unet_2d_condition.py 51.4 KB
Newer Older
Patrick von Platen's avatar
Patrick von Platen committed
1
# Copyright 2023 The HuggingFace Team. All rights reserved.
2
3
4
5
6
7
8
9
10
11
12
13
#
# 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.
14
from dataclasses import dataclass
15
from typing import Any, Dict, List, Optional, Tuple, Union
Patrick von Platen's avatar
Patrick von Platen committed
16
17
18

import torch
import torch.nn as nn
19
import torch.utils.checkpoint
Patrick von Platen's avatar
Patrick von Platen committed
20
21

from ..configuration_utils import ConfigMixin, register_to_config
22
from ..loaders import UNet2DConditionLoadersMixin
23
from ..utils import BaseOutput, logging
24
from .activations import get_activation
25
from .attention_processor import AttentionProcessor, AttnProcessor
YiYi Xu's avatar
YiYi Xu committed
26
27
from .embeddings import (
    GaussianFourierProjection,
YiYi Xu's avatar
YiYi Xu committed
28
29
30
    ImageHintTimeEmbedding,
    ImageProjection,
    ImageTimeEmbedding,
31
    PositionNet,
YiYi Xu's avatar
YiYi Xu committed
32
33
34
35
36
37
    TextImageProjection,
    TextImageTimeEmbedding,
    TextTimeEmbedding,
    TimestepEmbedding,
    Timesteps,
)
38
from .modeling_utils import ModelMixin
39
from .unet_2d_blocks import (
40
    UNetMidBlock2DCrossAttn,
Will Berman's avatar
Will Berman committed
41
    UNetMidBlock2DSimpleCrossAttn,
42
43
44
    get_down_block,
    get_up_block,
)
Patrick von Platen's avatar
Patrick von Platen committed
45
46


47
48
49
logger = logging.get_logger(__name__)  # pylint: disable=invalid-name


50
51
52
@dataclass
class UNet2DConditionOutput(BaseOutput):
    """
Steven Liu's avatar
Steven Liu committed
53
54
    The output of [`UNet2DConditionModel`].

55
56
    Args:
        sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`):
Steven Liu's avatar
Steven Liu committed
57
            The hidden states output conditioned on `encoder_hidden_states` input. Output of last layer of model.
58
59
    """

60
    sample: torch.FloatTensor = None
61
62


63
class UNet2DConditionModel(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin):
Kashif Rasul's avatar
Kashif Rasul committed
64
    r"""
Steven Liu's avatar
Steven Liu committed
65
66
    A conditional 2D UNet model that takes a noisy sample, conditional state, and a timestep and returns a sample
    shaped output.
Kashif Rasul's avatar
Kashif Rasul committed
67

Steven Liu's avatar
Steven Liu committed
68
69
    This model inherits from [`ModelMixin`]. Check the superclass documentation for it's generic methods implemented
    for all models (such as downloading or saving).
Kashif Rasul's avatar
Kashif Rasul committed
70
71

