Skip to content

GitLab

Commit a44e71bd authored by weishb's avatar weishb
Browse files

首次提交

parent df12fdb3
Show whitespace changes
Inline Side-by-side
Showing with 2136 additions and 1 deletion
LICENSE 0 → 100644
View file @ a44e71bd
Apache License
Version 2.0, January 2004
http://www.apache.org/licenses/
TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
1. Definitions.
"License" shall mean the terms and conditions for use, reproduction,
and distribution as defined by Sections 1 through 9 of this document.
"Licensor" shall mean the copyright owner or entity authorized by
the copyright owner that is granting the License.
"Legal Entity" shall mean the union of the acting entity and all
other entities that control, are controlled by, or are under common
control with that entity. For the purposes of this definition,
"control" means (i) the power, direct or indirect, to cause the
direction or management of such entity, whether by contract or
otherwise, or (ii) ownership of fifty percent (50%) or more of the
outstanding shares, or (iii) beneficial ownership of such entity.
"You" (or "Your") shall mean an individual or Legal Entity
exercising permissions granted by this License.
"Source" form shall mean the preferred form for making modifications,
including but not limited to software source code, documentation
source, and configuration files.
"Object" form shall mean any form resulting from mechanical
transformation or translation of a Source form, including but
not limited to compiled object code, generated documentation,
and conversions to other media types.
"Work" shall mean the work of authorship, whether in Source or
Object form, made available under the License, as indicated by a
copyright notice that is included in or attached to the work
(an example is provided in the Appendix below).
"Derivative Works" shall mean any work, whether in Source or Object
form, that is based on (or derived from) the Work and for which the
editorial revisions, annotations, elaborations, or other modifications
represent, as a whole, an original work of authorship. For the purposes
of this License, Derivative Works shall not include works that remain
separable from, or merely link (or bind by name) to the interfaces of,
the Work and Derivative Works thereof.
"Contribution" shall mean any work of authorship, including
the original version of the Work and any modifications or additions
to that Work or Derivative Works thereof, that is intentionally
submitted to Licensor for inclusion in the Work by the copyright owner
or by an individual or Legal Entity authorized to submit on behalf of
the copyright owner. For the purposes of this definition, "submitted"
means any form of electronic, verbal, or written communication sent
to the Licensor or its representatives, including but not limited to
communication on electronic mailing lists, source code control systems,
and issue tracking systems that are managed by, or on behalf of, the
Licensor for the purpose of discussing and improving the Work, but
excluding communication that is conspicuously marked or otherwise
designated in writing by the copyright owner as "Not a Contribution."
"Contributor" shall mean Licensor and any individual or Legal Entity
on behalf of whom a Contribution has been received by Licensor and
subsequently incorporated within the Work.
2. Grant of Copyright License. Subject to the terms and conditions of
this License, each Contributor hereby grants to You a perpetual,
worldwide, non-exclusive, no-charge, royalty-free, irrevocable
copyright license to reproduce, prepare Derivative Works of,
publicly display, publicly perform, sublicense, and distribute the
Work and such Derivative Works in Source or Object form.
3. Grant of Patent License. Subject to the terms and conditions of
this License, each Contributor hereby grants to You a perpetual,
worldwide, non-exclusive, no-charge, royalty-free, irrevocable
(except as stated in this section) patent license to make, have made,
use, offer to sell, sell, import, and otherwise transfer the Work,
where such license applies only to those patent claims licensable
by such Contributor that are necessarily infringed by their
Contribution(s) alone or by combination of their Contribution(s)
with the Work to which such Contribution(s) was submitted. If You
institute patent litigation against any entity (including a
cross-claim or counterclaim in a lawsuit) alleging that the Work
or a Contribution incorporated within the Work constitutes direct
or contributory patent infringement, then any patent licenses
granted to You under this License for that Work shall terminate
as of the date such litigation is filed.
4. Redistribution. You may reproduce and distribute copies of the
Work or Derivative Works thereof in any medium, with or without
modifications, and in Source or Object form, provided that You
meet the following conditions:
(a) You must give any other recipients of the Work or
Derivative Works a copy of this License; and
(b) You must cause any modified files to carry prominent notices
stating that You changed the files; and
(c) You must retain, in the Source form of any Derivative Works
that You distribute, all copyright, patent, trademark, and
attribution notices from the Source form of the Work,
excluding those notices that do not pertain to any part of
the Derivative Works; and
(d) If the Work includes a "NOTICE" text file as part of its
distribution, then any Derivative Works that You distribute must
include a readable copy of the attribution notices contained
within such NOTICE file, excluding those notices that do not
pertain to any part of the Derivative Works, in at least one
of the following places: within a NOTICE text file distributed
as part of the Derivative Works; within the Source form or
documentation, if provided along with the Derivative Works; or,
within a display generated by the Derivative Works, if and
wherever such third-party notices normally appear. The contents
of the NOTICE file are for informational purposes only and
do not modify the License. You may add Your own attribution
notices within Derivative Works that You distribute, alongside
or as an addendum to the NOTICE text from the Work, provided
that such additional attribution notices cannot be construed
as modifying the License.
You may add Your own copyright statement to Your modifications and
may provide additional or different license terms and conditions
for use, reproduction, or distribution of Your modifications, or
for any such Derivative Works as a whole, provided Your use,
reproduction, and distribution of the Work otherwise complies with
the conditions stated in this License.
5. Submission of Contributions. Unless You explicitly state otherwise,
any Contribution intentionally submitted for inclusion in the Work
by You to the Licensor shall be under the terms and conditions of
this License, without any additional terms or conditions.
Notwithstanding the above, nothing herein shall supersede or modify
the terms of any separate license agreement you may have executed
with Licensor regarding such Contributions.
6. Trademarks. This License does not grant permission to use the trade
names, trademarks, service marks, or product names of the Licensor,
except as required for reasonable and customary use in describing the
origin of the Work and reproducing the content of the NOTICE file.
7. Disclaimer of Warranty. Unless required by applicable law or
agreed to in writing, Licensor provides the Work (and each
Contributor provides its Contributions) on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
implied, including, without limitation, any warranties or conditions
of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
PARTICULAR PURPOSE. You are solely responsible for determining the
appropriateness of using or redistributing the Work and assume any
risks associated with Your exercise of permissions under this License.
8. Limitation of Liability. In no event and under no legal theory,
whether in tort (including negligence), contract, or otherwise,
unless required by applicable law (such as deliberate and grossly
negligent acts) or agreed to in writing, shall any Contributor be
liable to You for damages, including any direct, indirect, special,
incidental, or consequential damages of any character arising as a
result of this License or out of the use or inability to use the
Work (including but not limited to damages for loss of goodwill,
work stoppage, computer failure or malfunction, or any and all
other commercial damages or losses), even if such Contributor
has been advised of the possibility of such damages.
9. Accepting Warranty or Additional Liability. While redistributing
the Work or Derivative Works thereof, You may choose to offer,
and charge a fee for, acceptance of support, warranty, indemnity,
or other liability obligations and/or rights consistent with this
License. However, in accepting such obligations, You may act only
on Your own behalf and on Your sole responsibility, not on behalf
of any other Contributor, and only if You agree to indemnify,
defend, and hold each Contributor harmless for any liability
incurred by, or claims asserted against, such Contributor by reason
of your accepting any such warranty or additional liability.
END OF TERMS AND CONDITIONS
APPENDIX: How to apply the Apache License to your work.
To apply the Apache License to your work, attach the following
boilerplate notice, with the fields enclosed by brackets "[]"
replaced with your own identifying information. (Don't include
the brackets!) The text should be enclosed in the appropriate
comment syntax for the file format. We also recommend that a
file or class name and description of purpose be included on the
same "printed page" as the copyright notice for easier
identification within third-party archives.
Copyright (c) 2025 Baidu, Inc. 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.
