_states = self.resnet_down(hidden_states, temb, scale=scale)
for downsampler in self.downsamplers:
skip_sample = downsampler(skip_sample)
hidden_states = self.skip_conv(skip_sample) + hidden_states
output_states += (hidden_states,)
return hidden_states, output_states, skip_sample
class SkipDownBlock2D(nn.Module):
def __init__(
self,
in_channels: int,
out_channels: int,
temb_channels: int,
dropout: float = 0.0,
num_layers: int = 1,
resnet_eps: float = 1e-6,
resnet_time_scale_shift: str = "default",
resnet_act_fn: str = "swish",
resnet_pre_norm: bool = True,
output_scale_factor: float = np.sqrt(2.0),
add_downsample: bool = True,
downsample_padding: int = 1,
):
super().__init__()
self.resnets = nn.ModuleList([])
for i in range(num_layers):
in_channels = in_channels if i == 0 else out_channels
self.resnets.append(
ResnetBlock2D(
in_channels=in_channels,
out_channels=out_channels,
temb_channels=temb_channels,
eps=resnet_eps,
groups=min(in_channels // 4, 32),
groups_out=min(out_channels // 4, 32),
dropout=dropout,
time_embedding_norm=resnet_time_scale_shift,
non_linearity=resnet_act_fn,
output_scale_factor=output_scale_factor,
pre_norm=resnet_pre_norm,
)
)
if add_downsample:
self.resnet_down = ResnetBlock2D(
in_channels=out_channels,
out_channels=out_channels,
temb_channels=temb_channels,
eps=resnet_eps,
groups=min(out_channels // 4, 32),
dropout=dropout,
time_embedding_norm=resnet_time_scale_shift,
non_linearity=resnet_act_fn,
output_scale_factor=output_scale_factor,
pre_norm=resnet_pre_norm,
use_in_shortcut=True,
down=True,
kernel="fir",
)
self.downsamplers = nn.ModuleList([FirDownsample2D(out_channels, out_channels=out_channels)])
self.skip_conv = nn.Conv2d(3, out_channels, kernel_size=(1, 1), stride=(1, 1))
else:
self.resnet_down = None
self.downsamplers = None
self.skip_conv = None
def forward(
self,
hidden_states: torch.FloatTensor,
temb: Optional[torch.FloatTensor] = None,
skip_sample: Optional[torch.FloatTensor] = None,
scale: float = 1.0,
) -> Tuple[torch.FloatTensor, Tuple[torch.FloatTensor, ...], torch.FloatTensor]:
output_states = ()
for resnet in self.resnets:
hidden_states = resnet(hidden_states, temb, scale)
output_states += (hidden_states,)
if self.downsamplers is not None:
hidden_states = self.resnet_down(hidden_states, temb, scale)
for downsampler in self.downsamplers:
skip_sample = downsampler(skip_sample)
hidden_states = self.skip_conv(skip_sample) + hidden_states
output_states += (hidden_states,)
return hidden_states, output_states, skip_sample
class ResnetDownsampleBlock2D(nn.Module):
def __init__(
self,
in_channels: int,
out_channels: int,
temb_channels: int,
dropout: float = 0.0,
num_layers: int = 1,
resnet_eps: float = 1e-6,
resnet_time_scale_shift: str = "default",
resnet_act_fn: str = "swish",
resnet_groups: int = 32,
resnet_pre_norm: bool = True,
output_scale_factor: float = 1.0,
add_downsample: bool = True,
skip_time_act: bool = False,
):
super().__init__()
resnets = []
for i in range(num_layers):
in_channels = in_channels if i == 0 else out_channels
resnets.append(
ResnetBlock2D(
in_channels=in_channels,
out_channels=out_channels,
temb_channels=temb_channels,
eps=resnet_eps,
groups=resnet_groups,
dropout=dropout,
time_embedding_norm=resnet_time_scale_shift,
non_linearity=resnet_act_fn,
output_scale_factor=output_scale_factor,
pre_norm=resnet_pre_norm,
skip_time_act=skip_time_act,
)
)
self.resnets = nn.ModuleList(resnets)
if add_downsample:
self.downsamplers = nn.ModuleList(
[
ResnetBlock2D(
in_channels=out_channels,
out_channels=out_channels,
temb_channels=temb_channels,
eps=resnet_eps,
groups=resnet_groups,
dropout=dropout,
time_embedding_norm=resnet_time_scale_shift,
non_linearity=resnet_act_fn,
output_scale_factor=output_scale_factor,
pre_norm=resnet_pre_norm,
skip_time_act=skip_time_act,
down=True,
)
]
)
else:
self.downsamplers = None
self.gradient_checkpointing = False
def forward(
self, hidden_states: torch.FloatTensor, temb: Optional[torch.FloatTensor] = None, scale: float = 1.0
) -> Tuple[torch.FloatTensor, Tuple[torch.FloatTensor, ...]]:
output_states = ()
for resnet in self.resnets:
if self.training and self.gradient_checkpointing:
def create_custom_forward(module):
def custom_forward(*inputs):
return module(*inputs)
return custom_forward
if is_torch_version(">=", "1.11.0"):
hidden_states = torch.utils.checkpoint.checkpoint(
create_custom_forward(resnet), hidden_states, temb, use_reentrant=False
)
else:
hidden_states = torch.utils.checkpoint.checkpoint(
create_custom_forward(resnet), hidden_states, temb
)
else:
hidden_states = resnet(hidden_states, temb, scale)
output_states = output_states + (hidden_states,)
if self.downsamplers is not None:
for downsampler in self.downsamplers:
hidden_states = downsampler(hidden_states, temb, scale)
output_states = output_states + (hidden_states,)
return hidden_states, output_states
class SimpleCrossAttnDownBlock2D(nn.Module):
def __init__(
self,
in_channels: int,
out_channels: int,
temb_channels: int,
dropout: float = 0.0,
num_layers: int = 1,
resnet_eps: float = 1e-6,
resnet_time_scale_shift: str = "default",
resnet_act_fn: str = "swish",
resnet_groups: int = 32,
resnet_pre_norm: bool = True,
attention_head_dim: int = 1,
cross_attention_dim: int = 1280,
output_scale_factor: float = 1.0,
add_downsample: bool = True,
skip_time_act: bool = False,
only_cross_attention: bool = False,
cross_attention_norm: Optional[str] = None,
):
super().__init__()
self.has_cross_attention = True
resnets = []
attentions = []
self.attention_head_dim = attention_head_dim
self.num_heads = out_channels // self.attention_head_dim
for i in range(num_layers):
in_channels = in_channels if i == 0 else out_channels
resnets.append(
ResnetBlock2D(
in_channels=in_channels,
out_channels=out_channels,
temb_channels=temb_channels,
eps=resnet_eps,
groups=resnet_groups,
dropout=dropout,
time_embedding_norm=resnet_time_scale_shift,
non_linearity=resnet_act_fn,
output_scale_factor=output_scale_factor,
pre_norm=resnet_pre_norm,
skip_time_act=skip_time_act,
)
)
processor = (
AttnAddedKVProcessor2_0() if hasattr(F, "scaled_dot_product_attention") else AttnAddedKVProcessor()
)
attentions.append(
Attention(
query_dim=out_channels,
cross_attention_dim=out_channels,
heads=self.num_heads,
dim_head=attention_head_dim,
added_kv_proj_dim=cross_attention_dim,
norm_num_groups=resnet_groups,
bias=True,
upcast_softmax=True,
only_cross_attention=only_cross_attention,
cross_attention_norm=cross_attention_norm,
processor=processor,
)
)
self.attentions = nn.ModuleList(attentions)
self.resnets = nn.ModuleList(resnets)
if add_downsample:
self.downsamplers = nn.ModuleList(
[
ResnetBlock2D(
in_channels=out_channels,
out_channels=out_channels,
temb_channels=temb_channels,
eps=resnet_eps,
groups=resnet_groups,
dropout=dropout,
time_embedding_norm=resnet_time_scale_shift,
non_linearity=resnet_act_fn,
output_scale_factor=output_scale_factor,
pre_norm=resnet_pre_norm,
skip_time_act=skip_time_act,
down=True,
)
]
)
else:
self.downsamplers = None
self.gradient_checkpointing = False
def forward(
self,
hidden_states: torch.FloatTensor,
temb: Optional[torch.FloatTensor] = None,
encoder_hidden_states: Optional[torch.FloatTensor] = None,
attention_mask: Optional[torch.FloatTensor] = None,
cross_attention_kwargs: Optional[Dict[str, Any]] = None,
encoder_attention_mask: Optional[torch.FloatTensor] = None,
) -> Tuple[torch.FloatTensor, Tuple[torch.FloatTensor, ...]]:
output_states = ()
cross_attention_kwargs = cross_attention_kwargs if cross_attention_kwargs is not None else {}
lora_scale = cross_attention_kwargs.get("scale", 1.0)
if attention_mask is None:
# if encoder_hidden_states is defined: we are doing cross-attn, so we should use cross-attn mask.
mask = None if encoder_hidden_states is None else encoder_attention_mask
else:
# when attention_mask is defined: we don't even check for encoder_attention_mask.
# this is to maintain compatibility with UnCLIP, which uses 'attention_mask' param for cross-attn masks.
# TODO: UnCLIP should express cross-attn mask via encoder_attention_mask param instead of via attention_mask.
# then we can simplify this whole if/else block to:
# mask = attention_mask if encoder_hidden_states is None else encoder_attention_mask
mask = attention_mask
for resnet, attn in zip(self.resnets, self.attentions):
if self.training and self.gradient_checkpointing:
def create_custom_forward(module, return_dict=None):
def custom_forward(*inputs):
if return_dict is not None:
return module(*inputs, return_dict=return_dict)
else:
return module(*inputs)
return custom_forward
hidden_states = torch.utils.checkpoint.checkpoint(create_custom_forward(resnet), hidden_states, temb)
hidden_states = attn(
hidden_states,
encoder_hidden_states=encoder_hidden_states,
attention_mask=mask,
**cross_attention_kwargs,
)
else:
hidden_states = resnet(hidden_states, temb, scale=lora_scale)
hidden_states = attn(
hidden_states,
encoder_hidden_states=encoder_hidden_states,
attention_mask=mask,
**cross_attention_kwargs,
)
output_states = output_states + (hidden_states,)
if self.downsamplers is not None:
for downsampler in self.downsamplers:
hidden_states = downsampler(hidden_states, temb, scale=lora_scale)
output_states = output_states + (hidden_states,)
return hidden_states, output_states
class KDownBlock2D(nn.Module):
def __init__(
self,
in_channels: int,
out_channels: int,
temb_channels: int,
dropout: float = 0.0,
num_layers: int = 4,
resnet_eps: float = 1e-5,
resnet_act_fn: str = "gelu",
resnet_group_size: int = 32,
add_downsample: bool = False,
):
super().__init__()
resnets = []
for i in range(num_layers):
in_channels = in_channels if i == 0 else out_channels
groups = in_channels // resnet_group_size
groups_out = out_channels // resnet_group_size
resnets.append(
ResnetBlock2D(
in_channels=in_channels,
out_channels=out_channels,
dropout=dropout,
temb_channels=temb_channels,
groups=groups,
groups_out=groups_out,
eps=resnet_eps,
non_linearity=resnet_act_fn,
time_embedding_norm="ada_group",
conv_shortcut_bias=False,
)
)
self.resnets = nn.ModuleList(resnets)
if add_downsample:
# YiYi's comments- might be able to use FirDownsample2D, look into details later
self.downsamplers = nn.ModuleList([KDownsample2D()])
else:
self.downsamplers = None
self.gradient_checkpointing = False
def forward(
self, hidden_states: torch.FloatTensor, temb: Optional[torch.FloatTensor] = None, scale: float = 1.0
) -> Tuple[torch.FloatTensor, Tuple[torch.FloatTensor, ...]]:
output_states = ()
for resnet in self.resnets:
if self.training and self.gradient_checkpointing:
def create_custom_forward(module):
def custom_forward(*inputs):
return module(*inputs)
return custom_forward
if is_torch_version(">=", "1.11.0"):
hidden_states = torch.utils.checkpoint.checkpoint(
create_custom_forward(resnet), hidden_states, temb, use_reentrant=False
)
else:
hidden_states = torch.utils.checkpoint.checkpoint(
create_custom_forward(resnet), hidden_states, temb
)
else:
hidden_states = resnet(hidden_states, temb, scale)
output_states += (hidden_states,)
if self.downsamplers is not None:
for downsampler in self.downsamplers:
hidden_states = downsampler(hidden_states)
return hidden_states, output_states
class KCrossAttnDownBlock2D(nn.Module):
def __init__(
self,
in_channels: int,
out_channels: int,
temb_channels: int,
cross_attention_dim: int,
dropout: float = 0.0,
num_layers: int = 4,
resnet_group_size: int = 32,
add_downsample: bool = True,
attention_head_dim: int = 64,
add_self_attention: bool = False,
resnet_eps: float = 1e-5,
resnet_act_fn: str = "gelu",
):
super().__init__()
resnets = []
attentions = []
self.has_cross_attention = True
for i in range(num_layers):
in_channels = in_channels if i == 0 else out_channels
groups = in_channels // resnet_group_size
groups_out = out_channels // resnet_group_size
resnets.append(
ResnetBlock2D(
in_channels=in_channels,
out_channels=out_channels,
dropout=dropout,
temb_channels=temb_channels,
groups=groups,
groups_out=groups_out,
eps=resnet_eps,
non_linearity=resnet_act_fn,
time_embedding_norm="ada_group",
conv_shortcut_bias=False,
)
)
attentions.append(
KAttentionBlock(
out_channels,
out_channels // attention_head_dim,
attention_head_dim,
cross_attention_dim=cross_attention_dim,
temb_channels=temb_channels,
attention_bias=True,
add_self_attention=add_self_attention,
cross_attention_norm="layer_norm",
group_size=resnet_group_size,
)
)
self.resnets = nn.ModuleList(resnets)
self.attentions = nn.ModuleList(attentions)
if add_downsample:
self.downsamplers = nn.ModuleList([KDownsample2D()])
else:
self.downsamplers = None
self.gradient_checkpointing = False
def forward(
self,
hidden_states: torch.FloatTensor,
temb: Optional[torch.FloatTensor] = None,
encoder_hidden_states: Optional[torch.FloatTensor] = None,
attention_mask: Optional[torch.FloatTensor] = None,
cross_attention_kwargs: Optional[Dict[str, Any]] = None,
encoder_attention_mask: Optional[torch.FloatTensor] = None,
) -> Tuple[torch.FloatTensor, Tuple[torch.FloatTensor, ...]]:
output_states = ()
lora_scale = cross_attention_kwargs.get("scale", 1.0) if cross_attention_kwargs is not None else 1.0
for resnet, attn in zip(self.resnets, self.attentions):
if self.training and self.gradient_checkpointing:
def create_custom_forward(module, return_dict=None):
def custom_forward(*inputs):
if return_dict is not None:
return module(*inputs, return_dict=return_dict)
else:
return module(*inputs)
return custom_forward
ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
hidden_states = torch.utils.checkpoint.checkpoint(
create_custom_forward(resnet),
hidden_states,
temb,
**ckpt_kwargs,
)
hidden_states = attn(
hidden_states,
encoder_hidden_states=encoder_hidden_states,
emb=temb,
attention_mask=attention_mask,
cross_attention_kwargs=cross_attention_kwargs,
encoder_attention_mask=encoder_attention_mask,
)
else:
hidden_states = resnet(hidden_states, temb, scale=lora_scale)
hidden_states = attn(
hidden_states,
encoder_hidden_states=encoder_hidden_states,
emb=temb,
attention_mask=attention_mask,
cross_attention_kwargs=cross_attention_kwargs,
encoder_attention_mask=encoder_attention_mask,
)
if self.downsamplers is None:
output_states += (None,)
else:
output_states += (hidden_states,)
if self.downsamplers is not None:
for downsampler in self.downsamplers:
hidden_states = downsampler(hidden_states)
return hidden_states, output_states
class AttnUpBlock2D(nn.Module):
def __init__(
self,
in_channels: int,
prev_output_channel: int,
out_channels: int,
temb_channels: int,
resolution_idx: int = None,
dropout: float = 0.0,
num_layers: int = 1,
resnet_eps: float = 1e-6,
resnet_time_scale_shift: str = "default",
resnet_act_fn: str = "swish",
resnet_groups: int = 32,
resnet_pre_norm: bool = True,
attention_head_dim: int = 1,
output_scale_factor: float = 1.0,
upsample_type: str = "conv",
):
super().__init__()
resnets = []
attentions = []
self.upsample_type = upsample_type
if attention_head_dim is None:
logger.warn(
f"It is not recommend to pass `attention_head_dim=None`. Defaulting `attention_head_dim` to `in_channels`: {out_channels}."