    Parameters:
72
73
        sample_size (`int` or `Tuple[int, int]`, *optional*, defaults to `None`):
            Height and width of input/output sample.
Steven Liu's avatar
Steven Liu committed
74
75
        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.
Kashif Rasul's avatar
Kashif Rasul committed
76
        center_input_sample (`bool`, *optional*, defaults to `False`): Whether to center the input sample.
Suraj Patil's avatar
Suraj Patil committed
77
        flip_sin_to_cos (`bool`, *optional*, defaults to `False`):
Kashif Rasul's avatar
Kashif Rasul committed
78
79
80
81
            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.
Will Berman's avatar
Will Berman committed
82
        mid_block_type (`str`, *optional*, defaults to `"UNetMidBlock2DCrossAttn"`):
Steven Liu's avatar
Steven Liu committed
83
84
85
            Block type for middle of UNet, it can be either `UNetMidBlock2DCrossAttn` or
            `UNetMidBlock2DSimpleCrossAttn`. If `None`, the mid block layer is skipped.
        up_block_types (`Tuple[str]`, *optional*, defaults to `("UpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D")`):
Kashif Rasul's avatar
Kashif Rasul committed
86
            The tuple of upsample blocks to use.
87
88
89
        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`].
Kashif Rasul's avatar
Kashif Rasul committed
90
91
92
93
94
95
96
        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.
        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.
Steven Liu's avatar
Steven Liu committed
97
            If `None`, normalization and activation layers is skipped in post-processing.
Kashif Rasul's avatar
Kashif Rasul committed
98
        norm_eps (`float`, *optional*, defaults to 1e-5): The epsilon to use for the normalization.
Sanchit Gandhi's avatar
Sanchit Gandhi committed
99
100
        cross_attention_dim (`int` or `Tuple[int]`, *optional*, defaults to 1280):
            The dimension of the cross attention features.
101
102
103
104
105
        transformer_layers_per_block (`int` or `Tuple[int]`, *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`].
        encoder_hid_dim (`int`, *optional*, defaults to None):
YiYi Xu's avatar
YiYi Xu committed
106
107
            If `encoder_hid_dim_type` is defined, `encoder_hidden_states` will be projected from `encoder_hid_dim`
            dimension to `cross_attention_dim`.
Steven Liu's avatar
Steven Liu committed
108
109
        encoder_hid_dim_type (`str`, *optional*, defaults to `None`):
            If given, the `encoder_hidden_states` and potentially other embeddings are down-projected to text
YiYi Xu's avatar
YiYi Xu committed
110
            embeddings of dimension `cross_attention` according to `encoder_hid_dim_type`.
Kashif Rasul's avatar
Kashif Rasul committed
111
        attention_head_dim (`int`, *optional*, defaults to 8): The dimension of the attention heads.
112
113
        num_attention_heads (`int`, *optional*):
            The number of attention heads. If not defined, defaults to `attention_head_dim`
Will Berman's avatar
Will Berman committed
114
        resnet_time_scale_shift (`str`, *optional*, defaults to `"default"`): Time scale shift config
Steven Liu's avatar
Steven Liu committed
115
116
            for ResNet blocks (see [`~models.resnet.ResnetBlock2D`]). Choose from `default` or `scale_shift`.
        class_embed_type (`str`, *optional*, defaults to `None`):
117
            The type of class embedding to use which is ultimately summed with the time embeddings. Choose from `None`,
Sanchit Gandhi's avatar
Sanchit Gandhi committed
118
            `"timestep"`, `"identity"`, `"projection"`, or `"simple_projection"`.
Steven Liu's avatar
Steven Liu committed
119
        addition_embed_type (`str`, *optional*, defaults to `None`):
Patrick von Platen's avatar
Patrick von Platen committed
120
121
            Configures an optional embedding which will be summed with the time embeddings. Choose from `None` or
            "text". "text" will use the `TextTimeEmbedding` layer.
122
123
        addition_time_embed_dim: (`int`, *optional*, defaults to `None`):
            Dimension for the timestep embeddings.
Steven Liu's avatar
Steven Liu committed
124
        num_class_embeds (`int`, *optional*, defaults to `None`):
125
126
            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`.
Steven Liu's avatar
Steven Liu committed
127
        time_embedding_type (`str`, *optional*, defaults to `positional`):
128
            The type of position embedding to use for timesteps. Choose from `positional` or `fourier`.
Steven Liu's avatar
Steven Liu committed
129
        time_embedding_dim (`int`, *optional*, defaults to `None`):
Patrick von Platen's avatar
Patrick von Platen committed
130
            An optional override for the dimension of the projected time embedding.
Steven Liu's avatar
Steven Liu committed
131
132
133
134
        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`):
135
            The second activation function to use in timestep embedding. Choose from `silu`, `mish` and `gelu`.
Steven Liu's avatar
Steven Liu committed
136
137
        time_cond_proj_dim (`int`, *optional*, defaults to `None`):
            The dimension of `cond_proj` layer in the timestep embedding.
138
        conv_in_kernel (`int`, *optional*, default to `3`): The kernel size of `conv_in` layer.
Will Berman's avatar
Will Berman committed
139
140
        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
Steven Liu's avatar
Steven Liu committed
141
            `class_embed_type="projection"`. Required when `class_embed_type="projection"`.
Sanchit Gandhi's avatar
Sanchit Gandhi committed
142
        class_embeddings_concat (`bool`, *optional*, defaults to `False`): Whether to concatenate the time
143
144
145
            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
Steven Liu's avatar
Steven Liu committed
146
147
148
            `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.
Kashif Rasul's avatar
Kashif Rasul committed
149
150
    """

151
152
    _supports_gradient_checkpointing = True

Patrick von Platen's avatar
Patrick von Platen committed
153
154
155
    @register_to_config
    def __init__(
        self,
Sid Sahai's avatar
Sid Sahai committed
156
157
158
159
160
161
162
163
164
165
166
167
        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",
        ),
168
        mid_block_type: Optional[str] = "UNetMidBlock2DCrossAttn",
Sid Sahai's avatar
Sid Sahai committed
169
        up_block_types: Tuple[str] = ("UpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D"),
170
        only_cross_attention: Union[bool, Tuple[bool]] = False,
Sid Sahai's avatar
Sid Sahai committed
171
        block_out_channels: Tuple[int] = (320, 640, 1280, 1280),
172
        layers_per_block: Union[int, Tuple[int]] = 2,
Sid Sahai's avatar
Sid Sahai committed
173
174
175
        downsample_padding: int = 1,
        mid_block_scale_factor: float = 1,
        act_fn: str = "silu",
176
        norm_num_groups: Optional[int] = 32,
Sid Sahai's avatar
Sid Sahai committed
177
        norm_eps: float = 1e-5,
Sanchit Gandhi's avatar
Sanchit Gandhi committed
178
        cross_attention_dim: Union[int, Tuple[int]] = 1280,
179
        transformer_layers_per_block: Union[int, Tuple[int]] = 1,
William Berman's avatar
William Berman committed
180
        encoder_hid_dim: Optional[int] = None,
YiYi Xu's avatar
YiYi Xu committed
181
        encoder_hid_dim_type: Optional[str] = None,
Suraj Patil's avatar
Suraj Patil committed
182
        attention_head_dim: Union[int, Tuple[int]] = 8,
183
        num_attention_heads: Optional[Union[int, Tuple[int]]] = None,
184
        dual_cross_attention: bool = False,
Suraj Patil's avatar
Suraj Patil committed
185
        use_linear_projection: bool = False,
Will Berman's avatar
Will Berman committed
186
        class_embed_type: Optional[str] = None,
Patrick von Platen's avatar
Patrick von Platen committed
187
        addition_embed_type: Optional[str] = None,
188
        addition_time_embed_dim: Optional[int] = None,
189
        num_class_embeds: Optional[int] = None,
190
        upcast_attention: bool = False,
Will Berman's avatar
Will Berman committed
191
        resnet_time_scale_shift: str = "default",
192
193
        resnet_skip_time_act: bool = False,
        resnet_out_scale_factor: int = 1.0,
194
        time_embedding_type: str = "positional",
Patrick von Platen's avatar
Patrick von Platen committed
195
        time_embedding_dim: Optional[int] = None,
196
        time_embedding_act_fn: Optional[str] = None,
197
198
199
200
        timestep_post_act: Optional[str] = None,
        time_cond_proj_dim: Optional[int] = None,
        conv_in_kernel: int = 3,
        conv_out_kernel: int = 3,
Will Berman's avatar
Will Berman committed
201
        projection_class_embeddings_input_dim: Optional[int] = None,
202
        attention_type: str = "default",
Sanchit Gandhi's avatar
Sanchit Gandhi committed
203
        class_embeddings_concat: bool = False,
204
        mid_block_only_cross_attention: Optional[bool] = None,
205
        cross_attention_norm: Optional[str] = None,
Patrick von Platen's avatar
Patrick von Platen committed
206
        addition_embed_type_num_heads=64,
Patrick von Platen's avatar
Patrick von Platen committed
207
208
209
210
211
    ):
        super().__init__()