PixArt.py 0 → 100644
View file @ a44e71bd
from abc import abstractmethod
from functools import partial
import math
from typing import Iterable
import numpy as np
import torch as th
import torch.nn as nn
import torch.nn.functional as F
from timm.models.vision_transformer import Attention, Mlp
from positional_encodings.torch_encodings import PositionalEncoding1D
from timm.models.layers import DropPath
from .utils import auto_grad_checkpoint, to_2tuple
from .PixArt_blocks import (
t2i_modulate,
WindowAttention,
MultiHeadCrossAttention,
T2IFinalLayer,
TimestepEmbedder,
FinalLayer,
)
import math
class PatchEmbed(nn.Module):
"""2D Image to Patch Embedding"""
def __init__(
self,
img_size=(256, 16),
patch_size=(16, 4),
overlap=(0, 0),
in_chans=128,
embed_dim=768,
norm_layer=None,
flatten=True,
bias=True,
):
super().__init__()
self.img_size = img_size
self.patch_size = patch_size
self.ol = overlap
self.grid_size = (
math.ceil((img_size[0] - patch_size[0]) / (patch_size[0] - overlap[0])) + 1,
math.ceil((img_size[1] - patch_size[1]) / (patch_size[1] - overlap[1])) + 1,
)
self.pad_size = (
(self.grid_size[0] - 1) * (self.patch_size[0] - overlap[0])
+ self.patch_size[0]
- self.img_size[0],
+(self.grid_size[1] - 1) * (self.patch_size[1] - overlap[1])
+ self.patch_size[1]
- self.img_size[1],
)
self.pad_size = (self.pad_size[0] // 2, self.pad_size[1] // 2)
self.num_patches = self.grid_size[0] * self.grid_size[1]
self.flatten = flatten
self.proj = nn.Conv2d(
in_chans,
embed_dim,
kernel_size=patch_size,
stride=(patch_size[0] - overlap[0], patch_size[1] - overlap[1]),
bias=bias,
)
self.norm = norm_layer(embed_dim) if norm_layer else nn.Identity()
def forward(self, x):
x = F.pad(
x,
(
self.pad_size[-1],
self.pad_size[-1],
self.pad_size[-2],
self.pad_size[-2],
),
"constant",
0,
)
x = self.proj(x)
if self.flatten:
x = x.flatten(2).transpose(1, 2) # BCHW -> BNC
x = self.norm(x)
return x
class PatchEmbed_1D(nn.Module):
def __init__(
self,
img_size=(256, 16),
in_chans=8,
embed_dim=1152,
norm_layer=None,
bias=True,
):
super().__init__()
self.proj = nn.Linear(in_chans * img_size[1], embed_dim, bias=bias)
self.norm = norm_layer(embed_dim) if norm_layer else nn.Identity()
def forward(self, x):
x = th.einsum("bctf->btfc", x)
x = x.flatten(2) # BTFC -> BTD
x = self.proj(x)
x = self.norm(x)
return x
def modulate(x, shift, scale):
return x * (1 + scale.unsqueeze(1)) + shift.unsqueeze(1)
def t2i_modulate(x, shift, scale):
return x * (1 + scale) + shift
class PixArtBlock(nn.Module):
"""
A PixArt block with adaptive layer norm (adaLN-single) conditioning.
"""
def __init__(
self,
hidden_size,
num_heads,
mlp_ratio=4.0,
drop_path=0.0,
window_size=0,
input_size=None,
use_rel_pos=False,
**block_kwargs
):
super().__init__()
self.hidden_size = hidden_size
self.norm1 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
self.attn = WindowAttention(
hidden_size,
num_heads=num_heads,
qkv_bias=True,
input_size=input_size if window_size == 0 else (window_size, window_size),
use_rel_pos=use_rel_pos,
**block_kwargs
)
self.cross_attn = MultiHeadCrossAttention(
hidden_size, num_heads, **block_kwargs
)
self.norm2 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
# to be compatible with lower version pytorch
approx_gelu = lambda: nn.GELU(approximate="tanh")
self.mlp = Mlp(
in_features=hidden_size,
hidden_features=int(hidden_size * mlp_ratio),
act_layer=approx_gelu,
drop=0,
)
self.drop_path = DropPath(drop_path) if drop_path > 0.0 else nn.Identity()
self.window_size = window_size
self.scale_shift_table = nn.Parameter(
th.randn(6, hidden_size) / hidden_size**0.5
)
def forward(self, x, y, t, mask=None, **kwargs):
B, N, C = x.shape
shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = (
self.scale_shift_table[None] + t.reshape(B, 6, -1)
).chunk(6, dim=1)
x = x + self.drop_path(
gate_msa
* self.attn(t2i_modulate(self.norm1(x), shift_msa, scale_msa)).reshape(
B, N, C
)
)
x = x + self.cross_attn(x, y, mask)
x = x + self.drop_path(
gate_mlp * self.mlp(t2i_modulate(self.norm2(x), shift_mlp, scale_mlp))
)
return x
from ldm.modules.diffusionmodules.attention import CrossAttention_1D
class PixArtBlock_Slow(nn.Module):
"""
A PixArt block with adaptive layer norm (adaLN-single) conditioning.
"""
def __init__(
self,
hidden_size,
num_heads,
mlp_ratio=4.0,
drop_path=0.0,
window_size=0,
input_size=None,
use_rel_pos=False,
**block_kwargs
):
super().__init__()
self.hidden_size = hidden_size
self.norm1 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
self.attn = CrossAttention_1D(
query_dim=hidden_size,
context_dim=hidden_size,
heads=num_heads,
dim_head=int(hidden_size / num_heads),
)
self.cross_attn = CrossAttention_1D(
query_dim=hidden_size,
context_dim=hidden_size,
heads=num_heads,
dim_head=int(hidden_size / num_heads),
)
self.norm2 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
# to be compatible with lower version pytorch
approx_gelu = lambda: nn.GELU(approximate="tanh")
self.mlp = Mlp(
in_features=hidden_size,
hidden_features=int(hidden_size * mlp_ratio),
act_layer=approx_gelu,
drop=0,
)
self.drop_path = DropPath(drop_path) if drop_path > 0.0 else nn.Identity()
self.window_size = window_size
self.scale_shift_table = nn.Parameter(
th.randn(6, hidden_size) / hidden_size**0.5
)
def forward(self, x, y, t, mask=None, **kwargs):
B, N, C = x.shape
shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = (
self.scale_shift_table[None] + t.reshape(B, 6, -1)
).chunk(6, dim=1)
x = x + self.drop_path(
gate_msa
* self.attn(t2i_modulate(self.norm1(x), shift_msa, scale_msa)).reshape(
B, N, C
)
)
x = x + self.cross_attn(x, y, mask)
x = x + self.drop_path(
gate_mlp * self.mlp(t2i_modulate(self.norm2(x), shift_mlp, scale_mlp))
)
return x
class PixArt(nn.Module):
"""
Diffusion model with a Transformer backbone.
"""
def __init__(
self,
input_size=(256, 16),
patch_size=(16, 4),
overlap=(0, 0),
in_channels=8,
hidden_size=1152,
depth=28,
num_heads=16,
mlp_ratio=4.0,
class_dropout_prob=0.1,
pred_sigma=True,
drop_path: float = 0.0,
window_size=0,
window_block_indexes=None,
use_rel_pos=False,
cond_dim=1024,
lewei_scale=1.0,
use_cfg=True,
cfg_scale=4.0,
config=None,
model_max_length=120,
**kwargs
):
if window_block_indexes is None:
window_block_indexes = []
super().__init__()
self.use_cfg = use_cfg
self.cfg_scale = cfg_scale
self.input_size = input_size
self.pred_sigma = pred_sigma
self.in_channels = in_channels
self.out_channels = in_channels * 2 if pred_sigma else in_channels
self.patch_size = patch_size
self.num_heads = num_heads
self.lewei_scale = (lewei_scale,)
self.x_embedder = PatchEmbed(
input_size, patch_size, overlap, in_channels, hidden_size, bias=True
)
self.t_embedder = TimestepEmbedder(hidden_size)
num_patches = self.x_embedder.num_patches
self.base_size = input_size[0] // self.patch_size[0] * 2
# Will use fixed sin-cos embedding:
self.register_buffer("pos_embed", th.zeros(1, num_patches, hidden_size))
approx_gelu = lambda: nn.GELU(approximate="tanh")
self.t_block = nn.Sequential(
nn.SiLU(), nn.Linear(hidden_size, 6 * hidden_size, bias=True)
)
self.y_embedder = nn.Linear(cond_dim, hidden_size)
drop_path = [
x.item() for x in th.linspace(0, drop_path, depth)
] # stochastic depth decay rule
self.blocks = nn.ModuleList(
[
PixArtBlock(
hidden_size,
num_heads,
mlp_ratio=mlp_ratio,
drop_path=drop_path[i],
input_size=(
self.x_embedder.grid_size[0],
self.x_embedder.grid_size[1],
),
window_size=0,
use_rel_pos=False,
)
for i in range(depth)
]
)
self.final_layer = T2IFinalLayer(hidden_size, patch_size, self.out_channels)
self.initialize_weights()
def forward(self, x, timestep, context_list, context_mask_list=None, **kwargs):
"""
Forward pass of PixArt.
x: (N, C, H, W) tensor of spatial inputs (images or latent representations of images)
t: (N,) tensor of diffusion timesteps
y: (N, 1, 120, C) tensor of class labels
"""
x = x.to(self.dtype)
timestep = timestep.to(self.dtype)
y = context_list[0].to(self.dtype)
pos_embed = self.pos_embed.to(self.dtype)
self.h, self.w = self.x_embedder.grid_size[0], self.x_embedder.grid_size[1]
x = self.x_embedder(x) + pos_embed
t = self.t_embedder(timestep.to(x.dtype))
t0 = self.t_block(t)
y = self.y_embedder(y)
mask = context_mask_list[0]
assert mask is not None
# if mask is not None:
y = y.masked_select(mask.unsqueeze(-1) != 0).view(1, -1, x.shape[-1])
y_lens = mask.sum(dim=1).tolist()
y_lens = [int(_) for _ in y_lens]
for block in self.blocks:
x = auto_grad_checkpoint(block, x, y, t0, y_lens)
x = self.final_layer(x, t)
x = self.unpatchify(x)
return x
def forward_with_dpmsolver(self, x, timestep, y, mask=None, **kwargs):
"""
dpm solver donnot need variance prediction
"""
# https://github.com/openai/glide-text2im/blob/main/notebooks/text2im.ipynb
model_out = self.forward(x, timestep, y, mask)
return model_out.chunk(2, dim=1)[0]
def forward_with_cfg(self, x, timestep, y, cfg_scale, mask=None, **kwargs):
"""
Forward pass of PixArt, but also batches the unconditional forward pass for classifier-free guidance.