)
attention_head_dim = out_channels
for i in range(num_layers):
res_skip_channels = in_channels if (i == num_layers - 1) else out_channels
resnet_in_channels = prev_output_channel if i == 0 else out_channels
resnets.append(
ResnetBlock2D(
in_channels=resnet_in_channels + res_skip_channels,
out_channels=out_channels,
temb_channels=temb_channels,
eps=resnet_eps,
groups=resnet_groups,
dropout=dropout,
time_embedding_norm=resnet_time_scale_shift,
non_linearity=resnet_act_fn,
output_scale_factor=output_scale_factor,
pre_norm=resnet_pre_norm,
)
)
attentions.append(
Attention(
out_channels,
heads=out_channels // attention_head_dim,
dim_head=attention_head_dim,
rescale_output_factor=output_scale_factor,
eps=resnet_eps,
norm_num_groups=resnet_groups,
residual_connection=True,
bias=True,
upcast_softmax=True,
_from_deprecated_attn_block=True,
)
)
self.attentions = nn.ModuleList(attentions)
self.resnets = nn.ModuleList(resnets)
if upsample_type == "conv":
self.upsamplers = nn.ModuleList([Upsample2D(out_channels, use_conv=True, out_channels=out_channels)])
elif upsample_type == "resnet":
self.upsamplers = nn.ModuleList(
[
ResnetBlock2D(
in_channels=out_channels,
out_channels=out_channels,
temb_channels=temb_channels,
eps=resnet_eps,
groups=resnet_groups,
dropout=dropout,
time_embedding_norm=resnet_time_scale_shift,
non_linearity=resnet_act_fn,
output_scale_factor=output_scale_factor,
pre_norm=resnet_pre_norm,
up=True,
)
]
)
else:
self.upsamplers = None
self.resolution_idx = resolution_idx
def forward(
self,
hidden_states: torch.FloatTensor,
res_hidden_states_tuple: Tuple[torch.FloatTensor, ...],
temb: Optional[torch.FloatTensor] = None,
upsample_size: Optional[int] = None,
scale: float = 1.0,
) -> torch.FloatTensor:
for resnet, attn in zip(self.resnets, self.attentions):
# pop res hidden states
res_hidden_states = res_hidden_states_tuple[-1]
res_hidden_states_tuple = res_hidden_states_tuple[:-1]
hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)
hidden_states = resnet(hidden_states, temb, scale=scale)
cross_attention_kwargs = {"scale": scale}
hidden_states = attn(hidden_states, **cross_attention_kwargs)
if self.upsamplers is not None:
for upsampler in self.upsamplers:
if self.upsample_type == "resnet":
hidden_states = upsampler(hidden_states, temb=temb, scale=scale)
else:
hidden_states = upsampler(hidden_states, scale=scale)
return hidden_states
class CrossAttnUpBlock2D(nn.Module):
def __init__(
self,
in_channels: int,
out_channels: int,
prev_output_channel: int,
temb_channels: int,
resolution_idx: Optional[int] = None,
dropout: float = 0.0,
num_layers: int = 1,
transformer_layers_per_block: Union[int, Tuple[int]] = 1,
resnet_eps: float = 1e-6,
resnet_time_scale_shift: str = "default",
resnet_act_fn: str = "swish",
resnet_groups: int = 32,
resnet_pre_norm: bool = True,
num_attention_heads: int = 1,
cross_attention_dim: int = 1280,
output_scale_factor: float = 1.0,
add_upsample: bool = True,
dual_cross_attention: bool = False,
use_linear_projection: bool = False,
only_cross_attention: bool = False,
upcast_attention: bool = False,
attention_type: str = "default",
):
super().__init__()
resnets = []
attentions = []
self.has_cross_attention = True
self.num_attention_heads = num_attention_heads
if isinstance(transformer_layers_per_block, int):
transformer_layers_per_block = [transformer_layers_per_block] * num_layers
for i in range(num_layers):
res_skip_channels = in_channels if (i == num_layers - 1) else out_channels
resnet_in_channels = prev_output_channel if i == 0 else out_channels
resnets.append(
ResnetBlock2D(
in_channels=resnet_in_channels + res_skip_channels,
out_channels=out_channels,
temb_channels=temb_channels,
eps=resnet_eps,
groups=resnet_groups,
dropout=dropout,
time_embedding_norm=resnet_time_scale_shift,
non_linearity=resnet_act_fn,
output_scale_factor=output_scale_factor,
pre_norm=resnet_pre_norm,
)
)
if not dual_cross_attention:
attentions.append(
Transformer2DModel(
num_attention_heads,
out_channels // num_attention_heads,
in_channels=out_channels,
num_layers=transformer_layers_per_block[i],
cross_attention_dim=cross_attention_dim,
norm_num_groups=resnet_groups,
use_linear_projection=use_linear_projection,
only_cross_attention=only_cross_attention,
upcast_attention=upcast_attention,
attention_type=attention_type,
)
)
else:
attentions.append(
DualTransformer2DModel(
num_attention_heads,
out_channels // num_attention_heads,
in_channels=out_channels,
num_layers=1,
cross_attention_dim=cross_attention_dim,
norm_num_groups=resnet_groups,
)
)
self.attentions = nn.ModuleList(attentions)
self.resnets = nn.ModuleList(resnets)
if add_upsample:
self.upsamplers = nn.ModuleList([Upsample2D(out_channels, use_conv=True, out_channels=out_channels)])
else:
self.upsamplers = None
self.gradient_checkpointing = False
self.resolution_idx = resolution_idx
def forward(
self,
hidden_states: torch.FloatTensor,
res_hidden_states_tuple: Tuple[torch.FloatTensor, ...],
spatial_attn_inputs = [],
spatial_attn_idx = 0,
temb: Optional[torch.FloatTensor] = None,
encoder_hidden_states: Optional[torch.FloatTensor] = None,
cross_attention_kwargs: Optional[Dict[str, Any]] = None,
upsample_size: Optional[int] = None,
attention_mask: Optional[torch.FloatTensor] = None,
encoder_attention_mask: Optional[torch.FloatTensor] = None,
) -> torch.FloatTensor:
lora_scale = cross_attention_kwargs.get("scale", 1.0) if cross_attention_kwargs is not None else 1.0
is_freeu_enabled = (
getattr(self, "s1", None)
and getattr(self, "s2", None)
and getattr(self, "b1", None)
and getattr(self, "b2", None)
)
for resnet, attn in zip(self.resnets, self.attentions):
# pop res hidden states
res_hidden_states = res_hidden_states_tuple[-1]
res_hidden_states_tuple = res_hidden_states_tuple[:-1]
# FreeU: Only operate on the first two stages
if is_freeu_enabled:
hidden_states, res_hidden_states = apply_freeu(
self.resolution_idx,
hidden_states,
res_hidden_states,
s1=self.s1,
s2=self.s2,
b1=self.b1,
b2=self.b2,
)
hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)
if self.training and self.gradient_checkpointing:
def create_custom_forward(module, return_dict=None):
def custom_forward(*inputs):
if return_dict is not None:
return module(*inputs, return_dict=return_dict)
else:
return module(*inputs)
return custom_forward
ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
hidden_states = torch.utils.checkpoint.checkpoint(
create_custom_forward(resnet),
hidden_states,
temb,
**ckpt_kwargs,
)
hidden_states, spatial_attn_inputs, spatial_attn_idx = attn(
hidden_states,
spatial_attn_inputs=spatial_attn_inputs,
spatial_attn_idx=spatial_attn_idx,
encoder_hidden_states=encoder_hidden_states,
cross_attention_kwargs=cross_attention_kwargs,
attention_mask=attention_mask,
encoder_attention_mask=encoder_attention_mask,
return_dict=False,
)
hidden_states = hidden_states[0]
else:
hidden_states = resnet(hidden_states, temb, scale=lora_scale)
hidden_states, spatial_attn_inputs, spatial_attn_idx = attn(
hidden_states,
spatial_attn_inputs=spatial_attn_inputs,
spatial_attn_idx=spatial_attn_idx,
encoder_hidden_states=encoder_hidden_states,
cross_attention_kwargs=cross_attention_kwargs,
attention_mask=attention_mask,
encoder_attention_mask=encoder_attention_mask,
return_dict=False,
)
hidden_states = hidden_states[0]
if self.upsamplers is not None:
for upsampler in self.upsamplers:
hidden_states = upsampler(hidden_states, upsample_size, scale=lora_scale)
return hidden_states, spatial_attn_inputs, spatial_attn_idx
class UpBlock2D(nn.Module):
def __init__(
self,
in_channels: int,
prev_output_channel: int,
out_channels: int,
temb_channels: int,
resolution_idx: Optional[int] = None,
dropout: float = 0.0,
num_layers: int = 1,
resnet_eps: float = 1e-6,
resnet_time_scale_shift: str = "default",
resnet_act_fn: str = "swish",
resnet_groups: int = 32,
resnet_pre_norm: bool = True,
output_scale_factor: float = 1.0,
add_upsample: bool = True,
):
super().__init__()
resnets = []
for i in range(num_layers):
res_skip_channels = in_channels if (i == num_layers - 1) else out_channels
resnet_in_channels = prev_output_channel if i == 0 else out_channels
resnets.append(
ResnetBlock2D(
in_channels=resnet_in_channels + res_skip_channels,
out_channels=out_channels,
temb_channels=temb_channels,
eps=resnet_eps,
groups=resnet_groups,
dropout=dropout,
time_embedding_norm=resnet_time_scale_shift,
non_linearity=resnet_act_fn,
output_scale_factor=output_scale_factor,
pre_norm=resnet_pre_norm,
)
)
self.resnets = nn.ModuleList(resnets)
if add_upsample:
self.upsamplers = nn.ModuleList([Upsample2D(out_channels, use_conv=True, out_channels=out_channels)])
else:
self.upsamplers = None
self.gradient_checkpointing = False
self.resolution_idx = resolution_idx
def forward(
self,
hidden_states: torch.FloatTensor,
res_hidden_states_tuple: Tuple[torch.FloatTensor, ...],
temb: Optional[torch.FloatTensor] = None,
upsample_size: Optional[int] = None,
scale: float = 1.0,
) -> torch.FloatTensor:
is_freeu_enabled = (
getattr(self, "s1", None)
and getattr(self, "s2", None)
and getattr(self, "b1", None)
and getattr(self, "b2", None)
)
for resnet in self.resnets:
# pop res hidden states
res_hidden_states = res_hidden_states_tuple[-1]
res_hidden_states_tuple = res_hidden_states_tuple[:-1]
# FreeU: Only operate on the first two stages
if is_freeu_enabled:
hidden_states, res_hidden_states = apply_freeu(
self.resolution_idx,
hidden_states,
res_hidden_states,
s1=self.s1,
s2=self.s2,
b1=self.b1,
b2=self.b2,
)
hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)
if self.training and self.gradient_checkpointing:
def create_custom_forward(module):
def custom_forward(*inputs):
return module(*inputs)
return custom_forward
if is_torch_version(">=", "1.11.0"):
hidden_states = torch.utils.checkpoint.checkpoint(
create_custom_forward(resnet), hidden_states, temb, use_reentrant=False
)
else:
hidden_states = torch.utils.checkpoint.checkpoint(
create_custom_forward(resnet), hidden_states, temb
)
else:
hidden_states = resnet(hidden_states, temb, scale=scale)
if self.upsamplers is not None:
for upsampler in self.upsamplers:
hidden_states = upsampler(hidden_states, upsample_size, scale=scale)
return hidden_states
class UpDecoderBlock2D(nn.Module):
def __init__(
self,
in_channels: int,
out_channels: int,
resolution_idx: Optional[int] = None,
dropout: float = 0.0,
num_layers: int = 1,
resnet_eps: float = 1e-6,
resnet_time_scale_shift: str = "default", # default, spatial
resnet_act_fn: str = "swish",
resnet_groups: int = 32,
resnet_pre_norm: bool = True,
output_scale_factor: float = 1.0,
add_upsample: bool = True,
temb_channels: Optional[int] = None,
):
super().__init__()
resnets = []
for i in range(num_layers):
input_channels = in_channels if i == 0 else out_channels
resnets.append(
ResnetBlock2D(
in_channels=input_channels,
out_channels=out_channels,
temb_channels=temb_channels,
eps=resnet_eps,
groups=resnet_groups,
dropout=dropout,
time_embedding_norm=resnet_time_scale_shift,
non_linearity=resnet_act_fn,
output_scale_factor=output_scale_factor,
pre_norm=resnet_pre_norm,
)
)
self.resnets = nn.ModuleList(resnets)
if add_upsample:
self.upsamplers = nn.ModuleList([Upsample2D(out_channels, use_conv=True, out_channels=out_channels)])
else:
self.upsamplers = None
self.resolution_idx = resolution_idx
def forward(
self, hidden_states: torch.FloatTensor, temb: Optional[torch.FloatTensor] = None, scale: float = 1.0
) -> torch.FloatTensor:
for resnet in self.resnets:
hidden_states = resnet(hidden_states, temb=temb, scale=scale)
if self.upsamplers is not None:
for upsampler in self.upsamplers:
hidden_states = upsampler(hidden_states)
return hidden_states
class AttnUpDecoderBlock2D(nn.Module):
def __init__(
self,
in_channels: int,
out_channels: int,
resolution_idx: Optional[int] = None,
dropout: float = 0.0,
num_layers: int = 1,
resnet_eps: float = 1e-6,
resnet_time_scale_shift: str = "default",
resnet_act_fn: str = "swish",
resnet_groups: int = 32,
resnet_pre_norm: bool = True,
attention_head_dim: int = 1,
output_scale_factor: float = 1.0,
add_upsample: bool = True,
temb_channels: Optional[int] = None,
):
super().__init__()
resnets = []
attentions = []
if attention_head_dim is None:
logger.warn(
f"It is not recommend to pass `attention_head_dim=None`. Defaulting `attention_head_dim` to `out_channels`: {out_channels}."
)
attention_head_dim = out_channels
for i in range(num_layers):
input_channels = in_channels if i == 0 else out_channels
resnets.append(
ResnetBlock2D(
in_channels=input_channels,
out_channels=out_channels,
temb_channels=temb_channels,
eps=resnet_eps,
groups=resnet_groups,
dropout=dropout,
time_embedding_norm=resnet_time_scale_shift,
non_linearity=resnet_act_fn,
output_scale_factor=output_scale_factor,
pre_norm=resnet_pre_norm,
)
)
attentions.append(
Attention(
out_channels,
heads=out_channels // attention_head_dim,
dim_head=attention_head_dim,
rescale_output_factor=output_scale_factor,
eps=resnet_eps,
norm_num_groups=resnet_groups if resnet_time_scale_shift != "spatial" else None,
spatial_norm_dim=temb_channels if resnet_time_scale_shift == "spatial" else None,
residual_connection=True,
bias=True,
upcast_softmax=True,
_from_deprecated_attn_block=True,
)
)
self.attentions = nn.ModuleList(attentions)
self.resnets = nn.ModuleList(resnets)
if add_upsample:
self.upsamplers = nn.ModuleList([Upsample2D(out_channels, use_conv=True, out_channels=out_channels)])
else:
self.upsamplers = None
self.resolution_idx = resolution_idx
def forward(
self, hidden_states: torch.FloatTensor, temb: Optional[torch.FloatTensor] = None, scale: float = 1.0
) -> torch.FloatTensor:
for resnet, attn in zip(self.resnets, self.attentions):
hidden_states = resnet(hidden_states, temb=temb, scale=scale)
cross_attention_kwargs = {"scale": scale}
hidden_states = attn(hidden_states, temb=temb, **cross_attention_kwargs)
if self.upsamplers is not None:
for upsampler in self.upsamplers:
hidden_states = upsampler(hidden_states, scale=scale)
return hidden_states
class AttnSkipUpBlock2D(nn.Module):
def __init__(
self,
in_channels: int,
prev_output_channel: int,
out_channels: int,
temb_channels: int,
resolution_idx: Optional[int] = None,
dropout: float = 0.0,
num_layers: int = 1,
resnet_eps: float = 1e-6,
resnet_time_scale_shift: str = "default",
resnet_act_fn: str = "swish",
resnet_pre_norm: bool = True,
attention_head_dim: int = 1,
output_scale_factor: float = np.sqrt(2.0),
add_upsample: bool = True,
):
super().__init__()
self.attentions = nn.ModuleList([])
self.resnets = nn.ModuleList([])
for i in range(num_layers):
res_skip_channels = in_channels if (i == num_layers - 1) else out_channels
resnet_in_channels = prev_output_channel if i == 0 else out_channels
self.resnets.append(
ResnetBlock2D(
in_channels=resnet_in_channels + res_skip_channels,
out_channels=out_channels,
temb_channels=temb_channels,
eps=resnet_eps,
groups=min(resnet_in_channels + res_skip_channels // 4, 32),
groups_out=min(out_channels // 4, 32),
dropout=dropout,
time_embedding_norm=resnet_time_scale_shift,
non_linearity=resnet_act_fn,
output_scale_factor=output_scale_factor,
pre_norm=resnet_pre_norm,
)
)
if attention_head_dim is None:
logger.warn(
f"It is not recommend to pass `attention_head_dim=None`. Defaulting `attention_head_dim` to `out_channels`: {out_channels}."