        self.sample_size = sample_size

212
213
214
215
216
        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."
            )

217
218
219
220
221
222
223
224
        # 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

Will Berman's avatar
Will Berman committed
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
        # 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}."
            )

241
242
243
244
245
        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}."
            )

Will Berman's avatar
Will Berman committed
246
247
248
249
250
        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}."
            )

Sanchit Gandhi's avatar
Sanchit Gandhi committed
251
252
253
254
255
        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}."
            )

256
257
258
259
260
        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}."
            )

Patrick von Platen's avatar
Patrick von Platen committed
261
        # input
262
263
264
265
        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
        )
Patrick von Platen's avatar
Patrick von Platen committed
266
267

        # time
268
        if time_embedding_type == "fourier":
Patrick von Platen's avatar
Patrick von Platen committed
269
            time_embed_dim = time_embedding_dim or block_out_channels[0] * 2
270
271
272
273
274
275
276
            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":
Patrick von Platen's avatar
Patrick von Platen committed
277
            time_embed_dim = time_embedding_dim or block_out_channels[0] * 4
278
279
280
281
282

            self.time_proj = Timesteps(block_out_channels[0], flip_sin_to_cos, freq_shift)
            timestep_input_dim = block_out_channels[0]
        else:
            raise ValueError(
Alexander Pivovarov's avatar
Alexander Pivovarov committed
283
                f"{time_embedding_type} does not exist. Please make sure to use one of `fourier` or `positional`."
284
            )
Patrick von Platen's avatar
Patrick von Platen committed
285

286
287
288
289
290
291
292
        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,
        )
Patrick von Platen's avatar
Patrick von Platen committed
293

YiYi Xu's avatar
YiYi Xu committed
294
295
        if encoder_hid_dim_type is None and encoder_hid_dim is not None:
            encoder_hid_dim_type = "text_proj"
296
            self.register_to_config(encoder_hid_dim_type=encoder_hid_dim_type)
YiYi Xu's avatar
YiYi Xu committed
297
298
299
300
301
302
303
304
            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":
William Berman's avatar
William Berman committed
305
            self.encoder_hid_proj = nn.Linear(encoder_hid_dim, cross_attention_dim)
YiYi Xu's avatar
YiYi Xu committed
306
307
308
309
310
311
312
313
314
        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,
            )
YiYi Xu's avatar
YiYi Xu committed
315
316
317
318
319
320
        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,
            )
YiYi Xu's avatar
YiYi Xu committed
321
322
323
324
        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'."
            )
William Berman's avatar
William Berman committed
325
326
327
        else:
            self.encoder_hid_proj = None

328
        # class embedding
Will Berman's avatar
Will Berman committed
329
        if class_embed_type is None and num_class_embeds is not None:
330
            self.class_embedding = nn.Embedding(num_class_embeds, time_embed_dim)
Will Berman's avatar
Will Berman committed
331
        elif class_embed_type == "timestep":
332
            self.class_embedding = TimestepEmbedding(timestep_input_dim, time_embed_dim, act_fn=act_fn)
Will Berman's avatar
Will Berman committed
333
334
        elif class_embed_type == "identity":
            self.class_embedding = nn.Identity(time_embed_dim, time_embed_dim)
Will Berman's avatar
Will Berman committed
335
336
337
338
339
340
341
342
343
344
345
346
347
        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)
Sanchit Gandhi's avatar
Sanchit Gandhi committed
348
349
350
351
352
353
        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)
Will Berman's avatar
Will Berman committed
354
355
        else:
            self.class_embedding = None
356