"""
# https://github.com/openai/glide-text2im/blob/main/notebooks/text2im.ipynb
half = x[: len(x) // 2]
combined = th.cat([half, half], dim=0)
model_out = self.forward(combined, timestep, y, mask)
model_out = model_out["x"] if isinstance(model_out, dict) else model_out
eps, rest = model_out[:, :8], model_out[:, 8:]
cond_eps, uncond_eps = th.split(eps, len(eps) // 2, dim=0)
half_eps = uncond_eps + cfg_scale * (cond_eps - uncond_eps)
eps = th.cat([half_eps, half_eps], dim=0)
return eps
def unpatchify(self, x):
"""
x: (N, T, patch_size 0 * patch_size 1 * C)
imgs: (Bs. 256. 16. 8)
"""
c = self.out_channels
p0 = self.x_embedder.patch_size[0]
p1 = self.x_embedder.patch_size[1]
h, w = self.x_embedder.grid_size[0], self.x_embedder.grid_size[1]
x = x.reshape(shape=(x.shape[0], h, w, p0, p1, c))
x = th.einsum("nhwpqc->nchpwq", x)
imgs = x.reshape(shape=(x.shape[0], c, h * p0, w * p1))
return imgs
def initialize_weights(self):
# Initialize transformer layers:
def _basic_init(module):
if isinstance(module, nn.Linear):
th.nn.init.xavier_uniform_(module.weight)
if module.bias is not None:
nn.init.constant_(module.bias, 0)
self.apply(_basic_init)
# Initialize (and freeze) pos_embed by sin-cos embedding:
pos_embed = get_2d_sincos_pos_embed(
self.pos_embed.shape[-1],
self.x_embedder.grid_size,
lewei_scale=self.lewei_scale,
base_size=self.base_size,
)
self.pos_embed.data.copy_(th.from_numpy(pos_embed).float().unsqueeze(0))
# Initialize patch_embed like nn.Linear (instead of nn.Conv2d):
w = self.x_embedder.proj.weight.data
nn.init.xavier_uniform_(w.view([w.shape[0], -1]))
# Initialize timestep embedding MLP:
nn.init.normal_(self.t_embedder.mlp[0].weight, std=0.02)
nn.init.normal_(self.t_embedder.mlp[2].weight, std=0.02)
nn.init.normal_(self.t_block[1].weight, std=0.02)
# Initialize caption embedding MLP:
nn.init.normal_(self.y_embedder.weight, std=0.02)
# Zero-out adaLN modulation layers in PixArt blocks:
for block in self.blocks:
nn.init.constant_(block.cross_attn.proj.weight, 0)
nn.init.constant_(block.cross_attn.proj.bias, 0)
# Zero-out output layers:
nn.init.constant_(self.final_layer.linear.weight, 0)
nn.init.constant_(self.final_layer.linear.bias, 0)
@property
def dtype(self):
return next(self.parameters()).dtype
class SwiGLU(nn.Module):
def __init__(
self,
dim: int,
hidden_dim: int,
multiple_of: int,
):
super().__init__()
hidden_dim = multiple_of * ((hidden_dim + multiple_of - 1) // multiple_of)
self.w1 = nn.Linear(dim, hidden_dim, bias=False)
self.w2 = nn.Linear(hidden_dim, dim, bias=False)
self.w3 = nn.Linear(dim, hidden_dim, bias=False)
def forward(self, x):
return self.w2(F.silu(self.w1(x)) * self.w3(x))
class MDTBlock(nn.Module):
"""
A PixArt block with adaptive layer norm (adaLN-single) conditioning.
"""
def __init__(
self,
hidden_size,
num_heads,
mlp_ratio=4.0,
FFN_type="SwiGLU",
drop_path=0.0,
window_size=0,
input_size=None,
use_rel_pos=False,
skip=False,
**block_kwargs
):
super().__init__()
self.hidden_size = hidden_size
self.norm1 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
self.attn = WindowAttention(
hidden_size,
num_heads=num_heads,
qkv_bias=True,
input_size=input_size if window_size == 0 else (window_size, window_size),
use_rel_pos=use_rel_pos,
**block_kwargs
)
self.cross_attn = MultiHeadCrossAttention(
hidden_size, num_heads, **block_kwargs
)
self.norm2 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
# to be compatible with lower version pytorch
approx_gelu = lambda: nn.GELU(approximate="tanh")
if FFN_type == "mlp":
self.mlp = Mlp(
in_features=hidden_size,
hidden_features=int(hidden_size * mlp_ratio),
act_layer=approx_gelu,
drop=0,
)
elif FFN_type == "SwiGLU":
self.mlp = SwiGLU(hidden_size, int(hidden_size * mlp_ratio), 1)
self.drop_path = DropPath(drop_path) if drop_path > 0.0 else nn.Identity()
self.window_size = window_size
self.scale_shift_table = nn.Parameter(
th.randn(6, hidden_size) / hidden_size**0.5
)
self.skip_linear = nn.Linear(2 * hidden_size, hidden_size) if skip else None
def forward(self, x, y, t, mask=None, skip=None, ids_keep=None, **kwargs):
B, N, C = x.shape
if self.skip_linear is not None:
x = self.skip_linear(th.cat([x, skip], dim=-1))
shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = (
self.scale_shift_table[None] + t.reshape(B, 6, -1)
).chunk(6, dim=1)
x = x + self.drop_path(
gate_msa
* self.attn(t2i_modulate(self.norm1(x), shift_msa, scale_msa)).reshape(
B, N, C
)
)
x = x + self.cross_attn(x, y, mask)
x = x + self.drop_path(
gate_mlp * self.mlp(t2i_modulate(self.norm2(x), shift_mlp, scale_mlp))
)
return x
class DEBlock(nn.Module):
"""
Decoder block with added SpecTNT transformer
"""
def __init__(
self,
hidden_size,
num_heads,
mlp_ratio=4.0,
FFN_type="SwiGLU",
drop_path=0.0,
window_size=0,
input_size=None,
use_rel_pos=False,
skip=False,
num_f=None,
num_t=None,
**block_kwargs
):
super().__init__()
self.hidden_size = hidden_size
self.norm1 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
self.attn = WindowAttention(
hidden_size,
num_heads=num_heads,
qkv_bias=True,
input_size=input_size if window_size == 0 else (window_size, window_size),
use_rel_pos=use_rel_pos,
**block_kwargs
)
self.cross_attn = MultiHeadCrossAttention(
hidden_size, num_heads, **block_kwargs
)
self.norm2 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
self.norm3 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
self.norm4 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
self.norm5 = nn.LayerNorm(
hidden_size * num_f, elementwise_affine=False, eps=1e-6
)
self.norm6 = nn.LayerNorm(
hidden_size * num_f, elementwise_affine=False, eps=1e-6
)
# to be compatible with lower version pytorch
approx_gelu = lambda: nn.GELU(approximate="tanh")
if FFN_type == "mlp":
self.mlp = Mlp(
in_features=hidden_size,
hidden_features=int(hidden_size * mlp_ratio),
act_layer=approx_gelu,
drop=0,
)
elif FFN_type == "SwiGLU":
self.mlp = SwiGLU(hidden_size, int(hidden_size * mlp_ratio), 1)
self.drop_path = DropPath(drop_path) if drop_path > 0.0 else nn.Identity()
self.window_size = window_size
self.scale_shift_table = nn.Parameter(
th.randn(6, hidden_size) / hidden_size**0.5
)
self.skip_linear = nn.Linear(2 * hidden_size, hidden_size) if skip else None
self.F_transformer = WindowAttention(
hidden_size,
num_heads=4,
qkv_bias=True,
input_size=input_size if window_size == 0 else (window_size, window_size),
use_rel_pos=use_rel_pos,
**block_kwargs
)
self.T_transformer = WindowAttention(
hidden_size * num_f,
num_heads=16,
qkv_bias=True,
input_size=input_size if window_size == 0 else (window_size, window_size),
use_rel_pos=use_rel_pos,
**block_kwargs
)
self.f_pos = nn.Embedding(num_f, hidden_size)
self.t_pos = nn.Embedding(num_t, hidden_size * num_f)
self.num_f = num_f
self.num_t = num_t
def forward(self, x, end, y, t, mask=None, skip=None, ids_keep=None, **kwargs):
B, D, C = x.shape
T = self.num_t
F_add_1 = self.num_f
x_normal = x
if self.skip_linear is not None:
x_normal = self.skip_linear(th.cat([x_normal, skip], dim=-1))
D = T * (F_add_1 - 1)
shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = (
self.scale_shift_table[None] + t.reshape(B, 6, -1)
).chunk(6, dim=1)
x_normal = x_normal + self.drop_path(
gate_msa
* self.attn(
t2i_modulate(self.norm1(x_normal), shift_msa, scale_msa)
).reshape(B, D, C)
)
x_normal = x_normal.reshape(B, T, F_add_1 - 1, C)
x_normal = th.cat((x_normal, end), 2)
x_normal = x_normal.reshape(B * T, F_add_1, C)
pos_f = th.arange(self.num_f, device=x.device).unsqueeze(0).expand(B * T, -1)
x_normal = x_normal + self.f_pos(pos_f)
x_normal = x_normal + self.F_transformer(self.norm3(x_normal))
x_normal = x_normal.reshape(B, T, F_add_1 * C)
pos_t = th.arange(self.num_t, device=x.device).unsqueeze(0).expand(B, -1)
x_normal = x_normal + self.t_pos(pos_t)
x_normal = x_normal + self.T_transformer(self.norm5(x_normal))
x_normal = x_normal.reshape(B, T, F_add_1, C)
end = x_normal[:, :, -1, :].unsqueeze(2)
x_normal = x_normal[:, :, :-1, :]
x_normal = x_normal.reshape(B, T * (F_add_1 - 1), C)
x_normal = x_normal + self.cross_attn(x_normal, y, mask)
x_normal = x_normal + self.drop_path(
gate_mlp
* self.mlp(t2i_modulate(self.norm2(x_normal), shift_mlp, scale_mlp))
)
return x_normal, end
class PixArt_MDT(nn.Module):
"""
Diffusion model with a Transformer backbone.