)
attention_head_dim = out_channels
self.attentions.append(
Attention(
out_channels,
heads=out_channels // attention_head_dim,
dim_head=attention_head_dim,
rescale_output_factor=output_scale_factor,
eps=resnet_eps,
norm_num_groups=32,
residual_connection=True,
bias=True,
upcast_softmax=True,
_from_deprecated_attn_block=True,
)
)
self.upsampler = FirUpsample2D(in_channels, out_channels=out_channels)
if add_upsample:
self.resnet_up = ResnetBlock2D(
in_channels=out_channels,
out_channels=out_channels,
temb_channels=temb_channels,
eps=resnet_eps,
groups=min(out_channels // 4, 32),
groups_out=min(out_channels // 4, 32),
dropout=dropout,
time_embedding_norm=resnet_time_scale_shift,
non_linearity=resnet_act_fn,
output_scale_factor=output_scale_factor,
pre_norm=resnet_pre_norm,
use_in_shortcut=True,
up=True,
kernel="fir",
)
self.skip_conv = nn.Conv2d(out_channels, 3, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
self.skip_norm = torch.nn.GroupNorm(
num_groups=min(out_channels // 4, 32), num_channels=out_channels, eps=resnet_eps, affine=True
)
self.act = nn.SiLU()
else:
self.resnet_up = None
self.skip_conv = None
self.skip_norm = None
self.act = None
self.resolution_idx = resolution_idx
def forward(
self,
hidden_states: torch.FloatTensor,
res_hidden_states_tuple: Tuple[torch.FloatTensor, ...],
temb: Optional[torch.FloatTensor] = None,
skip_sample=None,
scale: float = 1.0,
) -> Tuple[torch.FloatTensor, torch.FloatTensor]:
for resnet in self.resnets:
# pop res hidden states
res_hidden_states = res_hidden_states_tuple[-1]
res_hidden_states_tuple = res_hidden_states_tuple[:-1]
hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)
hidden_states = resnet(hidden_states, temb, scale=scale)
cross_attention_kwargs = {"scale": scale}
hidden_states = self.attentions[0](hidden_states, **cross_attention_kwargs)
if skip_sample is not None:
skip_sample = self.upsampler(skip_sample)
else:
skip_sample = 0
if self.resnet_up is not None:
skip_sample_states = self.skip_norm(hidden_states)
skip_sample_states = self.act(skip_sample_states)
skip_sample_states = self.skip_conv(skip_sample_states)
skip_sample = skip_sample + skip_sample_states
hidden_states = self.resnet_up(hidden_states, temb, scale=scale)
return hidden_states, skip_sample
class SkipUpBlock2D(nn.Module):
def __init__(
self,
in_channels: int,
prev_output_channel: int,
out_channels: int,
temb_channels: int,
resolution_idx: Optional[int] = None,
dropout: float = 0.0,
num_layers: int = 1,
resnet_eps: float = 1e-6,
resnet_time_scale_shift: str = "default",
resnet_act_fn: str = "swish",
resnet_pre_norm: bool = True,
output_scale_factor: float = np.sqrt(2.0),
add_upsample: bool = True,
upsample_padding: int = 1,
):
super().__init__()
self.resnets = nn.ModuleList([])
for i in range(num_layers):
res_skip_channels = in_channels if (i == num_layers - 1) else out_channels
resnet_in_channels = prev_output_channel if i == 0 else out_channels
self.resnets.append(
ResnetBlock2D(
in_channels=resnet_in_channels + res_skip_channels,
out_channels=out_channels,
temb_channels=temb_channels,
eps=resnet_eps,
groups=min((resnet_in_channels + res_skip_channels) // 4, 32),
groups_out=min(out_channels // 4, 32),
dropout=dropout,
time_embedding_norm=resnet_time_scale_shift,
non_linearity=resnet_act_fn,
output_scale_factor=output_scale_factor,
pre_norm=resnet_pre_norm,
)
)
self.upsampler = FirUpsample2D(in_channels, out_channels=out_channels)
if add_upsample:
self.resnet_up = ResnetBlock2D(
in_channels=out_channels,
out_channels=out_channels,
temb_channels=temb_channels,
eps=resnet_eps,
groups=min(out_channels // 4, 32),
groups_out=min(out_channels // 4, 32),
dropout=dropout,
time_embedding_norm=resnet_time_scale_shift,
non_linearity=resnet_act_fn,
output_scale_factor=output_scale_factor,
pre_norm=resnet_pre_norm,
use_in_shortcut=True,
up=True,
kernel="fir",
)
self.skip_conv = nn.Conv2d(out_channels, 3, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
self.skip_norm = torch.nn.GroupNorm(
num_groups=min(out_channels // 4, 32), num_channels=out_channels, eps=resnet_eps, affine=True
)
self.act = nn.SiLU()
else:
self.resnet_up = None
self.skip_conv = None
self.skip_norm = None
self.act = None
self.resolution_idx = resolution_idx
def forward(
self,
hidden_states: torch.FloatTensor,
res_hidden_states_tuple: Tuple[torch.FloatTensor, ...],
temb: Optional[torch.FloatTensor] = None,
skip_sample=None,
scale: float = 1.0,
) -> Tuple[torch.FloatTensor, torch.FloatTensor]:
for resnet in self.resnets:
# pop res hidden states
res_hidden_states = res_hidden_states_tuple[-1]
res_hidden_states_tuple = res_hidden_states_tuple[:-1]
hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)
hidden_states = resnet(hidden_states, temb, scale=scale)
if skip_sample is not None:
skip_sample = self.upsampler(skip_sample)
else:
skip_sample = 0
if self.resnet_up is not None:
skip_sample_states = self.skip_norm(hidden_states)
skip_sample_states = self.act(skip_sample_states)
skip_sample_states = self.skip_conv(skip_sample_states)
skip_sample = skip_sample + skip_sample_states
hidden_states = self.resnet_up(hidden_states, temb, scale=scale)
return hidden_states, skip_sample
class ResnetUpsampleBlock2D(nn.Module):
def __init__(
self,
in_channels: int,
prev_output_channel: int,
out_channels: int,
temb_channels: int,
resolution_idx: Optional[int] = None,
dropout: float = 0.0,
num_layers: int = 1,
resnet_eps: float = 1e-6,
resnet_time_scale_shift: str = "default",
resnet_act_fn: str = "swish",
resnet_groups: int = 32,
resnet_pre_norm: bool = True,
output_scale_factor: float = 1.0,
add_upsample: bool = True,
skip_time_act: bool = False,
):
super().__init__()
resnets = []
for i in range(num_layers):
res_skip_channels = in_channels if (i == num_layers - 1) else out_channels
resnet_in_channels = prev_output_channel if i == 0 else out_channels
resnets.append(
ResnetBlock2D(
in_channels=resnet_in_channels + res_skip_channels,
out_channels=out_channels,
temb_channels=temb_channels,
eps=resnet_eps,
groups=resnet_groups,
dropout=dropout,
time_embedding_norm=resnet_time_scale_shift,
non_linearity=resnet_act_fn,
output_scale_factor=output_scale_factor,
pre_norm=resnet_pre_norm,
skip_time_act=skip_time_act,
)
)
self.resnets = nn.ModuleList(resnets)
if add_upsample:
self.upsamplers = nn.ModuleList(
[
ResnetBlock2D(
in_channels=out_channels,
out_channels=out_channels,
temb_channels=temb_channels,
eps=resnet_eps,
groups=resnet_groups,
dropout=dropout,
time_embedding_norm=resnet_time_scale_shift,
non_linearity=resnet_act_fn,
output_scale_factor=output_scale_factor,
pre_norm=resnet_pre_norm,
skip_time_act=skip_time_act,
up=True,
)
]
)
else:
self.upsamplers = None
self.gradient_checkpointing = False
self.resolution_idx = resolution_idx
def forward(
self,
hidden_states: torch.FloatTensor,
res_hidden_states_tuple: Tuple[torch.FloatTensor, ...],
temb: Optional[torch.FloatTensor] = None,
upsample_size: Optional[int] = None,
scale: float = 1.0,
) -> torch.FloatTensor:
for resnet in self.resnets:
# pop res hidden states
res_hidden_states = res_hidden_states_tuple[-1]
res_hidden_states_tuple = res_hidden_states_tuple[:-1]
hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)
if self.training and self.gradient_checkpointing:
def create_custom_forward(module):
def custom_forward(*inputs):
return module(*inputs)
return custom_forward
if is_torch_version(">=", "1.11.0"):
hidden_states = torch.utils.checkpoint.checkpoint(
create_custom_forward(resnet), hidden_states, temb, use_reentrant=False
)
else:
hidden_states = torch.utils.checkpoint.checkpoint(
create_custom_forward(resnet), hidden_states, temb
)
else:
hidden_states = resnet(hidden_states, temb, scale=scale)
if self.upsamplers is not None:
for upsampler in self.upsamplers:
hidden_states = upsampler(hidden_states, temb, scale=scale)
return hidden_states
class SimpleCrossAttnUpBlock2D(nn.Module):
def __init__(
self,
in_channels: int,
out_channels: int,
prev_output_channel: int,
temb_channels: int,
resolution_idx: Optional[int] = None,
dropout: float = 0.0,
num_layers: int = 1,
resnet_eps: float = 1e-6,
resnet_time_scale_shift: str = "default",
resnet_act_fn: str = "swish",
resnet_groups: int = 32,
resnet_pre_norm: bool = True,
attention_head_dim: int = 1,
cross_attention_dim: int = 1280,
output_scale_factor: float = 1.0,
add_upsample: bool = True,
skip_time_act: bool = False,
only_cross_attention: bool = False,
cross_attention_norm: Optional[str] = None,
):
super().__init__()
resnets = []
attentions = []
self.has_cross_attention = True
self.attention_head_dim = attention_head_dim
self.num_heads = out_channels // self.attention_head_dim
for i in range(num_layers):
res_skip_channels = in_channels if (i == num_layers - 1) else out_channels
resnet_in_channels = prev_output_channel if i == 0 else out_channels
resnets.append(
ResnetBlock2D(
in_channels=resnet_in_channels + res_skip_channels,
out_channels=out_channels,
temb_channels=temb_channels,
eps=resnet_eps,
groups=resnet_groups,
dropout=dropout,
time_embedding_norm=resnet_time_scale_shift,
non_linearity=resnet_act_fn,
output_scale_factor=output_scale_factor,
pre_norm=resnet_pre_norm,
skip_time_act=skip_time_act,
)
)
processor = (
AttnAddedKVProcessor2_0() if hasattr(F, "scaled_dot_product_attention") else AttnAddedKVProcessor()
)
attentions.append(
Attention(
query_dim=out_channels,
cross_attention_dim=out_channels,
heads=self.num_heads,
dim_head=self.attention_head_dim,
added_kv_proj_dim=cross_attention_dim,
norm_num_groups=resnet_groups,
bias=True,
upcast_softmax=True,
only_cross_attention=only_cross_attention,
cross_attention_norm=cross_attention_norm,
processor=processor,
)
)
self.attentions = nn.ModuleList(attentions)
self.resnets = nn.ModuleList(resnets)
if add_upsample:
self.upsamplers = nn.ModuleList(
[
ResnetBlock2D(
in_channels=out_channels,
out_channels=out_channels,
temb_channels=temb_channels,
eps=resnet_eps,
groups=resnet_groups,
dropout=dropout,
time_embedding_norm=resnet_time_scale_shift,
non_linearity=resnet_act_fn,
output_scale_factor=output_scale_factor,
pre_norm=resnet_pre_norm,
skip_time_act=skip_time_act,
up=True,
)
]
)
else:
self.upsamplers = None
self.gradient_checkpointing = False
self.resolution_idx = resolution_idx
def forward(
self,
hidden_states: torch.FloatTensor,
res_hidden_states_tuple: Tuple[torch.FloatTensor, ...],
temb: Optional[torch.FloatTensor] = None,
encoder_hidden_states: Optional[torch.FloatTensor] = None,
upsample_size: Optional[int] = None,
attention_mask: Optional[torch.FloatTensor] = None,
cross_attention_kwargs: Optional[Dict[str, Any]] = None,
encoder_attention_mask: Optional[torch.FloatTensor] = None,
) -> torch.FloatTensor:
cross_attention_kwargs = cross_attention_kwargs if cross_attention_kwargs is not None else {}
lora_scale = cross_attention_kwargs.get("scale", 1.0)
if attention_mask is None:
# if encoder_hidden_states is defined: we are doing cross-attn, so we should use cross-attn mask.
mask = None if encoder_hidden_states is None else encoder_attention_mask
else:
# when attention_mask is defined: we don't even check for encoder_attention_mask.
# this is to maintain compatibility with UnCLIP, which uses 'attention_mask' param for cross-attn masks.
# TODO: UnCLIP should express cross-attn mask via encoder_attention_mask param instead of via attention_mask.
# then we can simplify this whole if/else block to:
# mask = attention_mask if encoder_hidden_states is None else encoder_attention_mask
mask = attention_mask
for resnet, attn in zip(self.resnets, self.attentions):
# resnet
# pop res hidden states
res_hidden_states = res_hidden_states_tuple[-1]
res_hidden_states_tuple = res_hidden_states_tuple[:-1]
hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)
if self.training and self.gradient_checkpointing:
def create_custom_forward(module, return_dict=None):
def custom_forward(*inputs):
if return_dict is not None:
return module(*inputs, return_dict=return_dict)
else:
return module(*inputs)
return custom_forward
hidden_states = torch.utils.checkpoint.checkpoint(create_custom_forward(resnet), hidden_states, temb)
hidden_states = attn(
hidden_states,
encoder_hidden_states=encoder_hidden_states,
attention_mask=mask,
**cross_attention_kwargs,
)
else:
hidden_states = resnet(hidden_states, temb, scale=lora_scale)
hidden_states = attn(
hidden_states,
encoder_hidden_states=encoder_hidden_states,
attention_mask=mask,
**cross_attention_kwargs,
)
if self.upsamplers is not None:
for upsampler in self.upsamplers:
hidden_states = upsampler(hidden_states, temb, scale=lora_scale)
return hidden_states
class KUpBlock2D(nn.Module):
def __init__(
self,
in_channels: int,
out_channels: int,
temb_channels: int,
resolution_idx: int,
dropout: float = 0.0,
num_layers: int = 5,
resnet_eps: float = 1e-5,
resnet_act_fn: str = "gelu",
resnet_group_size: Optional[int] = 32,
add_upsample: bool = True,
):
super().__init__()
resnets = []
k_in_channels = 2 * out_channels
k_out_channels = in_channels
num_layers = num_layers - 1
for i in range(num_layers):
in_channels = k_in_channels if i == 0 else out_channels
groups = in_channels // resnet_group_size
groups_out = out_channels // resnet_group_size
resnets.append(
ResnetBlock2D(
in_channels=in_channels,
out_channels=k_out_channels if (i == num_layers - 1) else out_channels,
temb_channels=temb_channels,
eps=resnet_eps,
groups=groups,
groups_out=groups_out,
dropout=dropout,
non_linearity=resnet_act_fn,
time_embedding_norm="ada_group",
conv_shortcut_bias=False,
)
)
self.resnets = nn.ModuleList(resnets)
if add_upsample:
self.upsamplers = nn.ModuleList([KUpsample2D()])
else:
self.upsamplers = None
self.gradient_checkpointing = False
self.resolution_idx = resolution_idx
def forward(
self,
hidden_states: torch.FloatTensor,
res_hidden_states_tuple: Tuple[torch.FloatTensor, ...],
temb: Optional[torch.FloatTensor] = None,
upsample_size: Optional[int] = None,
scale: float = 1.0,
) -> torch.FloatTensor:
res_hidden_states_tuple = res_hidden_states_tuple[-1]
if res_hidden_states_tuple is not None:
hidden_states = torch.cat([hidden_states, res_hidden_states_tuple], dim=1)
for resnet in self.resnets:
if self.training and self.gradient_checkpointing:
def create_custom_forward(module):
def custom_forward(*inputs):
return module(*inputs)
return custom_forward
if is_torch_version(">=", "1.11.0"):
hidden_states = torch.utils.checkpoint.checkpoint(
create_custom_forward(resnet), hidden_states, temb, use_reentrant=False
)
else:
hidden_states = torch.utils.checkpoint.checkpoint(
create_custom_forward(resnet), hidden_states, temb
)
else:
hidden_states = resnet(hidden_states, temb, scale=scale)
if self.upsamplers is not None:
for upsampler in self.upsamplers:
hidden_states = upsampler(hidden_states)
return hidden_states
class KCrossAttnUpBlock2D(nn.Module):
def __init__(
self,
in_channels: int,
out_channels: int,
temb_channels: int,
resolution_idx: int,
dropout: float = 0.0,
num_layers: int = 4,
resnet_eps: float = 1e-5,
resnet_act_fn: str = "gelu",
resnet_group_size: int = 32,
attention_head_dim: int = 1, # attention dim_head
cross_attention_dim: int = 768,
add_upsample: bool = True,
upcast_attention: bool = False,
):
super().__init__()
resnets = []
attentions = []
is_first_block = in_channels == out_channels == temb_channels
is_middle_block = in_channels != out_channels
add_self_attention = True if is_first_block else False
self.has_cross_attention = True
self.attention_head_dim = attention_head_dim
# in_channels, and out_channels for the block (k-unet)
k_in_channels = out_channels if is_first_block else 2 * out_channels
k_out_channels = in_channels
num_layers = num_layers - 1
for i in range(num_layers):
in_channels = k_in_channels if i == 0 else out_channels
groups = in_channels // resnet_group_size
groups_out = out_channels // resnet_group_size
if is_middle_block and (i == num_layers - 1):
conv_2d_out_channels = k_out_channels
else:
conv_2d_out_channels = None
resnets.append(
ResnetBlock2D(
in_channels=in_channels,
out_channels=out_channels,
conv_2d_out_channels=conv_2d_out_channels,
temb_channels=temb_channels,
eps=resnet_eps,
groups=groups,
groups_out=groups_out,
dropout=dropout,
non_linearity=resnet_act_fn,
time_embedding_norm="ada_group",
conv_shortcut_bias=False,
)
)
attentions.append(
KAttentionBlock(
k_out_channels if (i == num_layers - 1) else out_channels,
k_out_channels // attention_head_dim
if (i == num_layers - 1)
else out_channels // attention_head_dim,
attention_head_dim,
cross_attention_dim=cross_attention_dim,
temb_channels=temb_channels,
attention_bias=True,
add_self_attention=add_self_attention,
cross_attention_norm="layer_norm",
upcast_attention=upcast_attention,
)
)
self.resnets = nn.ModuleList(resnets)
self.attentions = nn.ModuleList(attentions)
if add_upsample:
self.upsamplers = nn.ModuleList([KUpsample2D()])
else:
self.upsamplers = None
self.gradient_checkpointing = False
self.resolution_idx = resolution_idx
def forward(
self,
hidden_states: torch.FloatTensor,
res_hidden_states_tuple: Tuple[torch.FloatTensor, ...],
temb: Optional[torch.FloatTensor] = None,
encoder_hidden_states: Optional[torch.FloatTensor] = None,
cross_attention_kwargs: Optional[Dict[str, Any]] = None,
upsample_size: Optional[int] = None,
attention_mask: Optional[torch.FloatTensor] = None,
encoder_attention_mask: Optional[torch.FloatTensor] = None,
) -> torch.FloatTensor:
res_hidden_states_tuple = res_hidden_states_tuple[-1]
if res_hidden_states_tuple is not None:
hidden_states = torch.cat([hidden_states, res_hidden_states_tuple], dim=1)
lora_scale = cross_attention_kwargs.get("scale", 1.0) if cross_attention_kwargs is not None else 1.0
for resnet, attn in zip(self.resnets, self.attentions):
if self.training and self.gradient_checkpointing:
def create_custom_forward(module, return_dict=None):
def custom_forward(*inputs):
if return_dict is not None:
return module(*inputs, return_dict=return_dict)
else:
return module(*inputs)
return custom_forward
ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
hidden_states = torch.utils.checkpoint.checkpoint(
create_custom_forward(resnet),
hidden_states,
temb,
**ckpt_kwargs,
)
hidden_states = attn(
hidden_states,
encoder_hidden_states=encoder_hidden_states,
emb=temb,
attention_mask=attention_mask,
cross_attention_kwargs=cross_attention_kwargs,
encoder_attention_mask=encoder_attention_mask,
)
else:
hidden_states = resnet(hidden_states, temb, scale=lora_scale)
hidden_states = attn(
hidden_states,
encoder_hidden_states=encoder_hidden_states,
emb=temb,
attention_mask=attention_mask,
cross_attention_kwargs=cross_attention_kwargs,
encoder_attention_mask=encoder_attention_mask,
)
if self.upsamplers is not None:
for upsampler in self.upsamplers:
hidden_states = upsampler(hidden_states)
return hidden_states
# can potentially later be renamed to `No-feed-forward` attention
class KAttentionBlock(nn.Module):
r"""
A basic Transformer block.