Patrick von Platen's avatar
Patrick von Platen committed
357
358
359
360
361
362
363
364
365
        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
            )
YiYi Xu's avatar
YiYi Xu committed
366
367
368
369
370
371
372
        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
            )
373
374
375
        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)
YiYi Xu's avatar
YiYi Xu committed
376
377
378
379
380
381
        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)
Patrick von Platen's avatar
Patrick von Platen committed
382
        elif addition_embed_type is not None:
YiYi Xu's avatar
YiYi Xu committed
383
            raise ValueError(f"addition_embed_type: {addition_embed_type} must be None, 'text' or 'text_image'.")
Patrick von Platen's avatar
Patrick von Platen committed
384

385
386
387
        if time_embedding_act_fn is None:
            self.time_embed_act = None
        else:
388
            self.time_embed_act = get_activation(time_embedding_act_fn)
389

Patrick von Platen's avatar
Patrick von Platen committed
390
391
392
        self.down_blocks = nn.ModuleList([])
        self.up_blocks = nn.ModuleList([])

393
        if isinstance(only_cross_attention, bool):
394
395
396
            if mid_block_only_cross_attention is None:
                mid_block_only_cross_attention = only_cross_attention

397
398
            only_cross_attention = [only_cross_attention] * len(down_block_types)

399
400
401
        if mid_block_only_cross_attention is None:
            mid_block_only_cross_attention = False

402
403
404
        if isinstance(num_attention_heads, int):
            num_attention_heads = (num_attention_heads,) * len(down_block_types)

Suraj Patil's avatar
Suraj Patil committed
405
406
407
        if isinstance(attention_head_dim, int):
            attention_head_dim = (attention_head_dim,) * len(down_block_types)

Sanchit Gandhi's avatar
Sanchit Gandhi committed
408
409
410
        if isinstance(cross_attention_dim, int):
            cross_attention_dim = (cross_attention_dim,) * len(down_block_types)

411
412
413
        if isinstance(layers_per_block, int):
            layers_per_block = [layers_per_block] * len(down_block_types)

414
415
416
        if isinstance(transformer_layers_per_block, int):
            transformer_layers_per_block = [transformer_layers_per_block] * len(down_block_types)

Sanchit Gandhi's avatar
Sanchit Gandhi committed
417
418
419
420
421
422
423
424
        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

Patrick von Platen's avatar
Patrick von Platen committed
425
426
427
428
429
430
431
432
433
        # 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,
434
                num_layers=layers_per_block[i],
435
                transformer_layers_per_block=transformer_layers_per_block[i],
Patrick von Platen's avatar
Patrick von Platen committed
436
437
                in_channels=input_channel,
                out_channels=output_channel,
Sanchit Gandhi's avatar
Sanchit Gandhi committed
438
                temb_channels=blocks_time_embed_dim,
Patrick von Platen's avatar
Patrick von Platen committed
439
440
441
                add_downsample=not is_final_block,
                resnet_eps=norm_eps,
                resnet_act_fn=act_fn,
442
                resnet_groups=norm_num_groups,
Sanchit Gandhi's avatar
Sanchit Gandhi committed
443
                cross_attention_dim=cross_attention_dim[i],
444
                num_attention_heads=num_attention_heads[i],
Patrick von Platen's avatar
Patrick von Platen committed
445
                downsample_padding=downsample_padding,
446
                dual_cross_attention=dual_cross_attention,
Suraj Patil's avatar
Suraj Patil committed
447
                use_linear_projection=use_linear_projection,
448
                only_cross_attention=only_cross_attention[i],
449
                upcast_attention=upcast_attention,
Will Berman's avatar
Will Berman committed
450
                resnet_time_scale_shift=resnet_time_scale_shift,
451
                attention_type=attention_type,
452
453
                resnet_skip_time_act=resnet_skip_time_act,
                resnet_out_scale_factor=resnet_out_scale_factor,
454
                cross_attention_norm=cross_attention_norm,
455
                attention_head_dim=attention_head_dim[i] if attention_head_dim[i] is not None else output_channel,
Patrick von Platen's avatar
Patrick von Platen committed
456
457
458
459
            )
            self.down_blocks.append(down_block)