"""
def __init__(
self,
input_size=(256, 16),
patch_size=(16, 4),
overlap=(0, 0),
in_channels=8,
hidden_size=1152,
depth=28,
num_heads=16,
mlp_ratio=4.0,
class_dropout_prob=0.1,
pred_sigma=False,
drop_path: float = 0.0,
window_size=0,
window_block_indexes=None,
use_rel_pos=False,
cond_dim=1024,
lewei_scale=1.0,
use_cfg=True,
cfg_scale=4.0,
config=None,
model_max_length=120,
mask_ratio=None,
decode_layer=4,
**kwargs
):
if window_block_indexes is None:
window_block_indexes = []
super().__init__()
self.use_cfg = use_cfg
self.cfg_scale = cfg_scale
self.input_size = input_size
self.pred_sigma = pred_sigma
self.in_channels = in_channels
self.out_channels = in_channels
self.patch_size = patch_size
self.num_heads = num_heads
self.lewei_scale = (lewei_scale,)
decode_layer = int(decode_layer)
self.x_embedder = PatchEmbed(
input_size, patch_size, overlap, in_channels, hidden_size, bias=True
)
self.t_embedder = TimestepEmbedder(hidden_size)
num_patches = self.x_embedder.num_patches
self.base_size = input_size[0] // self.patch_size[0] * 2
# Will use fixed sin-cos embedding:
self.register_buffer("pos_embed", th.zeros(1, num_patches, hidden_size))
approx_gelu = lambda: nn.GELU(approximate="tanh")
self.t_block = nn.Sequential(
nn.SiLU(), nn.Linear(hidden_size, 6 * hidden_size, bias=True)
)
self.y_embedder = nn.Linear(cond_dim, hidden_size)
half_depth = (depth - decode_layer) // 2
self.half_depth = half_depth
drop_path_half = [
x.item() for x in th.linspace(0, drop_path, half_depth)
] # stochastic depth decay rule
drop_path_decode = [x.item() for x in th.linspace(0, drop_path, decode_layer)]
self.en_inblocks = nn.ModuleList(
[
MDTBlock(
hidden_size,
num_heads,
mlp_ratio=mlp_ratio,
drop_path=drop_path_half[i],
input_size=(
self.x_embedder.grid_size[0],
self.x_embedder.grid_size[1],
),
window_size=0,
use_rel_pos=False,
FFN_type="mlp",
)
for i in range(half_depth)
]
)
self.en_outblocks = nn.ModuleList(
[
MDTBlock(
hidden_size,
num_heads,
mlp_ratio=mlp_ratio,
drop_path=drop_path_half[i],
input_size=(
self.x_embedder.grid_size[0],
self.x_embedder.grid_size[1],
),
window_size=0,
use_rel_pos=False,
skip=True,
FFN_type="mlp",
)
for i in range(half_depth)
]
)
self.de_blocks = nn.ModuleList(
[
MDTBlock(
hidden_size,
num_heads,
mlp_ratio=mlp_ratio,
drop_path=drop_path_decode[i],
input_size=(
self.x_embedder.grid_size[0],
self.x_embedder.grid_size[1],
),
window_size=0,
use_rel_pos=False,
skip=True,
FFN_type="mlp",
)
for i in range(decode_layer)
]
)
self.sideblocks = nn.ModuleList(
[
MDTBlock(
hidden_size,
num_heads,
mlp_ratio=mlp_ratio,
input_size=(
self.x_embedder.grid_size[0],
self.x_embedder.grid_size[1],
),
window_size=0,
use_rel_pos=False,
FFN_type="mlp",
)
for _ in range(1)
]
)
self.final_layer = T2IFinalLayer(hidden_size, patch_size, self.out_channels)
self.decoder_pos_embed = nn.Parameter(
th.zeros(1, num_patches, hidden_size), requires_grad=True
)
if mask_ratio is not None:
self.mask_token = nn.Parameter(th.zeros(1, 1, hidden_size))
self.mask_ratio = float(mask_ratio)
self.decode_layer = int(decode_layer)
else:
self.mask_token = nn.Parameter(
th.zeros(1, 1, hidden_size), requires_grad=False
)
self.mask_ratio = None
self.decode_layer = int(decode_layer)
self.initialize_weights()
def forward(self, x, t, context, mask=None, enable_mask=False, **kwargs):
"""
Forward pass of PixArt.
x: (N, C, H, W) tensor of spatial inputs (images or latent representations of images)
t: (N,) tensor of diffusion timesteps
y: (N, 1, 120, C) tensor of class labels
"""
x = x.to(self.dtype)
t = t.to(self.dtype)
y = context.to(self.dtype)
pos_embed = self.pos_embed.to(self.dtype)
self.h, self.w = self.x_embedder.grid_size[0], self.x_embedder.grid_size[1]
x = self.x_embedder(x) + pos_embed
t = self.t_embedder(t.to(x.dtype))
t0 = self.t_block(t)
y = self.y_embedder(y)
try:
mask = mask
except:
mask = th.ones(x.shape[0], 1).to(x.device)
print("MASK !!!!!!!!!!!!!!!!!!!!!!!!!")
assert mask is not None
y = y.masked_select(mask.unsqueeze(-1) != 0).view(1, -1, x.shape[-1])
y_lens = mask.sum(dim=1).tolist()
y_lens = [int(_) for _ in y_lens]
input_skip = x
masked_stage = False
skips = []
# TODO : masking op for training
if self.mask_ratio is not None and self.training:
rand_mask_ratio = th.rand(1, device=x.device)
rand_mask_ratio = rand_mask_ratio * 0.2 + self.mask_ratio
x, mask, ids_restore, ids_keep = self.random_masking(x, rand_mask_ratio)
masked_stage = True
for block in self.en_inblocks:
if masked_stage:
x = auto_grad_checkpoint(block, x, y, t0, y_lens, ids_keep=ids_keep)
else:
x = auto_grad_checkpoint(block, x, y, t0, y_lens, ids_keep=None)
skips.append(x)
for block in self.en_outblocks:
if masked_stage:
x = auto_grad_checkpoint(
block, x, y, t0, y_lens, skip=skips.pop(), ids_keep=ids_keep
)
else:
x = auto_grad_checkpoint(
block, x, y, t0, y_lens, skip=skips.pop(), ids_keep=None
)
if self.mask_ratio is not None and self.training:
x = self.forward_side_interpolater(x, y, t0, y_lens, mask, ids_restore)
masked_stage = False
else:
# add pos embed
x = x + self.decoder_pos_embed
for i in range(len(self.de_blocks)):
block = self.de_blocks[i]
this_skip = input_skip
x = auto_grad_checkpoint(
block, x, y, t0, y_lens, skip=this_skip, ids_keep=None
)
x = self.final_layer(x, t)
x = self.unpatchify(x)
return x
def forward_with_dpmsolver(self, x, timestep, y, mask=None, **kwargs):
"""
dpm solver donnot need variance prediction
"""
# https://github.com/openai/glide-text2im/blob/main/notebooks/text2im.ipynb
model_out = self.forward(x, timestep, y, mask)
return model_out.chunk(2, dim=1)[0]
def forward_with_cfg(
self, x, timestep, context_list, context_mask_list=None, cfg_scale=4.0, **kwargs
):
"""
Forward pass of PixArt, but also batches the unconditional forward pass for classifier-free guidance.