Parameters:
dim (`int`): The number of channels in the input and output.
num_attention_heads (`int`): The number of heads to use for multi-head attention.
attention_head_dim (`int`): The number of channels in each head.
dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
cross_attention_dim (`int`, *optional*): The size of the encoder_hidden_states vector for cross attention.
attention_bias (`bool`, *optional*, defaults to `False`):
Configure if the attention layers should contain a bias parameter.
upcast_attention (`bool`, *optional*, defaults to `False`):
Set to `True` to upcast the attention computation to `float32`.
temb_channels (`int`, *optional*, defaults to 768):
The number of channels in the token embedding.
add_self_attention (`bool`, *optional*, defaults to `False`):
Set to `True` to add self-attention to the block.
cross_attention_norm (`str`, *optional*, defaults to `None`):
The type of normalization to use for the cross attention. Can be `None`, `layer_norm`, or `group_norm`.
group_size (`int`, *optional*, defaults to 32):
The number of groups to separate the channels into for group normalization.
"""
def __init__(
self,
dim: int,
num_attention_heads: int,
attention_head_dim: int,
dropout: float = 0.0,
cross_attention_dim: Optional[int] = None,
attention_bias: bool = False,
upcast_attention: bool = False,
temb_channels: int = 768, # for ada_group_norm
add_self_attention: bool = False,
cross_attention_norm: Optional[str] = None,
group_size: int = 32,
):
super().__init__()
self.add_self_attention = add_self_attention
# 1. Self-Attn
if add_self_attention:
self.norm1 = AdaGroupNorm(temb_channels, dim, max(1, dim // group_size))
self.attn1 = Attention(
query_dim=dim,
heads=num_attention_heads,
dim_head=attention_head_dim,
dropout=dropout,
bias=attention_bias,
cross_attention_dim=None,
cross_attention_norm=None,
)
# 2. Cross-Attn
self.norm2 = AdaGroupNorm(temb_channels, dim, max(1, dim // group_size))
self.attn2 = Attention(
query_dim=dim,
cross_attention_dim=cross_attention_dim,
heads=num_attention_heads,
dim_head=attention_head_dim,
dropout=dropout,
bias=attention_bias,
upcast_attention=upcast_attention,
cross_attention_norm=cross_attention_norm,
)
def _to_3d(self, hidden_states: torch.FloatTensor, height: int, weight: int) -> torch.FloatTensor:
return hidden_states.permute(0, 2, 3, 1).reshape(hidden_states.shape[0], height * weight, -1)
def _to_4d(self, hidden_states: torch.FloatTensor, height: int, weight: int) -> torch.FloatTensor:
return hidden_states.permute(0, 2, 1).reshape(hidden_states.shape[0], -1, height, weight)
def forward(
self,
hidden_states: torch.FloatTensor,
encoder_hidden_states: Optional[torch.FloatTensor] = None,
# TODO: mark emb as non-optional (self.norm2 requires it).
# requires assessing impact of change to positional param interface.
emb: Optional[torch.FloatTensor] = None,
attention_mask: Optional[torch.FloatTensor] = None,
cross_attention_kwargs: Optional[Dict[str, Any]] = None,
encoder_attention_mask: Optional[torch.FloatTensor] = None,
) -> torch.FloatTensor:
cross_attention_kwargs = cross_attention_kwargs if cross_attention_kwargs is not None else {}
# 1. Self-Attention
if self.add_self_attention:
norm_hidden_states = self.norm1(hidden_states, emb)
height, weight = norm_hidden_states.shape[2:]
norm_hidden_states = self._to_3d(norm_hidden_states, height, weight)
attn_output = self.attn1(
norm_hidden_states,
encoder_hidden_states=None,
attention_mask=attention_mask,
**cross_attention_kwargs,
)
attn_output = self._to_4d(attn_output, height, weight)
hidden_states = attn_output + hidden_states
# 2. Cross-Attention/None
norm_hidden_states = self.norm2(hidden_states, emb)
height, weight = norm_hidden_states.shape[2:]
norm_hidden_states = self._to_3d(norm_hidden_states, height, weight)
attn_output = self.attn2(
norm_hidden_states,
encoder_hidden_states=encoder_hidden_states,
attention_mask=attention_mask if encoder_hidden_states is None else encoder_attention_mask,
**cross_attention_kwargs,
)
attn_output = self._to_4d(attn_output, height, weight)
hidden_states = attn_output + hidden_states
return hidden_states
ootd/pipelines_ootd/unet_vton_2d_condition.py 0 → 100644
View file @ 54a066bf
# Copyright 2023 The HuggingFace Team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# Modified by Yuhao Xu for OOTDiffusion (https://github.com/levihsu/OOTDiffusion)
from dataclasses import dataclass
from typing import Any, Dict, List, Optional, Tuple, Union
import torch
import torch.nn as nn
import torch.utils.checkpoint
from .unet_vton_2d_blocks import (
UNetMidBlock2D,
UNetMidBlock2DCrossAttn,
UNetMidBlock2DSimpleCrossAttn,
get_down_block,
get_up_block,
)
from diffusers.configuration_utils import ConfigMixin, register_to_config
from diffusers.loaders import UNet2DConditionLoadersMixin
from diffusers.utils import USE_PEFT_BACKEND, BaseOutput, deprecate, logging, scale_lora_layers, unscale_lora_layers
from diffusers.models.activations import get_activation
from diffusers.models.attention_processor import (
ADDED_KV_ATTENTION_PROCESSORS,
CROSS_ATTENTION_PROCESSORS,
AttentionProcessor,
AttnAddedKVProcessor,
AttnProcessor,
)
from diffusers.models.embeddings import (
GaussianFourierProjection,
ImageHintTimeEmbedding,
ImageProjection,
ImageTimeEmbedding,
PositionNet,
TextImageProjection,
TextImageTimeEmbedding,
TextTimeEmbedding,
TimestepEmbedding,
Timesteps,
)
from diffusers.models.modeling_utils import ModelMixin
# from ..diffusers.src.diffusers.models.unet_2d_blocks import (
# UNetMidBlock2D,
# UNetMidBlock2DCrossAttn,
# UNetMidBlock2DSimpleCrossAttn,
# get_down_block,
# get_up_block,
# )
logger = logging.get_logger(__name__) # pylint: disable=invalid-name
@dataclass
class UNet2DConditionOutput(BaseOutput):
"""
The output of [`UNet2DConditionModel`].
Args:
sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`):
The hidden states output conditioned on `encoder_hidden_states` input. Output of last layer of model.
"""
sample: torch.FloatTensor = None
class UNetVton2DConditionModel(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin):
r"""
A conditional 2D UNet model that takes a noisy sample, conditional state, and a timestep and returns a sample
shaped output.
This model inherits from [`ModelMixin`]. Check the superclass documentation for it's generic methods implemented
for all models (such as downloading or saving).
Parameters:
sample_size (`int` or `Tuple[int, int]`, *optional*, defaults to `None`):
Height and width of input/output sample.
in_channels (`int`, *optional*, defaults to 4): Number of channels in the input sample.
out_channels (`int`, *optional*, defaults to 4): Number of channels in the output.
center_input_sample (`bool`, *optional*, defaults to `False`): Whether to center the input sample.
flip_sin_to_cos (`bool`, *optional*, defaults to `False`):
Whether to flip the sin to cos in the time embedding.
freq_shift (`int`, *optional*, defaults to 0): The frequency shift to apply to the time embedding.
down_block_types (`Tuple[str]`, *optional*, defaults to `("CrossAttnDownBlock2D", "CrossAttnDownBlock2D", "CrossAttnDownBlock2D", "DownBlock2D")`):
The tuple of downsample blocks to use.
mid_block_type (`str`, *optional*, defaults to `"UNetMidBlock2DCrossAttn"`):
Block type for middle of UNet, it can be one of `UNetMidBlock2DCrossAttn`, `UNetMidBlock2D`, or
`UNetMidBlock2DSimpleCrossAttn`. If `None`, the mid block layer is skipped.
up_block_types (`Tuple[str]`, *optional*, defaults to `("UpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D")`):
The tuple of upsample blocks to use.
only_cross_attention(`bool` or `Tuple[bool]`, *optional*, default to `False`):
Whether to include self-attention in the basic transformer blocks, see
[`~models.attention.BasicTransformerBlock`].
block_out_channels (`Tuple[int]`, *optional*, defaults to `(320, 640, 1280, 1280)`):
The tuple of output channels for each block.
layers_per_block (`int`, *optional*, defaults to 2): The number of layers per block.
downsample_padding (`int`, *optional*, defaults to 1): The padding to use for the downsampling convolution.
mid_block_scale_factor (`float`, *optional*, defaults to 1.0): The scale factor to use for the mid block.
dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
act_fn (`str`, *optional*, defaults to `"silu"`): The activation function to use.
norm_num_groups (`int`, *optional*, defaults to 32): The number of groups to use for the normalization.
If `None`, normalization and activation layers is skipped in post-processing.
norm_eps (`float`, *optional*, defaults to 1e-5): The epsilon to use for the normalization.
cross_attention_dim (`int` or `Tuple[int]`, *optional*, defaults to 1280):
The dimension of the cross attention features.
transformer_layers_per_block (`int`, `Tuple[int]`, or `Tuple[Tuple]` , *optional*, defaults to 1):
The number of transformer blocks of type [`~models.attention.BasicTransformerBlock`]. Only relevant for
[`~models.unet_2d_blocks.CrossAttnDownBlock2D`], [`~models.unet_2d_blocks.CrossAttnUpBlock2D`],
[`~models.unet_2d_blocks.UNetMidBlock2DCrossAttn`].
reverse_transformer_layers_per_block : (`Tuple[Tuple]`, *optional*, defaults to None):
The number of transformer blocks of type [`~models.attention.BasicTransformerBlock`], in the upsampling
blocks of the U-Net. Only relevant if `transformer_layers_per_block` is of type `Tuple[Tuple]` and for
[`~models.unet_2d_blocks.CrossAttnDownBlock2D`], [`~models.unet_2d_blocks.CrossAttnUpBlock2D`],
[`~models.unet_2d_blocks.UNetMidBlock2DCrossAttn`].
encoder_hid_dim (`int`, *optional*, defaults to None):
If `encoder_hid_dim_type` is defined, `encoder_hidden_states` will be projected from `encoder_hid_dim`
dimension to `cross_attention_dim`.
encoder_hid_dim_type (`str`, *optional*, defaults to `None`):
If given, the `encoder_hidden_states` and potentially other embeddings are down-projected to text
embeddings of dimension `cross_attention` according to `encoder_hid_dim_type`.
attention_head_dim (`int`, *optional*, defaults to 8): The dimension of the attention heads.
num_attention_heads (`int`, *optional*):
The number of attention heads. If not defined, defaults to `attention_head_dim`
resnet_time_scale_shift (`str`, *optional*, defaults to `"default"`): Time scale shift config
for ResNet blocks (see [`~models.resnet.ResnetBlock2D`]). Choose from `default` or `scale_shift`.
class_embed_type (`str`, *optional*, defaults to `None`):
The type of class embedding to use which is ultimately summed with the time embeddings. Choose from `None`,
`"timestep"`, `"identity"`, `"projection"`, or `"simple_projection"`.
addition_embed_type (`str`, *optional*, defaults to `None`):
Configures an optional embedding which will be summed with the time embeddings. Choose from `None` or
"text". "text" will use the `TextTimeEmbedding` layer.
addition_time_embed_dim: (`int`, *optional*, defaults to `None`):
Dimension for the timestep embeddings.
num_class_embeds (`int`, *optional*, defaults to `None`):
Input dimension of the learnable embedding matrix to be projected to `time_embed_dim`, when performing
class conditioning with `class_embed_type` equal to `None`.
time_embedding_type (`str`, *optional*, defaults to `positional`):
The type of position embedding to use for timesteps. Choose from `positional` or `fourier`.
time_embedding_dim (`int`, *optional*, defaults to `None`):
An optional override for the dimension of the projected time embedding.
time_embedding_act_fn (`str`, *optional*, defaults to `None`):
Optional activation function to use only once on the time embeddings before they are passed to the rest of
the UNet. Choose from `silu`, `mish`, `gelu`, and `swish`.
timestep_post_act (`str`, *optional*, defaults to `None`):
The second activation function to use in timestep embedding. Choose from `silu`, `mish` and `gelu`.
time_cond_proj_dim (`int`, *optional*, defaults to `None`):
The dimension of `cond_proj` layer in the timestep embedding.
conv_in_kernel (`int`, *optional*, default to `3`): The kernel size of `conv_in` layer. conv_out_kernel (`int`,
*optional*, default to `3`): The kernel size of `conv_out` layer. projection_class_embeddings_input_dim (`int`,
*optional*): The dimension of the `class_labels` input when
`class_embed_type="projection"`. Required when `class_embed_type="projection"`.
class_embeddings_concat (`bool`, *optional*, defaults to `False`): Whether to concatenate the time
embeddings with the class embeddings.
mid_block_only_cross_attention (`bool`, *optional*, defaults to `None`):
Whether to use cross attention with the mid block when using the `UNetMidBlock2DSimpleCrossAttn`. If
`only_cross_attention` is given as a single boolean and `mid_block_only_cross_attention` is `None`, the
`only_cross_attention` value is used as the value for `mid_block_only_cross_attention`. Default to `False`
otherwise.
"""
_supports_gradient_checkpointing = True
@register_to_config
def __init__(
self,
sample_size: Optional[int] = None,
in_channels: int = 4,
out_channels: int = 4,
center_input_sample: bool = False,
flip_sin_to_cos: bool = True,
freq_shift: int = 0,
down_block_types: Tuple[str] = (
"CrossAttnDownBlock2D",
"CrossAttnDownBlock2D",
"CrossAttnDownBlock2D",
"DownBlock2D",
),
mid_block_type: Optional[str] = "UNetMidBlock2DCrossAttn",
up_block_types: Tuple[str] = ("UpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D"),
only_cross_attention: Union[bool, Tuple[bool]] = False,
block_out_channels: Tuple[int] = (320, 640, 1280, 1280),
layers_per_block: Union[int, Tuple[int]] = 2,
downsample_padding: int = 1,
mid_block_scale_factor: float = 1,
dropout: float = 0.0,
act_fn: str = "silu",
norm_num_groups: Optional[int] = 32,
norm_eps: float = 1e-5,
cross_attention_dim: Union[int, Tuple[int]] = 1280,
transformer_layers_per_block: Union[int, Tuple[int], Tuple[Tuple]] = 1,
reverse_transformer_layers_per_block: Optional[Tuple[Tuple[int]]] = None,
encoder_hid_dim: Optional[int] = None,
encoder_hid_dim_type: Optional[str] = None,
attention_head_dim: Union[int, Tuple[int]] = 8,
num_attention_heads: Optional[Union[int, Tuple[int]]] = None,
dual_cross_attention: bool = False,
use_linear_projection: bool = False,
class_embed_type: Optional[str] = None,
addition_embed_type: Optional[str] = None,
addition_time_embed_dim: Optional[int] = None,
num_class_embeds: Optional[int] = None,
upcast_attention: bool = False,
resnet_time_scale_shift: str = "default",
resnet_skip_time_act: bool = False,
resnet_out_scale_factor: int = 1.0,
time_embedding_type: str = "positional",
time_embedding_dim: Optional[int] = None,
time_embedding_act_fn: Optional[str] = None,
timestep_post_act: Optional[str] = None,
time_cond_proj_dim: Optional[int] = None,
conv_in_kernel: int = 3,
conv_out_kernel: int = 3,
projection_class_embeddings_input_dim: Optional[int] = None,
attention_type: str = "default",
class_embeddings_concat: bool = False,
mid_block_only_cross_attention: Optional[bool] = None,
cross_attention_norm: Optional[str] = None,
addition_embed_type_num_heads=64,
):
super().__init__()
self.sample_size = sample_size
if num_attention_heads is not None:
raise ValueError(
"At the moment it is not possible to define the number of attention heads via `num_attention_heads` because of a naming issue as described in https://github.com/huggingface/diffusers/issues/2011#issuecomment-1547958131. Passing `num_attention_heads` will only be supported in diffusers v0.19."
)
# If `num_attention_heads` is not defined (which is the case for most models)
# it will default to `attention_head_dim`. This looks weird upon first reading it and it is.
# The reason for this behavior is to correct for incorrectly named variables that were introduced
# when this library was created. The incorrect naming was only discovered much later in https://github.com/huggingface/diffusers/issues/2011#issuecomment-1547958131
# Changing `attention_head_dim` to `num_attention_heads` for 40,000+ configurations is too backwards breaking
# which is why we correct for the naming here.
num_attention_heads = num_attention_heads or attention_head_dim
# Check inputs
if len(down_block_types) != len(up_block_types):
raise ValueError(
f"Must provide the same number of `down_block_types` as `up_block_types`. `down_block_types`: {down_block_types}. `up_block_types`: {up_block_types}."
)
if len(block_out_channels) != len(down_block_types):
raise ValueError(
f"Must provide the same number of `block_out_channels` as `down_block_types`. `block_out_channels`: {block_out_channels}. `down_block_types`: {down_block_types}."
)
if not isinstance(only_cross_attention, bool) and len(only_cross_attention) != len(down_block_types):
raise ValueError(
f"Must provide the same number of `only_cross_attention` as `down_block_types`. `only_cross_attention`: {only_cross_attention}. `down_block_types`: {down_block_types}."
)
if not isinstance(num_attention_heads, int) and len(num_attention_heads) != len(down_block_types):
raise ValueError(
f"Must provide the same number of `num_attention_heads` as `down_block_types`. `num_attention_heads`: {num_attention_heads}. `down_block_types`: {down_block_types}."
)
if not isinstance(attention_head_dim, int) and len(attention_head_dim) != len(down_block_types):
raise ValueError(
f"Must provide the same number of `attention_head_dim` as `down_block_types`. `attention_head_dim`: {attention_head_dim}. `down_block_types`: {down_block_types}."
)
if isinstance(cross_attention_dim, list) and len(cross_attention_dim) != len(down_block_types):
raise ValueError(
f"Must provide the same number of `cross_attention_dim` as `down_block_types`. `cross_attention_dim`: {cross_attention_dim}. `down_block_types`: {down_block_types}."
)
if not isinstance(layers_per_block, int) and len(layers_per_block) != len(down_block_types):
raise ValueError(
f"Must provide the same number of `layers_per_block` as `down_block_types`. `layers_per_block`: {layers_per_block}. `down_block_types`: {down_block_types}."