        # mid
Will Berman's avatar
Will Berman committed
460
461
        if mid_block_type == "UNetMidBlock2DCrossAttn":
            self.mid_block = UNetMidBlock2DCrossAttn(
462
                transformer_layers_per_block=transformer_layers_per_block[-1],
Will Berman's avatar
Will Berman committed
463
                in_channels=block_out_channels[-1],
Sanchit Gandhi's avatar
Sanchit Gandhi committed
464
                temb_channels=blocks_time_embed_dim,
Will Berman's avatar
Will Berman committed
465
466
467
468
                resnet_eps=norm_eps,
                resnet_act_fn=act_fn,
                output_scale_factor=mid_block_scale_factor,
                resnet_time_scale_shift=resnet_time_scale_shift,
Sanchit Gandhi's avatar
Sanchit Gandhi committed
469
                cross_attention_dim=cross_attention_dim[-1],
470
                num_attention_heads=num_attention_heads[-1],
Will Berman's avatar
Will Berman committed
471
472
473
474
                resnet_groups=norm_num_groups,
                dual_cross_attention=dual_cross_attention,
                use_linear_projection=use_linear_projection,
                upcast_attention=upcast_attention,
475
                attention_type=attention_type,
Will Berman's avatar
Will Berman committed
476
477
478
479
            )
        elif mid_block_type == "UNetMidBlock2DSimpleCrossAttn":
            self.mid_block = UNetMidBlock2DSimpleCrossAttn(
                in_channels=block_out_channels[-1],
Sanchit Gandhi's avatar
Sanchit Gandhi committed
480
                temb_channels=blocks_time_embed_dim,
Will Berman's avatar
Will Berman committed
481
482
483
                resnet_eps=norm_eps,
                resnet_act_fn=act_fn,
                output_scale_factor=mid_block_scale_factor,
Sanchit Gandhi's avatar
Sanchit Gandhi committed
484
                cross_attention_dim=cross_attention_dim[-1],
485
                attention_head_dim=attention_head_dim[-1],
Will Berman's avatar
Will Berman committed
486
487
                resnet_groups=norm_num_groups,
                resnet_time_scale_shift=resnet_time_scale_shift,
488
                skip_time_act=resnet_skip_time_act,
489
                only_cross_attention=mid_block_only_cross_attention,
490
                cross_attention_norm=cross_attention_norm,
Will Berman's avatar
Will Berman committed
491
            )
492
493
        elif mid_block_type is None:
            self.mid_block = None
Will Berman's avatar
Will Berman committed
494
495
        else:
            raise ValueError(f"unknown mid_block_type : {mid_block_type}")
Patrick von Platen's avatar
Patrick von Platen committed
496

497
498
499
        # count how many layers upsample the images
        self.num_upsamplers = 0

Patrick von Platen's avatar
Patrick von Platen committed
500
501
        # up
        reversed_block_out_channels = list(reversed(block_out_channels))
502
        reversed_num_attention_heads = list(reversed(num_attention_heads))
503
        reversed_layers_per_block = list(reversed(layers_per_block))
Sanchit Gandhi's avatar
Sanchit Gandhi committed
504
        reversed_cross_attention_dim = list(reversed(cross_attention_dim))
505
        reversed_transformer_layers_per_block = list(reversed(transformer_layers_per_block))
506
        only_cross_attention = list(reversed(only_cross_attention))
507

Patrick von Platen's avatar
Patrick von Platen committed
508
509
        output_channel = reversed_block_out_channels[0]
        for i, up_block_type in enumerate(up_block_types):
510
511
            is_final_block = i == len(block_out_channels) - 1

Patrick von Platen's avatar
Patrick von Platen committed
512
513
514
515
            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)]

516
517
518
519
520
521
            # add upsample block for all BUT final layer
            if not is_final_block:
                add_upsample = True
                self.num_upsamplers += 1
            else:
                add_upsample = False
Patrick von Platen's avatar
Patrick von Platen committed
522
523
524

            up_block = get_up_block(
                up_block_type,
525
                num_layers=reversed_layers_per_block[i] + 1,
526
                transformer_layers_per_block=reversed_transformer_layers_per_block[i],
Patrick von Platen's avatar
Patrick von Platen committed
527
528
529
                in_channels=input_channel,
                out_channels=output_channel,
                prev_output_channel=prev_output_channel,
Sanchit Gandhi's avatar
Sanchit Gandhi committed
530
                temb_channels=blocks_time_embed_dim,
531
                add_upsample=add_upsample,
Patrick von Platen's avatar
Patrick von Platen committed
532
533
                resnet_eps=norm_eps,
                resnet_act_fn=act_fn,
534
                resnet_groups=norm_num_groups,
Sanchit Gandhi's avatar
Sanchit Gandhi committed
535
                cross_attention_dim=reversed_cross_attention_dim[i],
536
                num_attention_heads=reversed_num_attention_heads[i],
537
                dual_cross_attention=dual_cross_attention,
Suraj Patil's avatar
Suraj Patil committed
538
                use_linear_projection=use_linear_projection,
539
                only_cross_attention=only_cross_attention[i],
540
                upcast_attention=upcast_attention,
Will Berman's avatar
Will Berman committed
541
                resnet_time_scale_shift=resnet_time_scale_shift,
542
                attention_type=attention_type,
543
544
                resnet_skip_time_act=resnet_skip_time_act,
                resnet_out_scale_factor=resnet_out_scale_factor,
545
                cross_attention_norm=cross_attention_norm,
546
                attention_head_dim=attention_head_dim[i] if attention_head_dim[i] is not None else output_channel,
Patrick von Platen's avatar
Patrick von Platen committed
547
548
549
550
551
            )
            self.up_blocks.append(up_block)
            prev_output_channel = output_channel

        # out
552
553
554
555
        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
            )
556

557
            self.conv_act = get_activation(act_fn)
558

559
560
561
562
563
564
565
566
        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
        )
Patrick von Platen's avatar
Patrick von Platen committed
567

568
569
570
571
572
573
574
575
        if attention_type == "gated":
            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]
            self.position_net = PositionNet(positive_len=positive_len, out_dim=cross_attention_dim)