"""
# https://github.com/openai/glide-text2im/blob/main/notebooks/text2im.ipynb
half = x[: len(x) // 2]
combined = th.cat([half, half], dim=0)
model_out = self.forward(
combined, timestep, context_list, context_mask_list=None
)
model_out = model_out["x"] if isinstance(model_out, dict) else model_out
eps, rest = model_out[:, :8], model_out[:, 8:]
cond_eps, uncond_eps = th.split(eps, len(eps) // 2, dim=0)
half_eps = uncond_eps + cfg_scale * (cond_eps - uncond_eps)
eps = th.cat([half_eps, half_eps], dim=0)
return eps
def unpatchify(self, x):
"""
x: (N, T, patch_size 0 * patch_size 1 * C)
imgs: (Bs. 256. 16. 8)
"""
if self.x_embedder.ol == (0, 0) or self.x_embedder.ol == [0, 0]:
c = self.out_channels
p0 = self.x_embedder.patch_size[0]
p1 = self.x_embedder.patch_size[1]
h, w = self.x_embedder.grid_size[0], self.x_embedder.grid_size[1]
x = x.reshape(shape=(x.shape[0], h, w, p0, p1, c))
x = th.einsum("nhwpqc->nchpwq", x)
imgs = x.reshape(shape=(x.shape[0], c, h * p0, w * p1))
return imgs
lf = self.x_embedder.grid_size[0]
rf = self.x_embedder.grid_size[1]
lp = self.x_embedder.patch_size[0]
rp = self.x_embedder.patch_size[1]
lo = self.x_embedder.ol[0]
ro = self.x_embedder.ol[1]
lm = self.x_embedder.img_size[0]
rm = self.x_embedder.img_size[1]
lpad = self.x_embedder.pad_size[0]
rpad = self.x_embedder.pad_size[1]
bs = x.shape[0]
torch_map = self.torch_map
c = self.out_channels
x = x.reshape(shape=(bs, lf, rf, lp, rp, c))
x = th.einsum("nhwpqc->nchwpq", x)
added_map = th.zeros(bs, c, lm + 2 * lpad, rm + 2 * rpad).to(x.device)
for i in range(lf):
for j in range(rf):
xx = (i) * (lp - lo)
yy = (j) * (rp - ro)
added_map[:, :, xx : (xx + lp), yy : (yy + rp)] += x[:, :, i, j, :, :]
added_map = added_map[:][:][lpad : lm + lpad, rpad : rm + rpad]
return th.mul(added_map.to(x.device), torch_map.to(x.device))
def random_masking(self, x, mask_ratio):
"""
Perform per-sample random masking by per-sample shuffling.
Per-sample shuffling is done by argsort random noise.
x: [N, L, D], sequence
"""
N, L, D = x.shape # batch, length, dim
len_keep = int(L * (1 - mask_ratio))
noise = th.rand(N, L, device=x.device)
# sort noise for each sample
# ascend: small is keep, large is remove
ids_shuffle = th.argsort(noise, dim=1)
ids_restore = th.argsort(ids_shuffle, dim=1)
# keep the first subset
ids_keep = ids_shuffle[:, :len_keep]
x_masked = th.gather(x, dim=1, index=ids_keep.unsqueeze(-1).repeat(1, 1, D))
# generate the binary mask: 0 is keep, 1 is remove
mask = th.ones([N, L], device=x.device)
mask[:, :len_keep] = 0
# unshuffle to get the binary mask
mask = th.gather(mask, dim=1, index=ids_restore)
return x_masked, mask, ids_restore, ids_keep
def forward_side_interpolater(self, x, y, t0, y_lens, mask, ids_restore):
# append mask tokens to sequence
mask_tokens = self.mask_token.repeat(
x.shape[0], ids_restore.shape[1] - x.shape[1], 1
)
x_ = th.cat([x, mask_tokens], dim=1)
x = th.gather(
x_, dim=1, index=ids_restore.unsqueeze(-1).repeat(1, 1, x.shape[2])
) # unshuffle
# add pos embed
x = x + self.decoder_pos_embed
# pass to the basic block
x_before = x
for sideblock in self.sideblocks:
x = sideblock(x, y, t0, y_lens, ids_keep=None)
# masked shortcut
mask = mask.unsqueeze(dim=-1)
x = x * mask + (1 - mask) * x_before
return x
def initialize_weights(self):
# Initialize transformer layers:
def _basic_init(module):
if isinstance(module, nn.Linear):
th.nn.init.xavier_uniform_(module.weight)
if module.bias is not None:
nn.init.constant_(module.bias, 0)
self.apply(_basic_init)
# Initialize (and freeze) pos_embed by sin-cos embedding:
pos_embed = get_2d_sincos_pos_embed(
self.pos_embed.shape[-1],
self.x_embedder.grid_size,
lewei_scale=self.lewei_scale,
base_size=self.base_size,
)
self.pos_embed.data.copy_(th.from_numpy(pos_embed).float().unsqueeze(0))
# Initialize patch_embed like nn.Linear (instead of nn.Conv2d):
w = self.x_embedder.proj.weight.data
nn.init.xavier_uniform_(w.view([w.shape[0], -1]))
# Initialize timestep embedding MLP:
nn.init.normal_(self.t_embedder.mlp[0].weight, std=0.02)
nn.init.normal_(self.t_embedder.mlp[2].weight, std=0.02)
nn.init.normal_(self.t_block[1].weight, std=0.02)
# Initialize caption embedding MLP:
nn.init.normal_(self.y_embedder.weight, std=0.02)
# Zero-out adaLN modulation layers in PixArt blocks:
for block in self.en_inblocks:
nn.init.constant_(block.cross_attn.proj.weight, 0)
nn.init.constant_(block.cross_attn.proj.bias, 0)
for block in self.en_outblocks:
nn.init.constant_(block.cross_attn.proj.weight, 0)
nn.init.constant_(block.cross_attn.proj.bias, 0)
for block in self.de_blocks:
nn.init.constant_(block.cross_attn.proj.weight, 0)
nn.init.constant_(block.cross_attn.proj.bias, 0)
for block in self.sideblocks:
nn.init.constant_(block.cross_attn.proj.weight, 0)
nn.init.constant_(block.cross_attn.proj.bias, 0)
# Zero-out output layers:
nn.init.constant_(self.final_layer.linear.weight, 0)
nn.init.constant_(self.final_layer.linear.bias, 0)
if self.x_embedder.ol == [0, 0] or self.x_embedder.ol == (0, 0):
return
lf = self.x_embedder.grid_size[0]
rf = self.x_embedder.grid_size[1]
lp = self.x_embedder.patch_size[0]
rp = self.x_embedder.patch_size[1]
lo = self.x_embedder.ol[0]
ro = self.x_embedder.ol[1]
lm = self.x_embedder.img_size[0]
rm = self.x_embedder.img_size[1]
lpad = self.x_embedder.pad_size[0]
rpad = self.x_embedder.pad_size[1]
torch_map = th.zeros(lm + 2 * lpad, rm + 2 * rpad).to("cuda")
for i in range(lf):
for j in range(rf):
xx = (i) * (lp - lo)
yy = (j) * (rp - ro)
torch_map[xx : (xx + lp), yy : (yy + rp)] += 1
torch_map = torch_map[lpad : lm + lpad, rpad : rm + rpad]
self.torch_map = th.reciprocal(torch_map)
@property
def dtype(self):
return next(self.parameters()).dtype
def get_2d_sincos_pos_embed(
embed_dim,
grid_size,
cls_token=False,
extra_tokens=0,
lewei_scale=1.0,
base_size_x=256 // 4,
base_size_y=16 // 4,
base_size=128,
):
"""
grid_size: int of the grid height and width
return:
pos_embed: [grid_size*grid_size, embed_dim] or [1+grid_size*grid_size, embed_dim] (w/ or w/o cls_token)
"""
if isinstance(grid_size, int):
grid_size = to_2tuple(grid_size)
grid_h = (
np.arange(grid_size[0], dtype=np.float32)
/ (grid_size[0] / base_size_x)
/ lewei_scale
)
grid_w = (
np.arange(grid_size[1], dtype=np.float32)
/ (grid_size[1] / base_size_y)
/ lewei_scale
)
grid = np.meshgrid(grid_w, grid_h) # here w goes first
grid = np.stack(grid, axis=0)
grid = grid.reshape([2, 1, grid_size[1], grid_size[0]])
pos_embed = get_2d_sincos_pos_embed_from_grid(embed_dim, grid)
if cls_token and extra_tokens > 0:
pos_embed = np.concatenate(
[np.zeros([extra_tokens, embed_dim]), pos_embed], axis=0
)
return pos_embed
def get_2d_sincos_pos_embed_from_grid(embed_dim, grid):
assert embed_dim % 2 == 0
# use half of dimensions to encode grid_h
emb_h = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[0])
emb_w = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[1])
return np.concatenate([emb_h, emb_w], axis=1)
def get_1d_sincos_pos_embed_from_grid(embed_dim, pos):
"""
embed_dim: output dimension for each position
pos: a list of positions to be encoded: size (M,)
out: (M, D)
"""
assert embed_dim % 2 == 0
omega = np.arange(embed_dim // 2, dtype=np.float64)
omega /= embed_dim / 2.0
omega = 1.0 / 10000**omega
pos = pos.reshape(-1)
out = np.einsum("m,d->md", pos, omega)
emb_sin = np.sin(out)
emb_cos = np.cos(out)
return np.concatenate([emb_sin, emb_cos], axis=1)
PixArt_blocks.py 0 → 100644
View file @ a44e71bd
# Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.