)
if isinstance(transformer_layers_per_block, list) and reverse_transformer_layers_per_block is None:
for layer_number_per_block in transformer_layers_per_block:
if isinstance(layer_number_per_block, list):
raise ValueError("Must provide 'reverse_transformer_layers_per_block` if using asymmetrical UNet.")
# input
conv_in_padding = (conv_in_kernel - 1) // 2
self.conv_in = nn.Conv2d(
in_channels, block_out_channels[0], kernel_size=conv_in_kernel, padding=conv_in_padding
)
# time
if time_embedding_type == "fourier":
time_embed_dim = time_embedding_dim or block_out_channels[0] * 2
if time_embed_dim % 2 != 0:
raise ValueError(f"`time_embed_dim` should be divisible by 2, but is {time_embed_dim}.")
self.time_proj = GaussianFourierProjection(
time_embed_dim // 2, set_W_to_weight=False, log=False, flip_sin_to_cos=flip_sin_to_cos
)
timestep_input_dim = time_embed_dim
elif time_embedding_type == "positional":
time_embed_dim = time_embedding_dim or block_out_channels[0] * 4
self.time_proj = Timesteps(block_out_channels[0], flip_sin_to_cos, freq_shift)
timestep_input_dim = block_out_channels[0]
else:
raise ValueError(
f"{time_embedding_type} does not exist. Please make sure to use one of `fourier` or `positional`."
)
self.time_embedding = TimestepEmbedding(
timestep_input_dim,
time_embed_dim,
act_fn=act_fn,
post_act_fn=timestep_post_act,
cond_proj_dim=time_cond_proj_dim,
)
if encoder_hid_dim_type is None and encoder_hid_dim is not None:
encoder_hid_dim_type = "text_proj"
self.register_to_config(encoder_hid_dim_type=encoder_hid_dim_type)
logger.info("encoder_hid_dim_type defaults to 'text_proj' as `encoder_hid_dim` is defined.")
if encoder_hid_dim is None and encoder_hid_dim_type is not None:
raise ValueError(
f"`encoder_hid_dim` has to be defined when `encoder_hid_dim_type` is set to {encoder_hid_dim_type}."
)
if encoder_hid_dim_type == "text_proj":
self.encoder_hid_proj = nn.Linear(encoder_hid_dim, cross_attention_dim)
elif encoder_hid_dim_type == "text_image_proj":
# image_embed_dim DOESN'T have to be `cross_attention_dim`. To not clutter the __init__ too much
# they are set to `cross_attention_dim` here as this is exactly the required dimension for the currently only use
# case when `addition_embed_type == "text_image_proj"` (Kadinsky 2.1)`
self.encoder_hid_proj = TextImageProjection(
text_embed_dim=encoder_hid_dim,
image_embed_dim=cross_attention_dim,
cross_attention_dim=cross_attention_dim,
)
elif encoder_hid_dim_type == "image_proj":
# Kandinsky 2.2
self.encoder_hid_proj = ImageProjection(
image_embed_dim=encoder_hid_dim,
cross_attention_dim=cross_attention_dim,
)
elif encoder_hid_dim_type is not None:
raise ValueError(
f"encoder_hid_dim_type: {encoder_hid_dim_type} must be None, 'text_proj' or 'text_image_proj'."
)
else:
self.encoder_hid_proj = None
# class embedding
if class_embed_type is None and num_class_embeds is not None:
self.class_embedding = nn.Embedding(num_class_embeds, time_embed_dim)
elif class_embed_type == "timestep":
self.class_embedding = TimestepEmbedding(timestep_input_dim, time_embed_dim, act_fn=act_fn)
elif class_embed_type == "identity":
self.class_embedding = nn.Identity(time_embed_dim, time_embed_dim)
elif class_embed_type == "projection":
if projection_class_embeddings_input_dim is None:
raise ValueError(
"`class_embed_type`: 'projection' requires `projection_class_embeddings_input_dim` be set"
)
# The projection `class_embed_type` is the same as the timestep `class_embed_type` except
# 1. the `class_labels` inputs are not first converted to sinusoidal embeddings
# 2. it projects from an arbitrary input dimension.
#
# Note that `TimestepEmbedding` is quite general, being mainly linear layers and activations.
# When used for embedding actual timesteps, the timesteps are first converted to sinusoidal embeddings.
# As a result, `TimestepEmbedding` can be passed arbitrary vectors.
self.class_embedding = TimestepEmbedding(projection_class_embeddings_input_dim, time_embed_dim)
elif class_embed_type == "simple_projection":
if projection_class_embeddings_input_dim is None:
raise ValueError(
"`class_embed_type`: 'simple_projection' requires `projection_class_embeddings_input_dim` be set"
)
self.class_embedding = nn.Linear(projection_class_embeddings_input_dim, time_embed_dim)
else:
self.class_embedding = None
if addition_embed_type == "text":
if encoder_hid_dim is not None:
text_time_embedding_from_dim = encoder_hid_dim
else:
text_time_embedding_from_dim = cross_attention_dim
self.add_embedding = TextTimeEmbedding(
text_time_embedding_from_dim, time_embed_dim, num_heads=addition_embed_type_num_heads
)
elif addition_embed_type == "text_image":
# text_embed_dim and image_embed_dim DON'T have to be `cross_attention_dim`. To not clutter the __init__ too much
# they are set to `cross_attention_dim` here as this is exactly the required dimension for the currently only use
# case when `addition_embed_type == "text_image"` (Kadinsky 2.1)`
self.add_embedding = TextImageTimeEmbedding(
text_embed_dim=cross_attention_dim, image_embed_dim=cross_attention_dim, time_embed_dim=time_embed_dim
)
elif addition_embed_type == "text_time":
self.add_time_proj = Timesteps(addition_time_embed_dim, flip_sin_to_cos, freq_shift)
self.add_embedding = TimestepEmbedding(projection_class_embeddings_input_dim, time_embed_dim)
elif addition_embed_type == "image":
# Kandinsky 2.2
self.add_embedding = ImageTimeEmbedding(image_embed_dim=encoder_hid_dim, time_embed_dim=time_embed_dim)
elif addition_embed_type == "image_hint":
# Kandinsky 2.2 ControlNet
self.add_embedding = ImageHintTimeEmbedding(image_embed_dim=encoder_hid_dim, time_embed_dim=time_embed_dim)
elif addition_embed_type is not None:
raise ValueError(f"addition_embed_type: {addition_embed_type} must be None, 'text' or 'text_image'.")
if time_embedding_act_fn is None:
self.time_embed_act = None
else:
self.time_embed_act = get_activation(time_embedding_act_fn)
self.down_blocks = nn.ModuleList([])
self.up_blocks = nn.ModuleList([])
if isinstance(only_cross_attention, bool):
if mid_block_only_cross_attention is None:
mid_block_only_cross_attention = only_cross_attention
only_cross_attention = [only_cross_attention] * len(down_block_types)
if mid_block_only_cross_attention is None:
mid_block_only_cross_attention = False
if isinstance(num_attention_heads, int):
num_attention_heads = (num_attention_heads,) * len(down_block_types)
if isinstance(attention_head_dim, int):
attention_head_dim = (attention_head_dim,) * len(down_block_types)
if isinstance(cross_attention_dim, int):
cross_attention_dim = (cross_attention_dim,) * len(down_block_types)
if isinstance(layers_per_block, int):
layers_per_block = [layers_per_block] * len(down_block_types)
if isinstance(transformer_layers_per_block, int):
transformer_layers_per_block = [transformer_layers_per_block] * len(down_block_types)
if class_embeddings_concat:
# The time embeddings are concatenated with the class embeddings. The dimension of the
# time embeddings passed to the down, middle, and up blocks is twice the dimension of the
# regular time embeddings
blocks_time_embed_dim = time_embed_dim * 2
else:
blocks_time_embed_dim = time_embed_dim
# down
output_channel = block_out_channels[0]
for i, down_block_type in enumerate(down_block_types):
input_channel = output_channel
output_channel = block_out_channels[i]
is_final_block = i == len(block_out_channels) - 1
down_block = get_down_block(
down_block_type,
num_layers=layers_per_block[i],
transformer_layers_per_block=transformer_layers_per_block[i],
in_channels=input_channel,
out_channels=output_channel,
temb_channels=blocks_time_embed_dim,
add_downsample=not is_final_block,
resnet_eps=norm_eps,
resnet_act_fn=act_fn,
resnet_groups=norm_num_groups,
cross_attention_dim=cross_attention_dim[i],
num_attention_heads=num_attention_heads[i],
downsample_padding=downsample_padding,
dual_cross_attention=dual_cross_attention,
use_linear_projection=use_linear_projection,
only_cross_attention=only_cross_attention[i],
upcast_attention=upcast_attention,
resnet_time_scale_shift=resnet_time_scale_shift,
attention_type=attention_type,
resnet_skip_time_act=resnet_skip_time_act,
resnet_out_scale_factor=resnet_out_scale_factor,
cross_attention_norm=cross_attention_norm,
attention_head_dim=attention_head_dim[i] if attention_head_dim[i] is not None else output_channel,
dropout=dropout,
)
self.down_blocks.append(down_block)
# mid
if mid_block_type == "UNetMidBlock2DCrossAttn":
self.mid_block = UNetMidBlock2DCrossAttn(
transformer_layers_per_block=transformer_layers_per_block[-1],
in_channels=block_out_channels[-1],
temb_channels=blocks_time_embed_dim,
dropout=dropout,
resnet_eps=norm_eps,
resnet_act_fn=act_fn,
output_scale_factor=mid_block_scale_factor,
resnet_time_scale_shift=resnet_time_scale_shift,
cross_attention_dim=cross_attention_dim[-1],
num_attention_heads=num_attention_heads[-1],
resnet_groups=norm_num_groups,
dual_cross_attention=dual_cross_attention,
use_linear_projection=use_linear_projection,
upcast_attention=upcast_attention,
attention_type=attention_type,
)
elif mid_block_type == "UNetMidBlock2DSimpleCrossAttn":
self.mid_block = UNetMidBlock2DSimpleCrossAttn(
in_channels=block_out_channels[-1],
temb_channels=blocks_time_embed_dim,
dropout=dropout,
resnet_eps=norm_eps,
resnet_act_fn=act_fn,
output_scale_factor=mid_block_scale_factor,
cross_attention_dim=cross_attention_dim[-1],
attention_head_dim=attention_head_dim[-1],
resnet_groups=norm_num_groups,
resnet_time_scale_shift=resnet_time_scale_shift,
skip_time_act=resnet_skip_time_act,
only_cross_attention=mid_block_only_cross_attention,
cross_attention_norm=cross_attention_norm,
)
elif mid_block_type == "UNetMidBlock2D":
self.mid_block = UNetMidBlock2D(
in_channels=block_out_channels[-1],
temb_channels=blocks_time_embed_dim,
dropout=dropout,
num_layers=0,
resnet_eps=norm_eps,
resnet_act_fn=act_fn,
output_scale_factor=mid_block_scale_factor,
resnet_groups=norm_num_groups,
resnet_time_scale_shift=resnet_time_scale_shift,
add_attention=False,
)
elif mid_block_type is None:
self.mid_block = None
else:
raise ValueError(f"unknown mid_block_type : {mid_block_type}")
# count how many layers upsample the images
self.num_upsamplers = 0
# up
reversed_block_out_channels = list(reversed(block_out_channels))
reversed_num_attention_heads = list(reversed(num_attention_heads))
reversed_layers_per_block = list(reversed(layers_per_block))
reversed_cross_attention_dim = list(reversed(cross_attention_dim))
reversed_transformer_layers_per_block = (
list(reversed(transformer_layers_per_block))
if reverse_transformer_layers_per_block is None
else reverse_transformer_layers_per_block
)
only_cross_attention = list(reversed(only_cross_attention))
output_channel = reversed_block_out_channels[0]
for i, up_block_type in enumerate(up_block_types):
is_final_block = i == len(block_out_channels) - 1
prev_output_channel = output_channel
output_channel = reversed_block_out_channels[i]
input_channel = reversed_block_out_channels[min(i + 1, len(block_out_channels) - 1)]
# add upsample block for all BUT final layer
if not is_final_block:
add_upsample = True
self.num_upsamplers += 1
else:
add_upsample = False
up_block = get_up_block(
up_block_type,
num_layers=reversed_layers_per_block[i] + 1,
transformer_layers_per_block=reversed_transformer_layers_per_block[i],
in_channels=input_channel,
out_channels=output_channel,
prev_output_channel=prev_output_channel,
temb_channels=blocks_time_embed_dim,
add_upsample=add_upsample,
resnet_eps=norm_eps,
resnet_act_fn=act_fn,
resolution_idx=i,
resnet_groups=norm_num_groups,
cross_attention_dim=reversed_cross_attention_dim[i],
num_attention_heads=reversed_num_attention_heads[i],
dual_cross_attention=dual_cross_attention,
use_linear_projection=use_linear_projection,
only_cross_attention=only_cross_attention[i],
upcast_attention=upcast_attention,
resnet_time_scale_shift=resnet_time_scale_shift,
attention_type=attention_type,
resnet_skip_time_act=resnet_skip_time_act,
resnet_out_scale_factor=resnet_out_scale_factor,
cross_attention_norm=cross_attention_norm,
attention_head_dim=attention_head_dim[i] if attention_head_dim[i] is not None else output_channel,
dropout=dropout,
)
self.up_blocks.append(up_block)
prev_output_channel = output_channel
# out
if norm_num_groups is not None:
self.conv_norm_out = nn.GroupNorm(
num_channels=block_out_channels[0], num_groups=norm_num_groups, eps=norm_eps
)
self.conv_act = get_activation(act_fn)
else:
self.conv_norm_out = None
self.conv_act = None
conv_out_padding = (conv_out_kernel - 1) // 2
self.conv_out = nn.Conv2d(
block_out_channels[0], out_channels, kernel_size=conv_out_kernel, padding=conv_out_padding
)
if attention_type in ["gated", "gated-text-image"]:
positive_len = 768
if isinstance(cross_attention_dim, int):
positive_len = cross_attention_dim
elif isinstance(cross_attention_dim, tuple) or isinstance(cross_attention_dim, list):
positive_len = cross_attention_dim[0]
feature_type = "text-only" if attention_type == "gated" else "text-image"
self.position_net = PositionNet(
positive_len=positive_len, out_dim=cross_attention_dim, feature_type=feature_type
)
@property
def attn_processors(self) -> Dict[str, AttentionProcessor]:
r"""
Returns:
`dict` of attention processors: A dictionary containing all attention processors used in the model with
indexed by its weight name.
"""
# set recursively
processors = {}
def fn_recursive_add_processors(name: str, module: torch.nn.Module, processors: Dict[str, AttentionProcessor]):
if hasattr(module, "get_processor"):
processors[f"{name}.processor"] = module.get_processor(return_deprecated_lora=True)
for sub_name, child in module.named_children():
fn_recursive_add_processors(f"{name}.{sub_name}", child, processors)
return processors
for name, module in self.named_children():
fn_recursive_add_processors(name, module, processors)
return processors
def set_attn_processor(
self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]], _remove_lora=False
):
r"""
Sets the attention processor to use to compute attention.
Parameters:
processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`):
The instantiated processor class or a dictionary of processor classes that will be set as the processor
for **all** `Attention` layers.
If `processor` is a dict, the key needs to define the path to the corresponding cross attention
processor. This is strongly recommended when setting trainable attention processors.
"""
count = len(self.attn_processors.keys())
if isinstance(processor, dict) and len(processor) != count:
raise ValueError(
f"A dict of processors was passed, but the number of processors {len(processor)} does not match the"
f" number of attention layers: {count}. Please make sure to pass {count} processor classes."
)
def fn_recursive_attn_processor(name: str, module: torch.nn.Module, processor):
if hasattr(module, "set_processor"):
if not isinstance(processor, dict):
module.set_processor(processor, _remove_lora=_remove_lora)
else:
module.set_processor(processor.pop(f"{name}.processor"), _remove_lora=_remove_lora)
for sub_name, child in module.named_children():
fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)
for name, module in self.named_children():
fn_recursive_attn_processor(name, module, processor)
def set_default_attn_processor(self):
"""
Disables custom attention processors and sets the default attention implementation.
"""
if all(proc.__class__ in ADDED_KV_ATTENTION_PROCESSORS for proc in self.attn_processors.values()):
processor = AttnAddedKVProcessor()
elif all(proc.__class__ in CROSS_ATTENTION_PROCESSORS for proc in self.attn_processors.values()):
processor = AttnProcessor()
else:
raise ValueError(
f"Cannot call `set_default_attn_processor` when attention processors are of type {next(iter(self.attn_processors.values()))}"
)
self.set_attn_processor(processor, _remove_lora=True)
def set_attention_slice(self, slice_size):
r"""
Enable sliced attention computation.
When this option is enabled, the attention module splits the input tensor in slices to compute attention in
several steps. This is useful for saving some memory in exchange for a small decrease in speed.
Args:
slice_size (`str` or `int` or `list(int)`, *optional*, defaults to `"auto"`):
When `"auto"`, input to the attention heads is halved, so attention is computed in two steps. If
`"max"`, maximum amount of memory is saved by running only one slice at a time. If a number is
provided, uses as many slices as `attention_head_dim // slice_size`. In this case, `attention_head_dim`
must be a multiple of `slice_size`.
"""
sliceable_head_dims = []
def fn_recursive_retrieve_sliceable_dims(module: torch.nn.Module):
if hasattr(module, "set_attention_slice"):
sliceable_head_dims.append(module.sliceable_head_dim)
for child in module.children():
fn_recursive_retrieve_sliceable_dims(child)
# retrieve number of attention layers
for module in self.children():
fn_recursive_retrieve_sliceable_dims(module)
num_sliceable_layers = len(sliceable_head_dims)
if slice_size == "auto":
# half the attention head size is usually a good trade-off between
# speed and memory
slice_size = [dim // 2 for dim in sliceable_head_dims]
elif slice_size == "max":
# make smallest slice possible
slice_size = num_sliceable_layers * [1]
slice_size = num_sliceable_layers * [slice_size] if not isinstance(slice_size, list) else slice_size
if len(slice_size) != len(sliceable_head_dims):
raise ValueError(
f"You have provided {len(slice_size)}, but {self.config} has {len(sliceable_head_dims)} different"
f" attention layers. Make sure to match `len(slice_size)` to be {len(sliceable_head_dims)}."