576
    @property
Patrick von Platen's avatar
Patrick von Platen committed
577
    def attn_processors(self) -> Dict[str, AttentionProcessor]:
578
579
580
581
582
        r"""
        Returns:
            `dict` of attention processors: A dictionary containing all attention processors used in the model with
            indexed by its weight name.
        """
583
        # set recursively
584
585
        processors = {}

Patrick von Platen's avatar
Patrick von Platen committed
586
        def fn_recursive_add_processors(name: str, module: torch.nn.Module, processors: Dict[str, AttentionProcessor]):
587
588
            if hasattr(module, "get_processor"):
                processors[f"{name}.processor"] = module.get_processor(return_deprecated_lora=True)
589
590
591
592
593
594
595
596
597
598
599

            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

Patrick von Platen's avatar
Patrick von Platen committed
600
    def set_attn_processor(self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]]):
601
        r"""
Steven Liu's avatar
Steven Liu committed
602
603
        Sets the attention processor to use to compute attention.

604
        Parameters:
Steven Liu's avatar
Steven Liu committed
605
            processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`):
606
                The instantiated processor class or a dictionary of processor classes that will be set as the processor
Steven Liu's avatar
Steven Liu committed
607
608
609
610
                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.
611
612
613
614
615
616
617
618
619
620
621

        """
        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):
622
            if hasattr(module, "set_processor"):
623
624
625
626
                if not isinstance(processor, dict):
                    module.set_processor(processor)
                else:
                    module.set_processor(processor.pop(f"{name}.processor"))
627

628
629
            for sub_name, child in module.named_children():
                fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)
630

631
632
        for name, module in self.named_children():
            fn_recursive_attn_processor(name, module, processor)
633

634
635
636
637
638
639
    def set_default_attn_processor(self):
        """
        Disables custom attention processors and sets the default attention implementation.
        """
        self.set_attn_processor(AttnProcessor())

640
    def set_attention_slice(self, slice_size):
641
642
        r"""
        Enable sliced attention computation.
643

Steven Liu's avatar
Steven Liu committed
644
645
        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.
646

647
648
        Args:
            slice_size (`str` or `int` or `list(int)`, *optional*, defaults to `"auto"`):
Steven Liu's avatar
Steven Liu committed
649
650
651
652
                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`.
653
654
655
        """
        sliceable_head_dims = []

Alexander Pivovarov's avatar
Alexander Pivovarov committed
656
        def fn_recursive_retrieve_sliceable_dims(module: torch.nn.Module):
657
658
659
660
            if hasattr(module, "set_attention_slice"):
                sliceable_head_dims.append(module.sliceable_head_dim)

            for child in module.children():
Alexander Pivovarov's avatar
Alexander Pivovarov committed
661
                fn_recursive_retrieve_sliceable_dims(child)
662
663
664

        # retrieve number of attention layers
        for module in self.children():
Alexander Pivovarov's avatar
Alexander Pivovarov committed
665
            fn_recursive_retrieve_sliceable_dims(module)
666

Alexander Pivovarov's avatar
Alexander Pivovarov committed
667
        num_sliceable_layers = len(sliceable_head_dims)
668
669
670
671
672
673
674

        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
Alexander Pivovarov's avatar
Alexander Pivovarov committed
675
            slice_size = num_sliceable_layers * [1]
676

Alexander Pivovarov's avatar
Alexander Pivovarov committed
677
        slice_size = num_sliceable_layers * [slice_size] if not isinstance(slice_size, list) else slice_size
678
679
680
681
682
683

        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)}."
            )
684

685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
        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)
704

705
    def _set_gradient_checkpointing(self, module, value=False):
706
        if hasattr(module, "gradient_checkpointing"):
707
708
            module.gradient_checkpointing = value

Patrick von Platen's avatar
Patrick von Platen committed
709
710
711
712
713
    def forward(
        self,
        sample: torch.FloatTensor,
        timestep: Union[torch.Tensor, float, int],
        encoder_hidden_states: torch.Tensor,
714
        class_labels: Optional[torch.Tensor] = None,
715
        timestep_cond: Optional[torch.Tensor] = None,
Will Berman's avatar
Will Berman committed
716
        attention_mask: Optional[torch.Tensor] = None,
717
        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
YiYi Xu's avatar
YiYi Xu committed
718
        added_cond_kwargs: Optional[Dict[str, torch.Tensor]] = None,
719
720
        down_block_additional_residuals: Optional[Tuple[torch.Tensor]] = None,
        mid_block_additional_residual: Optional[torch.Tensor] = None,
721
        encoder_attention_mask: Optional[torch.Tensor] = None,
722
723
        return_dict: bool = True,
    ) -> Union[UNet2DConditionOutput, Tuple]:
724
        r"""
Steven Liu's avatar
Steven Liu committed
725
726
        The [`UNet2DConditionModel`] forward method.