# This source code is licensed under the license found in the
# LICENSE file in the root directory of this source tree.
# --------------------------------------------------------
# References:
# GLIDE: https://github.com/openai/glide-text2im
# MAE: https://github.com/facebookresearch/mae/blob/main/models_mae.py
# --------------------------------------------------------
import math
import torch
import torch.nn as nn
from timm.models.vision_transformer import Mlp, Attention as Attention_
from einops import rearrange, repeat
import torch.nn.functional as F
from .utils import add_decomposed_rel_pos
def modulate(x, shift, scale):
return x * (1 + scale.unsqueeze(1)) + shift.unsqueeze(1)
def t2i_modulate(x, shift, scale):
return x * (1 + scale) + shift
class MultiHeadCrossAttention_org(nn.Module):
def __init__(self, n_feat, n_head, attn_drop=0.0, proj_drop=0):
"""Construct an MultiHeadedAttention object."""
super(MultiHeadCrossAttention, self).__init__()
assert n_feat % n_head == 0
# We assume d_v always equals d_k
self.d_k = n_feat // n_head
self.h = n_head
self.linear_q = nn.Linear(n_feat, n_feat)
self.linear_k = nn.Linear(n_feat, n_feat)
self.linear_v = nn.Linear(n_feat, n_feat)
self.proj = nn.Linear(n_feat, n_feat)
self.attn = None
self.dropout = nn.Dropout(p=attn_drop)
self.proj_drop = nn.Dropout(proj_drop)
def forward_qkv(self, query, key, value):
"""Transform query, key and value.
Args:
query (torch.Tensor): Query tensor (#batch, time1, size).
key (torch.Tensor): Key tensor (#batch, time2, size).
value (torch.Tensor): Value tensor (#batch, time2, size).
Returns:
torch.Tensor: Transformed query tensor (#batch, n_head, time1, d_k).
torch.Tensor: Transformed key tensor (#batch, n_head, time2, d_k).
torch.Tensor: Transformed value tensor (#batch, n_head, time2, d_k).
"""
n_batch = query.size(0)
q = self.linear_q(query).view(n_batch, -1, self.h, self.d_k)
k = self.linear_k(key).view(n_batch, -1, self.h, self.d_k)
v = self.linear_v(value).view(n_batch, -1, self.h, self.d_k)
q = q.transpose(1, 2)
k = k.transpose(1, 2)
v = v.transpose(1, 2)
return q, k, v
def forward_attention(self, value, scores, mask):
"""Compute attention context vector.
Args:
value (torch.Tensor): Transformed value (#batch, n_head, time2, d_k).
scores (torch.Tensor): Attention score (#batch, n_head, time1, time2).
mask (torch.Tensor): Mask (#batch, 1, time2) or (#batch, time1, time2).
Returns:
torch.Tensor: Transformed value (#batch, time1, d_model)
weighted by the attention score (#batch, time1, time2).
"""
n_batch = value.size(0)
if mask is not None:
mask = mask.unsqueeze(1).eq(0)
min_value = torch.finfo(scores.dtype).min
scores = scores.masked_fill(mask, min_value)
self.attn = torch.softmax(scores, dim=-1).masked_fill(mask, 0.0)
else:
self.attn = torch.softmax(scores, dim=-1)
p_attn = self.dropout(self.attn)
x = torch.matmul(p_attn, value)
x = x.transpose(1, 2).contiguous().view(n_batch, -1, self.h * self.d_k)
return self.proj_drop(self.proj(x))
def forward(self, x, cond, mask):
"""Compute scaled dot product attention.
Args:
query (torch.Tensor): Query tensor (#batch, time1, size).
key (torch.Tensor): Key tensor (#batch, time2, size).
value (torch.Tensor): Value tensor (#batch, time2, size).
mask (torch.Tensor): Mask tensor (#batch, 1, time2) or
(#batch, time1, time2).
Returns:
torch.Tensor: Output tensor (#batch, time1, d_model).
"""
q, k, v = self.forward_qkv(x, cond, cond)
scores = torch.matmul(q, k.transpose(-2, -1)) / math.sqrt(self.d_k)
return self.forward_attention(v, scores, mask)
class MultiHeadCrossAttention(nn.Module):
def __init__(
self, d_model, num_heads, attn_drop=0.0, proj_drop=0.0, **block_kwargs
):
super(MultiHeadCrossAttention, self).__init__()
assert d_model % num_heads == 0, "d_model must be divisible by num_heads"
self.d_model = d_model
self.num_heads = num_heads
self.head_dim = d_model // num_heads
self.q_linear = nn.Linear(d_model, d_model)
self.kv_linear = nn.Linear(d_model, d_model * 2)
self.attn_drop = nn.Dropout(attn_drop)
self.proj = nn.Linear(d_model, d_model)
self.proj_drop = nn.Dropout(proj_drop)
def forward(self, x, cond, mask=None):
# query/value: img tokens; key: condition; mask: if padding tokens
B, N, C = x.shape
assert mask is not None
q = self.q_linear(x).view(B, N, self.num_heads, self.head_dim)
if isinstance(mask, list):
# mask = y_lens: list of kv sequence lengths per batch item.
# cond is (1, total_valid_tokens, C) — batch flattened, padding removed.
kv = self.kv_linear(cond).view(1, -1, 2, self.num_heads, self.head_dim)
k, v = kv.unbind(2)
total_kv = k.shape[1]
# (B, N, heads, dim) → (B, heads, N, dim)
q = q.permute(0, 2, 1, 3)
# (1, total_kv, heads, dim) → (B, heads, total_kv, dim)
k = k.permute(0, 2, 1, 3).expand(B, -1, -1, -1)
v = v.permute(0, 2, 1, 3).expand(B, -1, -1, -1)
# Build block-diagonal attention mask from sequence lengths
attn_mask = torch.full(
(B, 1, N, total_kv), float("-inf"), dtype=q.dtype, device=q.device
)
offset = 0
for i, kv_len in enumerate(mask):
attn_mask[i, :, :, offset:offset + kv_len] = 0
offset += kv_len
else:
# mask is a padding mask tensor (B, 1, N_kv)
kv = self.kv_linear(cond).view(B, -1, 2, self.num_heads, self.head_dim)
k, v = kv.unbind(2)
# (B, N, heads, dim) → (B, heads, N, dim)
q = q.permute(0, 2, 1, 3)
k = k.permute(0, 2, 1, 3)
v = v.permute(0, 2, 1, 3)
attn_mask = torch.zeros_like(mask, dtype=q.dtype)
attn_mask.masked_fill_(mask == 0, float("-inf"))
attn_mask = attn_mask.unsqueeze(1)
x = F.scaled_dot_product_attention(
q, k, v, dropout_p=self.attn_drop.p, attn_mask=attn_mask
)
# (B, heads, N, dim) → (B, N, C)
x = x.permute(0, 2, 1, 3).reshape(B, N, C)
x = self.proj(x)
x = self.proj_drop(x)
return x
class WindowAttention(Attention_):
"""Multi-head Attention block with relative position embeddings."""
def __init__(
self,
dim,
num_heads=8,
qkv_bias=True,
use_rel_pos=False,
rel_pos_zero_init=True,
input_size=None,
**block_kwargs,
):
"""
Args:
dim (int): Number of input channels.
num_heads (int): Number of attention heads.
qkv_bias (bool: If True, add a learnable bias to query, key, value.
rel_pos (bool): If True, add relative positional embeddings to the attention map.
rel_pos_zero_init (bool): If True, zero initialize relative positional parameters.
input_size (int or None): Input resolution for calculating the relative positional
parameter size.