)
for i in range(len(slice_size)):
size = slice_size[i]
dim = sliceable_head_dims[i]
if size is not None and size > dim:
raise ValueError(f"size {size} has to be smaller or equal to {dim}.")
# Recursively walk through all the children.
# Any children which exposes the set_attention_slice method
# gets the message
def fn_recursive_set_attention_slice(module: torch.nn.Module, slice_size: List[int]):
if hasattr(module, "set_attention_slice"):
module.set_attention_slice(slice_size.pop())
for child in module.children():
fn_recursive_set_attention_slice(child, slice_size)
reversed_slice_size = list(reversed(slice_size))
for module in self.children():
fn_recursive_set_attention_slice(module, reversed_slice_size)
def _set_gradient_checkpointing(self, module, value=False):
if hasattr(module, "gradient_checkpointing"):
module.gradient_checkpointing = value
def enable_freeu(self, s1, s2, b1, b2):
r"""Enables the FreeU mechanism from https://arxiv.org/abs/2309.11497.
The suffixes after the scaling factors represent the stage blocks where they are being applied.
Please refer to the [official repository](https://github.com/ChenyangSi/FreeU) for combinations of values that
are known to work well for different pipelines such as Stable Diffusion v1, v2, and Stable Diffusion XL.
Args:
s1 (`float`):
Scaling factor for stage 1 to attenuate the contributions of the skip features. This is done to
mitigate the "oversmoothing effect" in the enhanced denoising process.
s2 (`float`):
Scaling factor for stage 2 to attenuate the contributions of the skip features. This is done to
mitigate the "oversmoothing effect" in the enhanced denoising process.
b1 (`float`): Scaling factor for stage 1 to amplify the contributions of backbone features.
b2 (`float`): Scaling factor for stage 2 to amplify the contributions of backbone features.
"""
for i, upsample_block in enumerate(self.up_blocks):
setattr(upsample_block, "s1", s1)
setattr(upsample_block, "s2", s2)
setattr(upsample_block, "b1", b1)
setattr(upsample_block, "b2", b2)
def disable_freeu(self):
"""Disables the FreeU mechanism."""
freeu_keys = {"s1", "s2", "b1", "b2"}
for i, upsample_block in enumerate(self.up_blocks):
for k in freeu_keys:
if hasattr(upsample_block, k) or getattr(upsample_block, k, None) is not None:
setattr(upsample_block, k, None)
def forward(
self,
sample: torch.FloatTensor,
spatial_attn_inputs,
timestep: Union[torch.Tensor, float, int],
encoder_hidden_states: torch.Tensor,
class_labels: Optional[torch.Tensor] = None,
timestep_cond: Optional[torch.Tensor] = None,
attention_mask: Optional[torch.Tensor] = None,
cross_attention_kwargs: Optional[Dict[str, Any]] = None,
added_cond_kwargs: Optional[Dict[str, torch.Tensor]] = None,
down_block_additional_residuals: Optional[Tuple[torch.Tensor]] = None,
mid_block_additional_residual: Optional[torch.Tensor] = None,
down_intrablock_additional_residuals: Optional[Tuple[torch.Tensor]] = None,
encoder_attention_mask: Optional[torch.Tensor] = None,
return_dict: bool = True,
) -> Union[UNet2DConditionOutput, Tuple]:
r"""
The [`UNet2DConditionModel`] forward method.
Args:
sample (`torch.FloatTensor`):
The noisy input tensor with the following shape `(batch, channel, height, width)`.
timestep (`torch.FloatTensor` or `float` or `int`): The number of timesteps to denoise an input.
encoder_hidden_states (`torch.FloatTensor`):
The encoder hidden states with shape `(batch, sequence_length, feature_dim)`.
class_labels (`torch.Tensor`, *optional*, defaults to `None`):
Optional class labels for conditioning. Their embeddings will be summed with the timestep embeddings.
timestep_cond: (`torch.Tensor`, *optional*, defaults to `None`):
Conditional embeddings for timestep. If provided, the embeddings will be summed with the samples passed
through the `self.time_embedding` layer to obtain the timestep embeddings.
attention_mask (`torch.Tensor`, *optional*, defaults to `None`):
An attention mask of shape `(batch, key_tokens)` is applied to `encoder_hidden_states`. If `1` the mask
is kept, otherwise if `0` it is discarded. Mask will be converted into a bias, which adds large
negative values to the attention scores corresponding to "discard" tokens.
cross_attention_kwargs (`dict`, *optional*):
A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
`self.processor` in
[diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
added_cond_kwargs: (`dict`, *optional*):
A kwargs dictionary containing additional embeddings that if specified are added to the embeddings that
are passed along to the UNet blocks.
down_block_additional_residuals: (`tuple` of `torch.Tensor`, *optional*):
A tuple of tensors that if specified are added to the residuals of down unet blocks.
mid_block_additional_residual: (`torch.Tensor`, *optional*):
A tensor that if specified is added to the residual of the middle unet block.
encoder_attention_mask (`torch.Tensor`):
A cross-attention mask of shape `(batch, sequence_length)` is applied to `encoder_hidden_states`. If
`True` the mask is kept, otherwise if `False` it is discarded. Mask will be converted into a bias,
which adds large negative values to the attention scores corresponding to "discard" tokens.
return_dict (`bool`, *optional*, defaults to `True`):
Whether or not to return a [`~models.unet_2d_condition.UNet2DConditionOutput`] instead of a plain
tuple.
cross_attention_kwargs (`dict`, *optional*):
A kwargs dictionary that if specified is passed along to the [`AttnProcessor`].
added_cond_kwargs: (`dict`, *optional*):
A kwargs dictionary containin additional embeddings that if specified are added to the embeddings that
are passed along to the UNet blocks.
down_block_additional_residuals (`tuple` of `torch.Tensor`, *optional*):
additional residuals to be added to UNet long skip connections from down blocks to up blocks for
example from ControlNet side model(s)
mid_block_additional_residual (`torch.Tensor`, *optional*):
additional residual to be added to UNet mid block output, for example from ControlNet side model
down_intrablock_additional_residuals (`tuple` of `torch.Tensor`, *optional*):
additional residuals to be added within UNet down blocks, for example from T2I-Adapter side model(s)
Returns:
[`~models.unet_2d_condition.UNet2DConditionOutput`] or `tuple`:
If `return_dict` is True, an [`~models.unet_2d_condition.UNet2DConditionOutput`] is returned, otherwise
a `tuple` is returned where the first element is the sample tensor.
"""
# By default samples have to be AT least a multiple of the overall upsampling factor.
# The overall upsampling factor is equal to 2 ** (# num of upsampling layers).
# However, the upsampling interpolation output size can be forced to fit any upsampling size
# on the fly if necessary.
default_overall_up_factor = 2**self.num_upsamplers
# upsample size should be forwarded when sample is not a multiple of `default_overall_up_factor`
forward_upsample_size = False
upsample_size = None
for dim in sample.shape[-2:]:
if dim % default_overall_up_factor != 0:
# Forward upsample size to force interpolation output size.
forward_upsample_size = True
break
# ensure attention_mask is a bias, and give it a singleton query_tokens dimension
# expects mask of shape:
# [batch, key_tokens]
# adds singleton query_tokens dimension:
# [batch, 1, key_tokens]
# this helps to broadcast it as a bias over attention scores, which will be in one of the following shapes:
# [batch, heads, query_tokens, key_tokens] (e.g. torch sdp attn)
# [batch * heads, query_tokens, key_tokens] (e.g. xformers or classic attn)
if attention_mask is not None:
# assume that mask is expressed as:
# (1 = keep, 0 = discard)
# convert mask into a bias that can be added to attention scores:
# (keep = +0, discard = -10000.0)
attention_mask = (1 - attention_mask.to(sample.dtype)) * -10000.0
attention_mask = attention_mask.unsqueeze(1)
# convert encoder_attention_mask to a bias the same way we do for attention_mask
if encoder_attention_mask is not None:
encoder_attention_mask = (1 - encoder_attention_mask.to(sample.dtype)) * -10000.0
encoder_attention_mask = encoder_attention_mask.unsqueeze(1)
# 0. center input if necessary
if self.config.center_input_sample:
sample = 2 * sample - 1.0
# 1. time
timesteps = timestep
if not torch.is_tensor(timesteps):
# TODO: this requires sync between CPU and GPU. So try to pass timesteps as tensors if you can
# This would be a good case for the `match` statement (Python 3.10+)
is_mps = sample.device.type == "mps"
if isinstance(timestep, float):
dtype = torch.float32 if is_mps else torch.float64
else:
dtype = torch.int32 if is_mps else torch.int64
timesteps = torch.tensor([timesteps], dtype=dtype, device=sample.device)
elif len(timesteps.shape) == 0:
timesteps = timesteps[None].to(sample.device)
# broadcast to batch dimension in a way that's compatible with ONNX/Core ML
timesteps = timesteps.expand(sample.shape[0])
t_emb = self.time_proj(timesteps)
# `Timesteps` does not contain any weights and will always return f32 tensors
# but time_embedding might actually be running in fp16. so we need to cast here.
# there might be better ways to encapsulate this.
t_emb = t_emb.to(dtype=sample.dtype)
emb = self.time_embedding(t_emb, timestep_cond)
aug_emb = None
if self.class_embedding is not None:
if class_labels is None:
raise ValueError("class_labels should be provided when num_class_embeds > 0")
if self.config.class_embed_type == "timestep":
class_labels = self.time_proj(class_labels)
# `Timesteps` does not contain any weights and will always return f32 tensors
# there might be better ways to encapsulate this.
class_labels = class_labels.to(dtype=sample.dtype)
class_emb = self.class_embedding(class_labels).to(dtype=sample.dtype)
if self.config.class_embeddings_concat:
emb = torch.cat([emb, class_emb], dim=-1)
else:
emb = emb + class_emb
if self.config.addition_embed_type == "text":
aug_emb = self.add_embedding(encoder_hidden_states)
elif self.config.addition_embed_type == "text_image":
# Kandinsky 2.1 - style
if "image_embeds" not in added_cond_kwargs:
raise ValueError(
f"{self.__class__} has the config param `addition_embed_type` set to 'text_image' which requires the keyword argument `image_embeds` to be passed in `added_cond_kwargs`"
)
image_embs = added_cond_kwargs.get("image_embeds")
text_embs = added_cond_kwargs.get("text_embeds", encoder_hidden_states)
aug_emb = self.add_embedding(text_embs, image_embs)
elif self.config.addition_embed_type == "text_time":
# SDXL - style
if "text_embeds" not in added_cond_kwargs:
raise ValueError(
f"{self.__class__} has the config param `addition_embed_type` set to 'text_time' which requires the keyword argument `text_embeds` to be passed in `added_cond_kwargs`"
)
text_embeds = added_cond_kwargs.get("text_embeds")
if "time_ids" not in added_cond_kwargs:
raise ValueError(
f"{self.__class__} has the config param `addition_embed_type` set to 'text_time' which requires the keyword argument `time_ids` to be passed in `added_cond_kwargs`"
)
time_ids = added_cond_kwargs.get("time_ids")
time_embeds = self.add_time_proj(time_ids.flatten())
time_embeds = time_embeds.reshape((text_embeds.shape[0], -1))
add_embeds = torch.concat([text_embeds, time_embeds], dim=-1)
add_embeds = add_embeds.to(emb.dtype)
aug_emb = self.add_embedding(add_embeds)
elif self.config.addition_embed_type == "image":
# Kandinsky 2.2 - style
if "image_embeds" not in added_cond_kwargs:
raise ValueError(
f"{self.__class__} has the config param `addition_embed_type` set to 'image' which requires the keyword argument `image_embeds` to be passed in `added_cond_kwargs`"
)
image_embs = added_cond_kwargs.get("image_embeds")
aug_emb = self.add_embedding(image_embs)
elif self.config.addition_embed_type == "image_hint":
# Kandinsky 2.2 - style
if "image_embeds" not in added_cond_kwargs or "hint" not in added_cond_kwargs:
raise ValueError(
f"{self.__class__} has the config param `addition_embed_type` set to 'image_hint' which requires the keyword arguments `image_embeds` and `hint` to be passed in `added_cond_kwargs`"
)
image_embs = added_cond_kwargs.get("image_embeds")
hint = added_cond_kwargs.get("hint")
aug_emb, hint = self.add_embedding(image_embs, hint)
sample = torch.cat([sample, hint], dim=1)
emb = emb + aug_emb if aug_emb is not None else emb
if self.time_embed_act is not None:
emb = self.time_embed_act(emb)
if self.encoder_hid_proj is not None and self.config.encoder_hid_dim_type == "text_proj":
encoder_hidden_states = self.encoder_hid_proj(encoder_hidden_states)
elif self.encoder_hid_proj is not None and self.config.encoder_hid_dim_type == "text_image_proj":
# Kadinsky 2.1 - style
if "image_embeds" not in added_cond_kwargs:
raise ValueError(
f"{self.__class__} has the config param `encoder_hid_dim_type` set to 'text_image_proj' which requires the keyword argument `image_embeds` to be passed in `added_conditions`"
)
image_embeds = added_cond_kwargs.get("image_embeds")
encoder_hidden_states = self.encoder_hid_proj(encoder_hidden_states, image_embeds)
elif self.encoder_hid_proj is not None and self.config.encoder_hid_dim_type == "image_proj":
# Kandinsky 2.2 - style
if "image_embeds" not in added_cond_kwargs:
raise ValueError(
f"{self.__class__} has the config param `encoder_hid_dim_type` set to 'image_proj' which requires the keyword argument `image_embeds` to be passed in `added_conditions`"
)
image_embeds = added_cond_kwargs.get("image_embeds")
encoder_hidden_states = self.encoder_hid_proj(image_embeds)
# 2. pre-process
sample = self.conv_in(sample)
# 2.5 GLIGEN position net
if cross_attention_kwargs is not None and cross_attention_kwargs.get("gligen", None) is not None:
cross_attention_kwargs = cross_attention_kwargs.copy()
gligen_args = cross_attention_kwargs.pop("gligen")
cross_attention_kwargs["gligen"] = {"objs": self.position_net(**gligen_args)}
# for spatial attention
spatial_attn_idx = 0
# 3. down
lora_scale = cross_attention_kwargs.get("scale", 1.0) if cross_attention_kwargs is not None else 1.0
if USE_PEFT_BACKEND:
# weight the lora layers by setting `lora_scale` for each PEFT layer
scale_lora_layers(self, lora_scale)
is_controlnet = mid_block_additional_residual is not None and down_block_additional_residuals is not None
# using new arg down_intrablock_additional_residuals for T2I-Adapters, to distinguish from controlnets
is_adapter = down_intrablock_additional_residuals is not None
# maintain backward compatibility for legacy usage, where
# T2I-Adapter and ControlNet both use down_block_additional_residuals arg
# but can only use one or the other
if not is_adapter and mid_block_additional_residual is None and down_block_additional_residuals is not None:
deprecate(
"T2I should not use down_block_additional_residuals",
"1.3.0",
"Passing intrablock residual connections with `down_block_additional_residuals` is deprecated \
and will be removed in diffusers 1.3.0. `down_block_additional_residuals` should only be used \
for ControlNet. Please make sure use `down_intrablock_additional_residuals` instead. ",
standard_warn=False,
)
down_intrablock_additional_residuals = down_block_additional_residuals
is_adapter = True
down_block_res_samples = (sample,)
for downsample_block in self.down_blocks:
if hasattr(downsample_block, "has_cross_attention") and downsample_block.has_cross_attention:
# For t2i-adapter CrossAttnDownBlock2D
additional_residuals = {}
if is_adapter and len(down_intrablock_additional_residuals) > 0:
additional_residuals["additional_residuals"] = down_intrablock_additional_residuals.pop(0)
sample, res_samples, spatial_attn_inputs, spatial_attn_idx = downsample_block(
hidden_states=sample,
spatial_attn_inputs=spatial_attn_inputs,
spatial_attn_idx=spatial_attn_idx,
temb=emb,
encoder_hidden_states=encoder_hidden_states,
attention_mask=attention_mask,
cross_attention_kwargs=cross_attention_kwargs,
encoder_attention_mask=encoder_attention_mask,
**additional_residuals,
)
else:
sample, res_samples = downsample_block(hidden_states=sample, temb=emb, scale=lora_scale)
if is_adapter and len(down_intrablock_additional_residuals) > 0:
sample += down_intrablock_additional_residuals.pop(0)
down_block_res_samples += res_samples
if is_controlnet:
new_down_block_res_samples = ()
for down_block_res_sample, down_block_additional_residual in zip(
down_block_res_samples, down_block_additional_residuals
):
down_block_res_sample = down_block_res_sample + down_block_additional_residual
new_down_block_res_samples = new_down_block_res_samples + (down_block_res_sample,)
down_block_res_samples = new_down_block_res_samples
# 4. mid
if self.mid_block is not None:
if hasattr(self.mid_block, "has_cross_attention") and self.mid_block.has_cross_attention:
sample, spatial_attn_inputs, spatial_attn_idx = self.mid_block(
sample,
spatial_attn_inputs=spatial_attn_inputs,
spatial_attn_idx=spatial_attn_idx,
temb=emb,
encoder_hidden_states=encoder_hidden_states,
attention_mask=attention_mask,
cross_attention_kwargs=cross_attention_kwargs,
encoder_attention_mask=encoder_attention_mask,
)
else:
sample = self.mid_block(sample, emb)
# To support T2I-Adapter-XL
if (
is_adapter
and len(down_intrablock_additional_residuals) > 0
and sample.shape == down_intrablock_additional_residuals[0].shape
):
sample += down_intrablock_additional_residuals.pop(0)
if is_controlnet:
sample = sample + mid_block_additional_residual
# 5. up
for i, upsample_block in enumerate(self.up_blocks):
is_final_block = i == len(self.up_blocks) - 1
res_samples = down_block_res_samples[-len(upsample_block.resnets) :]
down_block_res_samples = down_block_res_samples[: -len(upsample_block.resnets)]
# if we have not reached the final block and need to forward the
# upsample size, we do it here
if not is_final_block and forward_upsample_size:
upsample_size = down_block_res_samples[-1].shape[2:]
if hasattr(upsample_block, "has_cross_attention") and upsample_block.has_cross_attention:
sample, spatial_attn_inputs, spatial_attn_idx = upsample_block(
hidden_states=sample,
spatial_attn_inputs=spatial_attn_inputs,
spatial_attn_idx=spatial_attn_idx,
temb=emb,
res_hidden_states_tuple=res_samples,