Kashif Rasul's avatar
Kashif Rasul committed
727
        Args:
Steven Liu's avatar
Steven Liu committed
728
729
730
731
732
            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)`.
733
            encoder_attention_mask (`torch.Tensor`):
Steven Liu's avatar
Steven Liu committed
734
735
736
                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.
Kashif Rasul's avatar
Kashif Rasul committed
737
            return_dict (`bool`, *optional*, defaults to `True`):
Steven Liu's avatar
Steven Liu committed
738
739
                Whether or not to return a [`~models.unet_2d_condition.UNet2DConditionOutput`] instead of a plain
                tuple.
740
            cross_attention_kwargs (`dict`, *optional*):
Steven Liu's avatar
Steven Liu committed
741
                A kwargs dictionary that if specified is passed along to the [`AttnProcessor`].
742
743
744
            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.
Kashif Rasul's avatar
Kashif Rasul committed
745
746
747

        Returns:
            [`~models.unet_2d_condition.UNet2DConditionOutput`] or `tuple`:
Steven Liu's avatar
Steven Liu committed
748
749
                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.
Kashif Rasul's avatar
Kashif Rasul committed
750
        """
751
        # By default samples have to be AT least a multiple of the overall upsampling factor.
Alexander Pivovarov's avatar
Alexander Pivovarov committed
752
        # The overall upsampling factor is equal to 2 ** (# num of upsampling layers).
753
754
755
756
757
758
759
760
761
762
763
764
        # 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

        if any(s % default_overall_up_factor != 0 for s in sample.shape[-2:]):
            logger.info("Forward upsample size to force interpolation output size.")
            forward_upsample_size = True

765
766
767
768
769
770
771
772
        # 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)
Will Berman's avatar
Will Berman committed
773
        if attention_mask is not None:
774
775
776
777
            # 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)
Will Berman's avatar
Will Berman committed
778
779
780
            attention_mask = (1 - attention_mask.to(sample.dtype)) * -10000.0
            attention_mask = attention_mask.unsqueeze(1)

781
782
783
784
785
        # 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)

Patrick von Platen's avatar
Patrick von Platen committed
786
787
788
789
790
791
792
        # 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):
793
            # TODO: this requires sync between CPU and GPU. So try to pass timesteps as tensors if you can
794
795
            # This would be a good case for the `match` statement (Python 3.10+)
            is_mps = sample.device.type == "mps"
Patrick von Platen's avatar
Patrick von Platen committed
796
            if isinstance(timestep, float):
797
798
799
800
801
                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:
802
            timesteps = timesteps[None].to(sample.device)
Patrick von Platen's avatar
Patrick von Platen committed
803

804
        # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
805
        timesteps = timesteps.expand(sample.shape[0])
806

Patrick von Platen's avatar
Patrick von Platen committed
807
        t_emb = self.time_proj(timesteps)
808

809
        # `Timesteps` does not contain any weights and will always return f32 tensors
810
811
        # but time_embedding might actually be running in fp16. so we need to cast here.
        # there might be better ways to encapsulate this.
812
        t_emb = t_emb.to(dtype=sample.dtype)
813
814

        emb = self.time_embedding(t_emb, timestep_cond)
815
        aug_emb = None
Patrick von Platen's avatar
Patrick von Platen committed
816

Will Berman's avatar
Will Berman committed
817
        if self.class_embedding is not None:
818
819
            if class_labels is None:
                raise ValueError("class_labels should be provided when num_class_embeds > 0")
Will Berman's avatar
Will Berman committed
820
821
822
823

            if self.config.class_embed_type == "timestep":
                class_labels = self.time_proj(class_labels)

824
825
826
827
                # `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)

828
            class_emb = self.class_embedding(class_labels).to(dtype=sample.dtype)
Sanchit Gandhi's avatar
Sanchit Gandhi committed
829
830
831
832
833

            if self.config.class_embeddings_concat:
                emb = torch.cat([emb, class_emb], dim=-1)
            else:
                emb = emb + class_emb
834

Patrick von Platen's avatar
Patrick von Platen committed
835
836
        if self.config.addition_embed_type == "text":
            aug_emb = self.add_embedding(encoder_hidden_states)
YiYi Xu's avatar
YiYi Xu committed
837
        elif self.config.addition_embed_type == "text_image":
YiYi Xu's avatar
YiYi Xu committed
838
            # Kandinsky 2.1 - style
YiYi Xu's avatar
YiYi Xu committed
839
840
841
842
843
844
845
846
            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)
847
        elif self.config.addition_embed_type == "text_time":
848
            # SDXL - style
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
            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)
YiYi Xu's avatar
YiYi Xu committed
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
        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)
883
884

        emb = emb + aug_emb if aug_emb is not None else emb
Patrick von Platen's avatar
Patrick von Platen committed
885

886
887
888
        if self.time_embed_act is not None:
            emb = self.time_embed_act(emb)

YiYi Xu's avatar
YiYi Xu committed
889
        if self.encoder_hid_proj is not None and self.config.encoder_hid_dim_type == "text_proj":
William Berman's avatar
William Berman committed
890
            encoder_hidden_states = self.encoder_hid_proj(encoder_hidden_states)
YiYi Xu's avatar
YiYi Xu committed
891
892
893
894
895
896
897
898
899
        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)
YiYi Xu's avatar
YiYi Xu committed
900
901
902
903
904
905
906
907
        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)
Patrick von Platen's avatar
Patrick von Platen committed
908
909
910
        # 2. pre-process
        sample = self.conv_in(sample)

911
912
913
914
915
916
        # 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)}

Patrick von Platen's avatar
Patrick von Platen committed
917
        # 3. down
Will Berman's avatar
Will Berman committed
918
919
920
921

        is_controlnet = mid_block_additional_residual is not None and down_block_additional_residuals is not None
        is_adapter = mid_block_additional_residual is None and down_block_additional_residuals is not None

Patrick von Platen's avatar
Patrick von Platen committed
922
923
        down_block_res_samples = (sample,)
        for downsample_block in self.down_blocks:
924
            if hasattr(downsample_block, "has_cross_attention") and downsample_block.has_cross_attention:
Will Berman's avatar
Will Berman committed
925
926
927
928
929
                # For t2i-adapter CrossAttnDownBlock2D
                additional_residuals = {}
                if is_adapter and len(down_block_additional_residuals) > 0:
                    additional_residuals["additional_residuals"] = down_block_additional_residuals.pop(0)

Patrick von Platen's avatar
Patrick von Platen committed
930
                sample, res_samples = downsample_block(
931
932
933
                    hidden_states=sample,
                    temb=emb,
                    encoder_hidden_states=encoder_hidden_states,
Will Berman's avatar
Will Berman committed
934
                    attention_mask=attention_mask,
935
                    cross_attention_kwargs=cross_attention_kwargs,
936
                    encoder_attention_mask=encoder_attention_mask,
Will Berman's avatar
Will Berman committed
937
                    **additional_residuals,
Patrick von Platen's avatar
Patrick von Platen committed
938
939
940
941
                )
            else:
                sample, res_samples = downsample_block(hidden_states=sample, temb=emb)

Will Berman's avatar
Will Berman committed
942
943
944
                if is_adapter and len(down_block_additional_residuals) > 0:
                    sample += down_block_additional_residuals.pop(0)

Patrick von Platen's avatar
Patrick von Platen committed
945
946
            down_block_res_samples += res_samples

Will Berman's avatar
Will Berman committed
947
        if is_controlnet:
948
949
950
951
952
            new_down_block_res_samples = ()

            for down_block_res_sample, down_block_additional_residual in zip(
                down_block_res_samples, down_block_additional_residuals
            ):
953
                down_block_res_sample = down_block_res_sample + down_block_additional_residual
954
                new_down_block_res_samples = new_down_block_res_samples + (down_block_res_sample,)
955
956
957

            down_block_res_samples = new_down_block_res_samples

Patrick von Platen's avatar
Patrick von Platen committed
958
        # 4. mid
959
960
961
962
963
964
965
        if self.mid_block is not None:
            sample = self.mid_block(
                sample,
                emb,
                encoder_hidden_states=encoder_hidden_states,
                attention_mask=attention_mask,
                cross_attention_kwargs=cross_attention_kwargs,
966
                encoder_attention_mask=encoder_attention_mask,
967
            )
968
969
970
971
972
973
974
            # To support T2I-Adapter-XL
            if (
                is_adapter
                and len(down_block_additional_residuals) > 0
                and sample.shape == down_block_additional_residuals[0].shape
            ):
                sample += down_block_additional_residuals.pop(0)
Patrick von Platen's avatar
Patrick von Platen committed
975

Will Berman's avatar
Will Berman committed
976
        if is_controlnet:
977
            sample = sample + mid_block_additional_residual
978

Patrick von Platen's avatar
Patrick von Platen committed
979
        # 5. up
980
981
982
        for i, upsample_block in enumerate(self.up_blocks):
            is_final_block = i == len(self.up_blocks) - 1

Patrick von Platen's avatar
Patrick von Platen committed
983
984
985
            res_samples = down_block_res_samples[-len(upsample_block.resnets) :]
            down_block_res_samples = down_block_res_samples[: -len(upsample_block.resnets)]

986
987
988
989
990
            # 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:]

991
            if hasattr(upsample_block, "has_cross_attention") and upsample_block.has_cross_attention:
Patrick von Platen's avatar
Patrick von Platen committed
992
993
994
995
996
                sample = upsample_block(
                    hidden_states=sample,
                    temb=emb,
                    res_hidden_states_tuple=res_samples,
                    encoder_hidden_states=encoder_hidden_states,
997
                    cross_attention_kwargs=cross_attention_kwargs,
998
                    upsample_size=upsample_size,
Will Berman's avatar
Will Berman committed
999
                    attention_mask=attention_mask,
1000
                    encoder_attention_mask=encoder_attention_mask,
Patrick von Platen's avatar
Patrick von Platen committed
1001
1002
                )
            else:
1003
1004
1005
                sample = upsample_block(
                    hidden_states=sample, temb=emb, res_hidden_states_tuple=res_samples, upsample_size=upsample_size
                )
1006

Patrick von Platen's avatar
Patrick von Platen committed
1007
        # 6. post-process
1008
1009
1010
        if self.conv_norm_out:
            sample = self.conv_norm_out(sample)
            sample = self.conv_act(sample)
Patrick von Platen's avatar
Patrick von Platen committed
1011
1012
        sample = self.conv_out(sample)

1013
1014
        if not return_dict:
            return (sample,)
Patrick von Platen's avatar
Patrick von Platen committed
1015

1016
        return UNet2DConditionOutput(sample=sample)