"""
super().__init__(dim, num_heads=num_heads, qkv_bias=qkv_bias, **block_kwargs)
self.use_rel_pos = use_rel_pos
if self.use_rel_pos:
# initialize relative positional embeddings
self.rel_pos_h = nn.Parameter(
torch.zeros(2 * input_size[0] - 1, self.head_dim)
)
self.rel_pos_w = nn.Parameter(
torch.zeros(2 * input_size[1] - 1, self.head_dim)
)
if not rel_pos_zero_init:
nn.init.trunc_normal_(self.rel_pos_h, std=0.02)
nn.init.trunc_normal_(self.rel_pos_w, std=0.02)
def forward(self, x, mask=None):
B, N, C = x.shape
qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, C // self.num_heads)
q, k, v = qkv.unbind(2)
if use_fp32_attention := getattr(self, "fp32_attention", False):
q, k, v = q.float(), k.float(), v.float()
# (B, N, heads, dim) → (B, heads, N, dim) for SDPA
q = q.permute(0, 2, 1, 3)
k = k.permute(0, 2, 1, 3)
v = v.permute(0, 2, 1, 3)
attn_mask = None
if mask is not None:
attn_mask = torch.zeros(
B * self.num_heads, q.shape[2], k.shape[2],
dtype=q.dtype, device=q.device,
)
attn_mask.masked_fill_(
mask.squeeze(1).repeat(self.num_heads, 1, 1) == 0, float("-inf")
)
attn_mask = attn_mask.view(B, self.num_heads, q.shape[2], k.shape[2])
x = F.scaled_dot_product_attention(
q, k, v, dropout_p=self.attn_drop.p, attn_mask=attn_mask
)
# (B, heads, N, dim) → (B, N, C)
x = x.permute(0, 2, 1, 3).reshape(B, N, C)
x = self.proj(x)
x = self.proj_drop(x)
return x
#################################################################################
# AMP attention with fp32 softmax to fix loss NaN problem during training #
#################################################################################
class Attention(Attention_):
def forward(self, x):
B, N, C = x.shape
qkv = (
self.qkv(x)
.reshape(B, N, 3, self.num_heads, C // self.num_heads)
.permute(2, 0, 3, 1, 4)
)
q, k, v = qkv.unbind(0) # make torchscript happy (cannot use tensor as tuple)
use_fp32_attention = getattr(self, "fp32_attention", False)
if use_fp32_attention:
q, k = q.float(), k.float()
with torch.cuda.amp.autocast(enabled=not use_fp32_attention):
attn = (q @ k.transpose(-2, -1)) * self.scale
attn = attn.softmax(dim=-1)
attn = self.attn_drop(attn)
x = (attn @ v).transpose(1, 2).reshape(B, N, C)
x = self.proj(x)
x = self.proj_drop(x)
return x
class FinalLayer(nn.Module):
"""
The final layer of PixArt.
"""
def __init__(self, hidden_size, patch_size, out_channels):
super().__init__()
self.norm_final = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
self.linear = nn.Linear(
hidden_size, patch_size * patch_size * out_channels, bias=True
)
self.adaLN_modulation = nn.Sequential(
nn.SiLU(), nn.Linear(hidden_size, 2 * hidden_size, bias=True)
)
def forward(self, x, c):
shift, scale = self.adaLN_modulation(c).chunk(2, dim=1)
x = modulate(self.norm_final(x), shift, scale)
x = self.linear(x)
return x
class T2IFinalLayer(nn.Module):
"""
The final layer of PixArt.
"""
def __init__(self, hidden_size, patch_size, out_channels):
super().__init__()
self.norm_final = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
self.linear = nn.Linear(
hidden_size, patch_size[0] * patch_size[1] * out_channels, bias=True
)
self.scale_shift_table = nn.Parameter(
torch.randn(2, hidden_size) / hidden_size**0.5
)
self.out_channels = out_channels
self.initialize_weights()
def initialize_weights(self):
def _basic_init(module):
if isinstance(module, nn.Linear):
torch.nn.init.xavier_uniform_(module.weight)
if module.bias is not None:
nn.init.constant_(module.bias, 0)
self.apply(_basic_init)
def forward(self, x, t):
shift, scale = (self.scale_shift_table[None] + t[:, None]).chunk(2, dim=1)
x = t2i_modulate(self.norm_final(x), shift, scale)
x = self.linear(x)
return x
class MaskFinalLayer(nn.Module):
"""
The final layer of PixArt.
"""
def __init__(self, final_hidden_size, c_emb_size, patch_size, out_channels):
super().__init__()
self.norm_final = nn.LayerNorm(
final_hidden_size, elementwise_affine=False, eps=1e-6
)
self.linear = nn.Linear(
final_hidden_size, patch_size * patch_size * out_channels, bias=True
)
self.adaLN_modulation = nn.Sequential(
nn.SiLU(), nn.Linear(c_emb_size, 2 * final_hidden_size, bias=True)
)
def forward(self, x, t):
shift, scale = self.adaLN_modulation(t).chunk(2, dim=1)
x = modulate(self.norm_final(x), shift, scale)
x = self.linear(x)
return x
class DecoderLayer(nn.Module):
"""
The final layer of PixArt.
"""
def __init__(self, hidden_size, decoder_hidden_size):
super().__init__()
self.norm_decoder = nn.LayerNorm(
hidden_size, elementwise_affine=False, eps=1e-6
)
self.linear = nn.Linear(hidden_size, decoder_hidden_size, bias=True)
self.adaLN_modulation = nn.Sequential(
nn.SiLU(), nn.Linear(hidden_size, 2 * hidden_size, bias=True)
)
def forward(self, x, t):
shift, scale = self.adaLN_modulation(t).chunk(2, dim=1)
x = modulate(self.norm_decoder(x), shift, scale)
x = self.linear(x)
return x
#################################################################################
# Embedding Layers for Timesteps and Class Labels #
#################################################################################
class TimestepEmbedder(nn.Module):
"""
Embeds scalar timesteps into vector representations.
"""
def __init__(self, hidden_size, frequency_embedding_size=256):
super().__init__()
self.mlp = nn.Sequential(
nn.Linear(frequency_embedding_size, hidden_size, bias=True),
nn.SiLU(),
nn.Linear(hidden_size, hidden_size, bias=True),
)
self.frequency_embedding_size = frequency_embedding_size
@staticmethod
def timestep_embedding(t, dim, max_period=10000):
"""
Create sinusoidal timestep embeddings.
:param t: a 1-D Tensor of N indices, one per batch element.
These may be fractional.
:param dim: the dimension of the output.
:param max_period: controls the minimum frequency of the embeddings.
:return: an (N, D) Tensor of positional embeddings.
"""
# https://github.com/openai/glide-text2im/blob/main/glide_text2im/nn.py
half = dim // 2
freqs = torch.exp(
-math.log(max_period)
* torch.arange(start=0, end=half, dtype=torch.float32, device=t.device)
/ half
)
args = t[:, None].float() * freqs[None]
embedding = torch.cat([torch.cos(args), torch.sin(args)], dim=-1)
if dim % 2:
embedding = torch.cat(
[embedding, torch.zeros_like(embedding[:, :1])], dim=-1
)
return embedding
def forward(self, t):
t_freq = self.timestep_embedding(t, self.frequency_embedding_size).to(
self.dtype
)
return self.mlp(t_freq)
@property
def dtype(self):
return next(self.parameters()).dtype
class SizeEmbedder(TimestepEmbedder):
"""
Embeds scalar timesteps into vector representations.
"""
def __init__(self, hidden_size, frequency_embedding_size=256):
super().__init__(
hidden_size=hidden_size, frequency_embedding_size=frequency_embedding_size
)
self.mlp = nn.Sequential(
nn.Linear(frequency_embedding_size, hidden_size, bias=True),
nn.SiLU(),
nn.Linear(hidden_size, hidden_size, bias=True),
)
self.frequency_embedding_size = frequency_embedding_size
self.outdim = hidden_size
def forward(self, s, bs):
if s.ndim == 1:
s = s[:, None]
assert s.ndim == 2
if s.shape[0] != bs:
s = s.repeat(bs // s.shape[0], 1)
assert s.shape[0] == bs
b, dims = s.shape[0], s.shape[1]
s = rearrange(s, "b d -> (b d)")
s_freq = self.timestep_embedding(s, self.frequency_embedding_size).to(
self.dtype
)
s_emb = self.mlp(s_freq)
s_emb = rearrange(s_emb, "(b d) d2 -> b (d d2)", b=b, d=dims, d2=self.outdim)
return s_emb
@property
def dtype(self):
return next(self.parameters()).dtype
class LabelEmbedder(nn.Module):
"""
Embeds class labels into vector representations. Also handles label dropout for classifier-free guidance.
"""
def __init__(self, num_classes, hidden_size, dropout_prob):
super().__init__()
use_cfg_embedding = dropout_prob > 0
self.embedding_table = nn.Embedding(
num_classes + use_cfg_embedding, hidden_size
)
self.num_classes = num_classes
self.dropout_prob = dropout_prob
def token_drop(self, labels, force_drop_ids=None):
"""
Drops labels to enable classifier-free guidance.
"""
if force_drop_ids is None:
drop_ids = torch.rand(labels.shape[0]).cuda() < self.dropout_prob
else:
drop_ids = force_drop_ids == 1
labels = torch.where(drop_ids, self.num_classes, labels)
return labels
def forward(self, labels, train, force_drop_ids=None):
use_dropout = self.dropout_prob > 0
if (train and use_dropout) or (force_drop_ids is not None):
labels = self.token_drop(labels, force_drop_ids)
return self.embedding_table(labels)
class CaptionEmbedder(nn.Module):
"""
Embeds class labels into vector representations. Also handles label dropout for classifier-free guidance.
"""
def __init__(
self,
in_channels,
hidden_size,
uncond_prob,
act_layer=nn.GELU(approximate="tanh"),
token_num=120,
):
super().__init__()
self.y_proj = Mlp(
in_features=in_channels,
hidden_features=hidden_size,
out_features=hidden_size,
act_layer=act_layer,
drop=0,
)
self.register_buffer(
"y_embedding",
nn.Parameter(torch.randn(token_num, in_channels) / in_channels**0.5),
)
self.uncond_prob = uncond_prob
def token_drop(self, caption, force_drop_ids=None):
"""
Drops labels to enable classifier-free guidance.
"""
if force_drop_ids is None:
drop_ids = torch.rand(caption.shape[0]).cuda() < self.uncond_prob
else:
drop_ids = force_drop_ids == 1
caption = torch.where(drop_ids[:, None, None, None], self.y_embedding, caption)
return caption
def forward(self, caption, train, force_drop_ids=None):
if train:
assert caption.shape[2:] == self.y_embedding.shape
use_dropout = self.uncond_prob > 0
if (train and use_dropout) or (force_drop_ids is not None):
caption = self.token_drop(caption, force_drop_ids)
caption = self.y_proj(caption)
return caption
class CaptionEmbedderDoubleBr(nn.Module):
"""
Embeds class labels into vector representations. Also handles label dropout for classifier-free guidance.
"""
def __init__(
self,
in_channels,
hidden_size,
uncond_prob,
act_layer=nn.GELU(approximate="tanh"),
token_num=120,
):
super().__init__()
self.proj = Mlp(
in_features=in_channels,
hidden_features=hidden_size,
out_features=hidden_size,
act_layer=act_layer,
drop=0,
)
self.embedding = nn.Parameter(torch.randn(1, in_channels) / 10**0.5)
self.y_embedding = nn.Parameter(torch.randn(token_num, in_channels) / 10**0.5)
self.uncond_prob = uncond_prob
def token_drop(self, global_caption, caption, force_drop_ids=None):
"""
Drops labels to enable classifier-free guidance.
"""
if force_drop_ids is None:
drop_ids = torch.rand(global_caption.shape[0]).cuda() < self.uncond_prob
else:
drop_ids = force_drop_ids == 1
global_caption = torch.where(drop_ids[:, None], self.embedding, global_caption)
caption = torch.where(drop_ids[:, None, None, None], self.y_embedding, caption)
return global_caption, caption
def forward(self, caption, train, force_drop_ids=None):
assert caption.shape[2:] == self.y_embedding.shape
global_caption = caption.mean(dim=2).squeeze()
use_dropout = self.uncond_prob > 0
if (train and use_dropout) or (force_drop_ids is not None):
global_caption, caption = self.token_drop(
global_caption, caption, force_drop_ids
)
y_embed = self.proj(global_caption)
return y_embed, caption
README.md
View file @ a44e71bd
# AudioFly
## 论文
暂无
## 模型简介
AudioFly 是一个音频生成模型。它根据文本描述合成音效。该模型可以以 44.1 kHz 的采样率生成高质量音频。生成的音频与提示文本有很强的一致性。
AudioFly 采用了潜在扩散模型架构。该模型拥有 10 亿个参数，并在大量多样化的语料库上进行了训练。训练数据包括开源数据集，如 AudioSet、AudioCaps 和 TUT，以及专有的内部数据。该模型在单一事件和多事件场景中表现良好。在这两种情况下，生成的音频都能准确反映所描述的内容。在 AudioCaps 数据集上，AudioF
\ No newline at end of file
AudioFly 采用了潜在扩散模型架构。该模型拥有 10 亿个参数，并在大量多样化的语料库上进行了训练。训练数据包括开源数据集，如 AudioSet、AudioCaps 和 TUT，以及专有的内部数据。该模型在单一事件和多事件场景中表现良好。在这两种情况下，生成的音频都能准确反映所描述的内容。在 AudioCaps 数据集上，AudioFly 的性能优于之前的音频生成模型。
## 环境依赖
| 软件 | 版本 |
| :------: | :------: |
| DTK | 26.04 |
| Python | 3.10 |
| Transformers | 4.56.1 |
| vLLM | 0.18.1+das.dtk2604 |
推荐使用镜像:
harbor.sourcefind.cn:5443/dcu/admin/base/custom:vllm0.18.1-ubuntu22.04-dtk26.04-py3.10-20260529-iflytek
- 挂载地址`-v` 根据实际模型情况修改
```bash
docker run -it \
--shm-size 60g \
--network=host \
--name Spark-X1 \
--privileged \
--device=/dev/kfd \
--device=/dev/dri \
--device=/dev/mkfd \
--group-add video \
--cap-add=SYS_PTRACE \
--security-opt seccomp=unconfined \
-u root \
-v /opt/hyhal/:/opt/hyhal/:ro \
-v /path/your_code_data/:/path/your_code_data/ \
harbor.sourcefind.cn:5443/dcu/admin/base/custom:vllm0.18.1-ubuntu22.04-dtk26.04-py3.10-20260529-iflytek bash
```
更多镜像可前往[光源](https://sourcefind.cn/#/service-list)下载使用。
关于本项目DCU显卡所需的特殊深度学习库可从[光合](https://developer.sourcefind.cn/tool/)开发者社区下载安装。
其它环境配置说明：
```bash
#下载模型以后，替换下载路径内的文件
cp PixArt_blocks.py AudioFly/ldm/modules/diffusionmodules/PixArt_blocks.py
cp PixArt.py AudioFly/ldm/modules/diffusionmodules/PixArt.py
```
## 预训练权重
**请根据`支持的DCU型号`选择对应模型下载，FP8模型仅在BW1100/BW1101上支持，其他型号请勿使用！**
| 模型名称 | 权重大小 | 数据类型 | 支持的DCU型号 | 最低卡数需求 | 下载地址 |
| :------: | :------: | :------: | :------------: | :----------: | :------: |
| AudioFly | 1B | BF16 | BW1000 | 1 | [ModelScope](https://modelscope.cn/models/iflytek/AudioFly) |
## 数据集
暂无
## 训练
暂无
## 推理
### Pytorch
#### 单机推理
```bash
cd AudioFly
python run.py
```
## 效果展示
输入：
'Fierce winds howl through the valley'
输出：
<audio controls src="./doc/result.wav"></audio>
### 精度
DCU与GPU精度一致，推理框架：pytorch
## 源码仓库及问题反馈
- https://developer.sourcefind.cn/codes/modelzoo/audiofly
## 参考资料
- https://modelscope.csdn.net/68da3b11a6dc56200e8ae2ae.html
- https://modelscope.cn/models/iflytek/AudioFly
\ No newline at end of file
doc/result.wav 0 → 100644
View file @ a44e71bd
File added
icon.png 0 → 100644
View file @ a44e71bd
icon.png

64.4 KB

model.properties 0 → 100644
View file @ a44e71bd
# 模型唯一标识
modelCode=15319
# 模型名称
modelName=AudioFly
# 模型描述
modelDescription=AudioFly 是一个音频生成模型。它根据文本描述合成音效。该模型可以以 44.1 kHz 的采样率生成高质量音频。
# 运行过程
processType=推理
# 算法类别
appCategory=语音合成
# 框架类型
frameType=pytorch
# 加速卡类型
accelerateType=BW1000
run.py 0 → 100644
View file @ a44e71bd
import yaml
import torch
from ldm.utils.util import instantiate_from_config
configs = yaml.load(open('/public/home/weishb/iflytek/AudioFly/config/config.yaml', "r"), Loader=yaml.FullLoader)
model = instantiate_from_config(configs["model"])
checkpoint = torch.load('/public/home/weishb/iflytek/AudioFly/models/ldm/model.ckpt')
model.load_state_dict(checkpoint, strict=False)
model.eval()
model = model.cuda()
text = 'Fierce winds howl through the valley'
name = 'result'
savedir = './result'
model.generate_sample(
textlist=[text],
name=name,
cfg=3.5,# Guidance scale (controls how strongly generation follows the text prompt）; not recommended to change
ddim_steps=200, # Number of denoising steps in the diffusion process; not recommended to change
outputdir=f"{savedir}")
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