encoder_hidden_states=encoder_hidden_states,
cross_attention_kwargs=cross_attention_kwargs,
upsample_size=upsample_size,
attention_mask=attention_mask,
encoder_attention_mask=encoder_attention_mask,
)
else:
sample = upsample_block(
hidden_states=sample,
temb=emb,
res_hidden_states_tuple=res_samples,
upsample_size=upsample_size,
scale=lora_scale,
)
# 6. post-process
if self.conv_norm_out:
sample = self.conv_norm_out(sample)
sample = self.conv_act(sample)
sample = self.conv_out(sample)
if USE_PEFT_BACKEND:
# remove `lora_scale` from each PEFT layer
unscale_lora_layers(self, lora_scale)
if not return_dict:
return (sample,)
return UNet2DConditionOutput(sample=sample)
preprocess/humanparsing/mhp_extension/coco_style_annotation_creator/human_to_coco.py 0 → 100644
View file @ 54a066bf
import argparse
import datetime
import json
import os
from PIL import Image
import numpy as np
import pycococreatortools
def get_arguments():
parser = argparse.ArgumentParser(description="transform mask annotation to coco annotation")
parser.add_argument("--dataset", type=str, default='CIHP', help="name of dataset (CIHP, MHPv2 or VIP)")
parser.add_argument("--json_save_dir", type=str, default='../data/msrcnn_finetune_annotations',
help="path to save coco-style annotation json file")
parser.add_argument("--use_val", type=bool, default=False,
help="use train+val set for finetuning or not")
parser.add_argument("--train_img_dir", type=str, default='../data/instance-level_human_parsing/Training/Images',
help="train image path")
parser.add_argument("--train_anno_dir", type=str,
default='../data/instance-level_human_parsing/Training/Human_ids',
help="train human mask path")
parser.add_argument("--val_img_dir", type=str, default='../data/instance-level_human_parsing/Validation/Images',
help="val image path")
parser.add_argument("--val_anno_dir", type=str,
default='../data/instance-level_human_parsing/Validation/Human_ids',
help="val human mask path")
return parser.parse_args()
def main(args):
INFO = {
"description": args.split_name + " Dataset",
"url": "",
"version": "",
"year": 2019,
"contributor": "xyq",
"date_created": datetime.datetime.utcnow().isoformat(' ')
}
LICENSES = [
{
"id": 1,
"name": "",
"url": ""
}
]
CATEGORIES = [
{
'id': 1,
'name': 'person',
'supercategory': 'person',
},
]
coco_output = {
"info": INFO,
"licenses": LICENSES,
"categories": CATEGORIES,
"images": [],
"annotations": []
}
image_id = 1
segmentation_id = 1
for image_name in os.listdir(args.train_img_dir):
image = Image.open(os.path.join(args.train_img_dir, image_name))
image_info = pycococreatortools.create_image_info(
image_id, image_name, image.size
)
coco_output["images"].append(image_info)
human_mask_name = os.path.splitext(image_name)[0] + '.png'
human_mask = np.asarray(Image.open(os.path.join(args.train_anno_dir, human_mask_name)))
human_gt_labels = np.unique(human_mask)
for i in range(1, len(human_gt_labels)):
category_info = {'id': 1, 'is_crowd': 0}
binary_mask = np.uint8(human_mask == i)
annotation_info = pycococreatortools.create_annotation_info(
segmentation_id, image_id, category_info, binary_mask,
image.size, tolerance=10
)
if annotation_info is not None:
coco_output["annotations"].append(annotation_info)
segmentation_id += 1
image_id += 1
if not os.path.exists(args.json_save_dir):
os.makedirs(args.json_save_dir)
if not args.use_val:
with open('{}/{}_train.json'.format(args.json_save_dir, args.split_name), 'w') as output_json_file:
json.dump(coco_output, output_json_file)
else:
for image_name in os.listdir(args.val_img_dir):
image = Image.open(os.path.join(args.val_img_dir, image_name))
image_info = pycococreatortools.create_image_info(
image_id, image_name, image.size
)
coco_output["images"].append(image_info)
human_mask_name = os.path.splitext(image_name)[0] + '.png'
human_mask = np.asarray(Image.open(os.path.join(args.val_anno_dir, human_mask_name)))
human_gt_labels = np.unique(human_mask)
for i in range(1, len(human_gt_labels)):
category_info = {'id': 1, 'is_crowd': 0}
binary_mask = np.uint8(human_mask == i)
annotation_info = pycococreatortools.create_annotation_info(
segmentation_id, image_id, category_info, binary_mask,
image.size, tolerance=10
)
if annotation_info is not None:
coco_output["annotations"].append(annotation_info)
segmentation_id += 1
image_id += 1
with open('{}/{}_trainval.json'.format(args.json_save_dir, args.split_name), 'w') as output_json_file:
json.dump(coco_output, output_json_file)
coco_output_val = {
"info": INFO,
"licenses": LICENSES,
"categories": CATEGORIES,
"images": [],
"annotations": []
}
image_id_val = 1
segmentation_id_val = 1
for image_name in os.listdir(args.val_img_dir):
image = Image.open(os.path.join(args.val_img_dir, image_name))
image_info = pycococreatortools.create_image_info(
image_id_val, image_name, image.size
)
coco_output_val["images"].append(image_info)
human_mask_name = os.path.splitext(image_name)[0] + '.png'
human_mask = np.asarray(Image.open(os.path.join(args.val_anno_dir, human_mask_name)))
human_gt_labels = np.unique(human_mask)
for i in range(1, len(human_gt_labels)):
category_info = {'id': 1, 'is_crowd': 0}
binary_mask = np.uint8(human_mask == i)
annotation_info = pycococreatortools.create_annotation_info(
segmentation_id_val, image_id_val, category_info, binary_mask,
image.size, tolerance=10
)
if annotation_info is not None:
coco_output_val["annotations"].append(annotation_info)
segmentation_id_val += 1
image_id_val += 1
with open('{}/{}_val.json'.format(args.json_save_dir, args.split_name), 'w') as output_json_file_val:
json.dump(coco_output_val, output_json_file_val)
if __name__ == "__main__":
args = get_arguments()
main(args)
preprocess/humanparsing/mhp_extension/coco_style_annotation_creator/pycococreatortools.py 0 → 100644
View file @ 54a066bf
import re
import datetime
import numpy as np
from itertools import groupby
from skimage import measure
from PIL import Image
from pycocotools import mask
convert = lambda text: int(text) if text.isdigit() else text.lower()
natrual_key = lambda key: [convert(c) for c in re.split('([0-9]+)', key)]
def resize_binary_mask(array, new_size):
image = Image.fromarray(array.astype(np.uint8) * 255)
image = image.resize(new_size)
return np.asarray(image).astype(np.bool_)
def close_contour(contour):
if not np.array_equal(contour[0], contour[-1]):
contour = np.vstack((contour, contour[0]))
return contour
def binary_mask_to_rle(binary_mask):
rle = {'counts': [], 'size': list(binary_mask.shape)}
counts = rle.get('counts')
for i, (value, elements) in enumerate(groupby(binary_mask.ravel(order='F'))):
if i == 0 and value == 1:
counts.append(0)
counts.append(len(list(elements)))
return rle
def binary_mask_to_polygon(binary_mask, tolerance=0):
"""Converts a binary mask to COCO polygon representation
Args:
binary_mask: a 2D binary numpy array where '1's represent the object
tolerance: Maximum distance from original points of polygon to approximated
polygonal chain. If tolerance is 0, the original coordinate array is returned.
"""
polygons = []
# pad mask to close contours of shapes which start and end at an edge
padded_binary_mask = np.pad(binary_mask, pad_width=1, mode='constant', constant_values=0)
contours = measure.find_contours(padded_binary_mask, 0.5)
contours = np.subtract(contours, 1)
for contour in contours:
contour = close_contour(contour)
contour = measure.approximate_polygon(contour, tolerance)
if len(contour) < 3:
continue
contour = np.flip(contour, axis=1)
segmentation = contour.ravel().tolist()
# after padding and subtracting 1 we may get -0.5 points in our segmentation
segmentation = [0 if i < 0 else i for i in segmentation]
polygons.append(segmentation)
return polygons
def create_image_info(image_id, file_name, image_size,
date_captured=datetime.datetime.utcnow().isoformat(' '),
license_id=1, coco_url="", flickr_url=""):
image_info = {
"id": image_id,
"file_name": file_name,
"width": image_size[0],
"height": image_size[1],
"date_captured": date_captured,
"license": license_id,
"coco_url": coco_url,
"flickr_url": flickr_url
}
return image_info
def create_annotation_info(annotation_id, image_id, category_info, binary_mask,
image_size=None, tolerance=2, bounding_box=None):
if image_size is not None:
binary_mask = resize_binary_mask(binary_mask, image_size)
binary_mask_encoded = mask.encode(np.asfortranarray(binary_mask.astype(np.uint8)))
area = mask.area(binary_mask_encoded)
if area < 1:
return None
if bounding_box is None:
bounding_box = mask.toBbox(binary_mask_encoded)
if category_info["is_crowd"]:
is_crowd = 1
segmentation = binary_mask_to_rle(binary_mask)
else:
is_crowd = 0
segmentation = binary_mask_to_polygon(binary_mask, tolerance)
if not segmentation:
return None
annotation_info = {
"id": annotation_id,
"image_id": image_id,
"category_id": category_info["id"],
"iscrowd": is_crowd,
"area": area.tolist(),
"bbox": bounding_box.tolist(),
"segmentation": segmentation,
"width": binary_mask.shape[1],
"height": binary_mask.shape[0],
}
return annotation_info
preprocess/humanparsing/mhp_extension/coco_style_annotation_creator/test_human2coco_format.py 0 → 100644
View file @ 54a066bf
import argparse
import datetime
import json
import os
from PIL import Image
import pycococreatortools
def get_arguments():
parser = argparse.ArgumentParser(description="transform mask annotation to coco annotation")
parser.add_argument("--dataset", type=str, default='CIHP', help="name of dataset (CIHP, MHPv2 or VIP)")
parser.add_argument("--json_save_dir", type=str, default='../data/CIHP/annotations',
help="path to save coco-style annotation json file")
parser.add_argument("--test_img_dir", type=str, default='../data/CIHP/Testing/Images',
help="test image path")
return parser.parse_args()
args = get_arguments()
INFO = {
"description": args.dataset + "Dataset",
"url": "",
"version": "",
"year": 2020,
"contributor": "yunqiuxu",
"date_created": datetime.datetime.utcnow().isoformat(' ')
}
LICENSES = [
{
"id": 1,
"name": "",
"url": ""
}
]
CATEGORIES = [
{
'id': 1,
'name': 'person',
'supercategory': 'person',
},
]
def main(args):
coco_output = {
"info": INFO,
"licenses": LICENSES,
"categories": CATEGORIES,
"images": [],
"annotations": []
}
image_id = 1
for image_name in os.listdir(args.test_img_dir):
image = Image.open(os.path.join(args.test_img_dir, image_name))
image_info = pycococreatortools.create_image_info(
image_id, image_name, image.size
)
coco_output["images"].append(image_info)
image_id += 1
if not os.path.exists(os.path.join(args.json_save_dir)):
os.mkdir(os.path.join(args.json_save_dir))
with open('{}/{}.json'.format(args.json_save_dir, args.dataset), 'w') as output_json_file:
json.dump(coco_output, output_json_file)
if __name__ == "__main__":
main(args)
preprocess/humanparsing/mhp_extension/detectron2/.circleci/config.yml 0 → 100644
View file @ 54a066bf
# Python CircleCI 2.0 configuration file
#
# Check https://circleci.com/docs/2.0/language-python/ for more details
#
version: 2
# -------------------------------------------------------------------------------------
# Environments to run the jobs in
# -------------------------------------------------------------------------------------
cpu: &cpu
docker:
- image: circleci/python:3.6.8-stretch
resource_class: medium
gpu: &gpu
machine:
image: ubuntu-1604:201903-01
docker_layer_caching: true
resource_class: gpu.small
# -------------------------------------------------------------------------------------
# Re-usable commands
# -------------------------------------------------------------------------------------
install_python: &install_python
- run:
name: Install Python
working_directory: ~/
command: |
pyenv install 3.6.1
pyenv global 3.6.1
setup_venv: &setup_venv
- run:
name: Setup Virtual Env
working_directory: ~/
command: |
python -m venv ~/venv
echo ". ~/venv/bin/activate" >> $BASH_ENV
. ~/venv/bin/activate
python --version
which python
which pip
pip install --upgrade pip
install_dep: &install_dep
- run:
name: Install Dependencies
command: |
pip install --progress-bar off -U 'git+https://github.com/facebookresearch/fvcore'
pip install --progress-bar off cython opencv-python
pip install --progress-bar off 'git+https://github.com/cocodataset/cocoapi.git#subdirectory=PythonAPI'
pip install --progress-bar off torch torchvision
install_detectron2: &install_detectron2
- run:
name: Install Detectron2
command: |
gcc --version
pip install -U --progress-bar off -e .[dev]
python -m detectron2.utils.collect_env
install_nvidia_driver: &install_nvidia_driver
- run:
name: Install nvidia driver
working_directory: ~/
command: |
wget -q 'https://s3.amazonaws.com/ossci-linux/nvidia_driver/NVIDIA-Linux-x86_64-430.40.run'
sudo /bin/bash ./NVIDIA-Linux-x86_64-430.40.run -s --no-drm
nvidia-smi
run_unittests: &run_unittests
- run:
name: Run Unit Tests
command: |
python -m unittest discover -v -s tests
# -------------------------------------------------------------------------------------
# Jobs to run
# -------------------------------------------------------------------------------------
jobs:
cpu_tests:
<<: *cpu
working_directory: ~/detectron2
steps:
- checkout
- <<: *setup_venv
# Cache the venv directory that contains dependencies
- restore_cache:
keys:
- cache-key-{{ .Branch }}-ID-20200425
- <<: *install_dep
- save_cache:
paths:
- ~/venv
key: cache-key-{{ .Branch }}-ID-20200425
- <<: *install_detectron2
- run:
name: isort
command: |
isort -c -sp .
- run:
name: black
command: |
black --check -l 100 .
- run:
name: flake8
command: |
flake8 .
- <<: *run_unittests
gpu_tests:
<<: *gpu
working_directory: ~/detectron2
steps:
- checkout
- <<: *install_nvidia_driver
- run:
name: Install nvidia-docker
working_directory: ~/
command: |
curl -s -L https://nvidia.github.io/nvidia-docker/gpgkey | sudo apt-key add -
distribution=$(. /etc/os-release;echo $ID$VERSION_ID)
curl -s -L https://nvidia.github.io/nvidia-docker/$distribution/nvidia-docker.list | \
sudo tee /etc/apt/sources.list.d/nvidia-docker.list
sudo apt-get update && sudo apt-get install -y nvidia-docker2
# reload the docker daemon configuration
sudo pkill -SIGHUP dockerd
- run:
name: Launch docker
working_directory: ~/detectron2/docker
command: |
nvidia-docker build -t detectron2:v0 -f Dockerfile-circleci .
nvidia-docker run -itd --name d2 detectron2:v0
docker exec -it d2 nvidia-smi
- run:
name: Build Detectron2
command: |
docker exec -it d2 pip install 'git+https://github.com/facebookresearch/fvcore'
docker cp ~/detectron2 d2:/detectron2
# This will build d2 for the target GPU arch only
docker exec -it d2 pip install -e /detectron2
docker exec -it d2 python3 -m detectron2.utils.collect_env
docker exec -it d2 python3 -c 'import torch; assert(torch.cuda.is_available())'
- run:
name: Run Unit Tests
command: |
docker exec -e CIRCLECI=true -it d2 python3 -m unittest discover -v -s /detectron2/tests
workflows:
version: 2
regular_test:
jobs:
- cpu_tests
- gpu_tests
#nightly_test:
#jobs:
#- gpu_tests
#triggers:
#- schedule:
#cron: "0 0 * * *"
#filters:
#branches:
#only:
#- master
preprocess/humanparsing/mhp_extension/detectron2/.clang-format 0 → 100644
View file @ 54a066bf
AccessModifierOffset: -1
AlignAfterOpenBracket: AlwaysBreak
AlignConsecutiveAssignments: false
AlignConsecutiveDeclarations: false
AlignEscapedNewlinesLeft: true
AlignOperands: false
AlignTrailingComments: false
AllowAllParametersOfDeclarationOnNextLine: false
AllowShortBlocksOnASingleLine: false
AllowShortCaseLabelsOnASingleLine: false
AllowShortFunctionsOnASingleLine: Empty
AllowShortIfStatementsOnASingleLine: false
AllowShortLoopsOnASingleLine: false
AlwaysBreakAfterReturnType: None
AlwaysBreakBeforeMultilineStrings: true
AlwaysBreakTemplateDeclarations: true
BinPackArguments: false
BinPackParameters: false
BraceWrapping:
AfterClass: false
AfterControlStatement: false
AfterEnum: false
AfterFunction: false
AfterNamespace: false
AfterObjCDeclaration: false
AfterStruct: false
AfterUnion: false
BeforeCatch: false
BeforeElse: false
IndentBraces: false
BreakBeforeBinaryOperators: None
BreakBeforeBraces: Attach
BreakBeforeTernaryOperators: true
BreakConstructorInitializersBeforeComma: false
BreakAfterJavaFieldAnnotations: false
BreakStringLiterals: false
ColumnLimit: 80
CommentPragmas: '^ IWYU pragma:'
ConstructorInitializerAllOnOneLineOrOnePerLine: true
ConstructorInitializerIndentWidth: 4
ContinuationIndentWidth: 4
Cpp11BracedListStyle: true
DerivePointerAlignment: false
DisableFormat: false
ForEachMacros: [ FOR_EACH, FOR_EACH_ENUMERATE, FOR_EACH_KV, FOR_EACH_R, FOR_EACH_RANGE, ]
IncludeCategories:
- Regex: '^<.*\.h(pp)?>'
Priority: 1
- Regex: '^<.*'
Priority: 2
- Regex: '.*'
Priority: 3
IndentCaseLabels: true
IndentWidth: 2
IndentWrappedFunctionNames: false
KeepEmptyLinesAtTheStartOfBlocks: false
MacroBlockBegin: ''
MacroBlockEnd: ''
MaxEmptyLinesToKeep: 1
NamespaceIndentation: None
ObjCBlockIndentWidth: 2
ObjCSpaceAfterProperty: false
ObjCSpaceBeforeProtocolList: false
PenaltyBreakBeforeFirstCallParameter: 1
PenaltyBreakComment: 300
PenaltyBreakFirstLessLess: 120
PenaltyBreakString: 1000
PenaltyExcessCharacter: 1000000
PenaltyReturnTypeOnItsOwnLine: 200
PointerAlignment: Left
ReflowComments: true
SortIncludes: true
SpaceAfterCStyleCast: false
SpaceBeforeAssignmentOperators: true
SpaceBeforeParens: ControlStatements
SpaceInEmptyParentheses: false
SpacesBeforeTrailingComments: 1
SpacesInAngles: false
SpacesInContainerLiterals: true
SpacesInCStyleCastParentheses: false
SpacesInParentheses: false
SpacesInSquareBrackets: false
Standard: Cpp11
TabWidth: 8
UseTab: Never
preprocess/humanparsing/mhp_extension/detectron2/.flake8 0 → 100644
View file @ 54a066bf
# This is an example .flake8 config, used when developing *Black* itself.
# Keep in sync with setup.cfg which is used for source packages.
[flake8]
ignore = W503, E203, E221, C901, C408, E741
max-line-length = 100
max-complexity = 18
select = B,C,E,F,W,T4,B9
exclude = build,__init__.py
preprocess/humanparsing/mhp_extension/detectron2/.github/CODE_OF_CONDUCT.md 0 → 100644
View file @ 54a066bf
# Code of Conduct
Facebook has adopted a Code of Conduct that we expect project participants to adhere to.
Please read the [full text](https://code.fb.com/codeofconduct/)
so that you can understand what actions will and will not be tolerated.
preprocess/humanparsing/mhp_extension/detectron2/.github/CONTRIBUTING.md 0 → 100644
View file @ 54a066bf
# Contributing to detectron2
## Issues
We use GitHub issues to track public bugs and questions.
Please make sure to follow one of the
[issue templates](https://github.com/facebookresearch/detectron2/issues/new/choose)
when reporting any issues.
Facebook has a [bounty program](https://www.facebook.com/whitehat/) for the safe
disclosure of security bugs. In those cases, please go through the process
outlined on that page and do not file a public issue.
## Pull Requests
We actively welcome your pull requests.
However, if you're adding any significant features (e.g. > 50 lines), please
make sure to have a corresponding issue to discuss your motivation and proposals,
before sending a PR. We do not always accept new features, and we take the following
factors into consideration:
1. Whether the same feature can be achieved without modifying detectron2.
Detectron2 is designed so that you can implement many extensions from the outside, e.g.
those in [projects](https://github.com/facebookresearch/detectron2/tree/master/projects).
If some part is not as extensible, you can also bring up the issue to make it more extensible.
2. Whether the feature is potentially useful to a large audience, or only to a small portion of users.
3. Whether the proposed solution has a good design / interface.
4. Whether the proposed solution adds extra mental/practical overhead to users who don't
need such feature.
5. Whether the proposed solution breaks existing APIs.
When sending a PR, please do:
1. If a PR contains multiple orthogonal changes, split it to several PRs.
2. If you've added code that should be tested, add tests.
3. For PRs that need experiments (e.g. adding a new model or new methods),
you don't need to update model zoo, but do provide experiment results in the description of the PR.
4. If APIs are changed, update the documentation.
5. Make sure your code lints with `./dev/linter.sh`.
## Contributor License Agreement ("CLA")
In order to accept your pull request, we need you to submit a CLA. You only need
to do this once to work on any of Facebook's open source projects.
Complete your CLA here: <https://code.facebook.com/cla>
## License
By contributing to detectron2, you agree that your contributions will be licensed
under the LICENSE file in the root directory of this source tree.
preprocess/humanparsing/mhp_extension/detectron2/.github/Detectron2-Logo-Horz.svg 0 → 100644
View file @ 54a066bf
<svg id="Layer_1" data-name="Layer 1" xmlns="http://www.w3.org/2000/svg" viewBox="0 0 1930.09 354.96"><defs><style>.cls-1{fill:#aab4bc;}.cls-2{fill:#d2d6d7;}.cls-3{fill:#9da2ab;}.cls-4{fill:#e7eef1;}.cls-5{fill:#5173f1;}.cls-6{opacity:0.7;}.cls-7{fill:#797f89;}.cls-8{fill:#e3e7e9;}.cls-9{fill:#161622;}.cls-10{fill:#3f4652;}.cls-11{fill:#fff;}</style></defs><title>Detectron2-Logo-Horz</title><path class="cls-1" d="M191.24,31h71.34a4.87,4.87,0,0,1,4.87,4.87v5a0,0,0,0,1,0,0H186.38a0,0,0,0,1,0,0v-5A4.87,4.87,0,0,1,191.24,31Z"/><path class="cls-2" d="M412.92,100.67V263.61c0,.69,0,1.33,0,2a59.73,59.73,0,0,1-59.73,57.74H100.73A59.8,59.8,0,0,1,40.9,263.61V100.67c0-.69,0-1.33,0-2a59.33,59.33,0,0,1,8.79-29.21c.76-1.24,1.57-2.46,2.42-3.64A59.76,59.76,0,0,1,100.73,40.9H353.15a59.78,59.78,0,0,1,59.77,59.77Z"/><rect class="cls-3" x="198.81" y="262.89" width="55.95" height="41.28" rx="10.15"/><path class="cls-4" d="M244.61,260.72H209A12.33,12.33,0,0,0,196.64,273v21A12.33,12.33,0,0,0,209,306.33h35.65A12.32,12.32,0,0,0,256.92,294V273A12.32,12.32,0,0,0,244.61,260.72ZM209,265.05h35.65a8,8,0,0,1,8,8v1.45H201V273A8,8,0,0,1,209,265.05Zm43.63,13.76v9.43H201v-9.43Zm-8,23.19H209a8,8,0,0,1-8-8v-1.44h51.61V294A8,8,0,0,1,244.61,302Z"/><path class="cls-1" d="M382.21,177.18h71.34a4.87,4.87,0,0,1,4.87,4.87v5a0,0,0,0,1,0,0H377.35a0,0,0,0,1,0,0v-5A4.87,4.87,0,0,1,382.21,177.18Z" transform="translate(600.02 -235.74) rotate(90)"/><path class="cls-1" d="M.28,177.18H71.62a4.87,4.87,0,0,1,4.87,4.87v5a0,0,0,0,1,0,0H-4.59a0,0,0,0,1,0,0v-5A4.87,4.87,0,0,1,.28,177.18Z" transform="translate(-146.19 218.09) rotate(-90)"/><circle class="cls-1" cx="83.04" cy="283.53" r="6.28"/><circle class="cls-1" cx="370.79" cy="283.53" r="6.28"/><circle class="cls-1" cx="226.91" cy="66.06" r="6.28"/><circle class="cls-5" cx="368.44" cy="82.89" r="20.49"/><polygon class="cls-1" points="412.92 179.98 316.61 179.98 312.27 179.98 141.55 179.98 137.21 179.98 40.9 179.98 40.9 184.3 137.21 184.3 137.21 323.38 141.55 323.38 141.55 184.3 312.27 184.3 312.27 323.38 316.61 323.38 316.61 184.3 412.92 184.3 412.92 179.98"/><g class="cls-6"><path class="cls-7" d="M403.72,193a81.13,81.13,0,1,1-81.15-81.1A81.12,81.12,0,0,1,403.72,193Z"/></g><path class="cls-8" d="M313.71,104.06a76.74,76.74,0,1,0,76.74,76.74A76.75,76.75,0,0,0,313.71,104.06Zm0,132.48a55.74,55.74,0,1,1,55.73-55.74A55.8,55.8,0,0,1,313.71,236.54Z"/><path class="cls-9" d="M376.27,180.79a62.57,62.57,0,1,1-125.13,0,61,61,0,0,1,1.93-15.33,62.55,62.55,0,0,1,123.2,15.33Z"/><path class="cls-3" d="M313.71,121.19a59.6,59.6,0,1,1-59.6,59.6A57.93,57.93,0,0,1,256,166.18a59.72,59.72,0,0,1,57.76-45m0-3.65a63.36,63.36,0,0,0-61.3,47.75,61.81,61.81,0,0,0-1.95,15.5,63.25,63.25,0,1,0,63.25-63.25Z"/><g class="cls-6"><path class="cls-7" d="M228.66,193a81.12,81.12,0,1,1-81.14-81.1A81.11,81.11,0,0,1,228.66,193Z"/></g><path class="cls-8" d="M138.65,104.06A76.74,76.74,0,1,0,215.4,180.8,76.74,76.74,0,0,0,138.65,104.06Zm0,132.48a55.74,55.74,0,1,1,55.74-55.74A55.8,55.8,0,0,1,138.65,236.54Z"/><path class="cls-9" d="M201.22,180.79a62.57,62.57,0,1,1-125.13,0A61,61,0,0,1,78,165.46a62.55,62.55,0,0,1,123.2,15.33Z"/><path class="cls-3" d="M138.65,121.19a59.6,59.6,0,1,1-59.6,59.6,58.38,58.38,0,0,1,1.84-14.61,59.72,59.72,0,0,1,57.76-45m0-3.65a63.39,63.39,0,0,0-61.3,47.75,62.28,62.28,0,0,0-1.94,15.5,63.25,63.25,0,1,0,63.24-63.25Z"/><circle class="cls-10" cx="313.71" cy="180.79" r="29"/><circle class="cls-10" cx="138.65" cy="180.79" r="29"/><circle class="cls-11" cx="154.83" cy="156.49" r="12.7"/><circle class="cls-11" cx="329.89" cy="156.49" r="12.7"/><path class="cls-1" d="M312.27,40.91V81.77a100.32,100.32,0,0,0-72.71,33.61H214.3A100.51,100.51,0,0,0,142,81.82V40.9h-4.33V81.77A99.56,99.56,0,0,0,86.17,97.06l-34-31.27c-.85,1.18-1.66,2.4-2.42,3.64l36,33.1,0,0a95.88,95.88,0,0,1,126,16.46l.65.74h29.18l.65-.74a96,96,0,0,1,72.27-32.89h2.17V40.91Z"/><path class="cls-5" d="M1899.11,280.92H1758.56V251.65l81.53-77.75q19.44-18.35,19.44-39.32,0-14.55-9-23.29t-24.15-8.74q-16.59,0-25,9.6t-8.44,25.32l.87,9.6h-35.21a77.72,77.72,0,0,1-.58-10.19q0-30,18.77-48.45T1826.36,70q32,0,50.48,17.75t18.48,46q0,20.08-8,35.49t-27.22,32.29l-52.95,46.87h92Z"/><path class="cls-10" d="M557.9,280.92H487.77V74.32H557.9q52.38,0,81.62,28.37t29.24,74.93q0,46.56-29.24,74.93T557.9,280.92Zm54.85-51.36q18.76-18.76,18.77-51.94t-18.77-51.94q-18.76-18.77-56-18.77H523V248.33h33.76Q594,248.33,612.75,229.56Z"/><path class="cls-10" d="M826.87,215.45H711.93q2,18,13.1,28.66t29.1,10.62a40.72,40.72,0,0,0,21.53-5.82,32.61,32.61,0,0,0,13.68-15.71h34.91a70.46,70.46,0,0,1-26,37.1q-19.07,14.11-45,14.11-33.75,0-54.56-22.41t-20.8-56.74q0-33.45,21-56.15T753,126.41q33.18,0,53.69,22.26t20.51,56ZM753,154.63q-16.29,0-27.06,9.61t-13.38,25.6h80.31q-2.34-16-12.81-25.6T753,154.63Z"/><path class="cls-10" d="M915.19,250.08v30q-6.41,1.74-18,1.74-44.24,0-44.23-44.52V157.54h-23V129.9h23V90.62h34.63V129.9h28.22v27.64H887.55v76.24q0,17.76,16.88,17.75Z"/><path class="cls-10" d="M1075.48,215.45H960.54q2,18,13.09,28.66t29.1,10.62a40.72,40.72,0,0,0,21.53-5.82,32.55,32.55,0,0,0,13.68-15.71h34.92a70.48,70.48,0,0,1-26,37.1q-19.05,14.11-44.95,14.11-33.76,0-54.56-22.41t-20.8-56.74q0-33.45,20.94-56.15t54.13-22.7q33.16,0,53.68,22.26t20.52,56Zm-73.91-60.82q-16.3,0-27.06,9.61t-13.39,25.6h80.31q-2.33-16-12.8-25.6T1001.57,154.63Z"/><path class="cls-10" d="M1086.1,205.56q0-33.47,21.24-56.31t54.13-22.84q31.13,0,49.61,17.6t22,40.59h-35.5a36,36,0,0,0-13-20.08q-9.75-7.56-23.42-7.56-18.33,0-29.39,13.53t-11.06,35.07q0,21.52,11.06,34.91t29.39,13.39q13.66,0,23.42-7.57a35.88,35.88,0,0,0,13-20.08h35.5q-3.49,23-22,40.6t-49.61,17.6q-32.9,0-54.13-22.84T1086.1,205.56Z"/><path class="cls-10" d="M1322.58,250.08v30q-6.4,1.74-18,1.74-44.22,0-44.23-44.52V157.54h-23V129.9h23V90.62h34.63V129.9h28.23v27.64h-28.23v76.24q0,17.76,16.88,17.75Z"/><path class="cls-10" d="M1428.44,128.74V161a55.4,55.4,0,0,0-7.85-.59q-39,0-39,41.91v78.56H1347v-151h32v20.95q12.8-22.41,44.51-22.41Z"/><path class="cls-10" d="M1507.79,284.41q-34.92,0-56.6-23t-21.67-55.86q0-32.9,21.67-56t56.6-23.13q35.2,0,56.89,23.13t21.68,56q0,32.88-21.68,55.86T1507.79,284.41Zm-43.65-78.85q0,21.52,12.37,34.91t31.28,13.39q19.2,0,31.57-13.39t12.37-34.91q0-21.84-12.37-35.21T1507.79,157q-18.91,0-31.28,13.39T1464.14,205.56Z"/><path class="cls-10" d="M1631.22,129.9V150q5.25-9.9,17.32-16.74t29.24-6.83q26.78,0,41.47,16.29t14.69,43.36v94.86h-34.63v-90.5q0-16-7.42-25.17t-22.55-9.16q-16.58,0-26,9.89t-9.46,27.35v87.59h-34.62v-151Z"/></svg>
\ No newline at end of file
preprocess/humanparsing/mhp_extension/detectron2/.github/ISSUE_TEMPLATE.md 0 → 100644
View file @ 54a066bf
Please select an issue template from
https://github.com/facebookresearch/detectron2/issues/new/choose .
Otherwise your issue will be closed.
preprocess/humanparsing/mhp_extension/detectron2/.github/ISSUE_TEMPLATE/bugs.md 0 → 100644
View file @ 54a066bf
---
name: "🐛 Bugs"
about: Report bugs in detectron2
title: Please read & provide the following
---
## Instructions To Reproduce the 🐛 Bug:
1. what changes you made (`git diff`) or what code you wrote
```
<put diff or code here>
```
2. what exact command you run:
3. what you observed (including __full logs__):
```
<put logs here>
```
4. please simplify the steps as much as possible so they do not require additional resources to
run, such as a private dataset.
## Expected behavior:
If there are no obvious error in "what you observed" provided above,
please tell us the expected behavior.
## Environment:
Provide your environment information using the following command:
```
wget -nc -q https://github.com/facebookresearch/detectron2/raw/master/detectron2/utils/collect_env.py && python collect_env.py
```
If your issue looks like an installation issue / environment issue,
please first try to solve it yourself with the instructions in
https://detectron2.readthedocs.io/tutorials/install.html#common-installation-issues
preprocess/humanparsing/mhp_extension/detectron2/.github/ISSUE_TEMPLATE/config.yml 0 → 100644
View file @ 54a066bf
# require an issue template to be chosen
blank_issues_enabled: false
# Unexpected behaviors & bugs are split to two templates.
# When they are one template, users think "it's not a bug" and don't choose the template.
#
# But the file name is still "unexpected-problems-bugs.md" so that old references
# to this issue template still works.
# It's ok since this template should be a superset of "bugs.md" (unexpected behaviors is a superset of bugs)
preprocess/humanparsing/mhp_extension/detectron2/.github/ISSUE_TEMPLATE/feature-request.md 0 → 100644
View file @ 54a066bf
---
name: "\U0001F680Feature Request"
about: Submit a proposal/request for a new detectron2 feature
---
## 🚀 Feature
A clear and concise description of the feature proposal.
## Motivation & Examples
Tell us why the feature is useful.
Describe what the feature would look like, if it is implemented.
Best demonstrated using **code examples** in addition to words.
## Note
We only consider adding new features if they are relevant to many users.
If you request implementation of research papers --
we only consider papers that have enough significance and prevalance in the object detection field.
We do not take requests for most projects in the `projects/` directory,
because they are research code release that is mainly for other researchers to reproduce results.
Instead of adding features inside detectron2,
you can implement many features by [extending detectron2](https://detectron2.readthedocs.io/tutorials/extend.html).
The [projects/](https://github.com/facebookresearch/detectron2/tree/master/projects/) directory contains many of such examples.
preprocess/humanparsing/mhp_extension/detectron2/.github/ISSUE_TEMPLATE/questions-help-support.md 0 → 100644
View file @ 54a066bf
---
name: "❓How to do something?"
about: How to do something using detectron2? What does an API do?
---
## ❓ How to do something using detectron2
Describe what you want to do, including:
1. what inputs you will provide, if any:
2. what outputs you are expecting:
## ❓ What does an API do and how to use it?
Please link to which API or documentation you're asking about from
https://detectron2.readthedocs.io/
NOTE:
1. Only general answers are provided.
If you want to ask about "why X did not work", please use the
[Unexpected behaviors](https://github.com/facebookresearch/detectron2/issues/new/choose) issue template.
2. About how to implement new models / new dataloader / new training logic, etc., check documentation first.
3. We do not answer general machine learning / computer vision questions that are not specific to detectron2, such as how a model works, how to improve your training/make it converge, or what algorithm/methods can be used to achieve X.
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment