unet_2d_condition.py 44.3 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
28
29
30
31
32
33
from .embeddings import (
    GaussianFourierProjection,
    TextImageProjection,
    TextImageTimeEmbedding,
    TextTimeEmbedding,
    TimestepEmbedding,
    Timesteps,
)
34
from .modeling_utils import ModelMixin
35
from .unet_2d_blocks import (
36
37
38
39
    CrossAttnDownBlock2D,
    CrossAttnUpBlock2D,
    DownBlock2D,
    UNetMidBlock2DCrossAttn,
Will Berman's avatar
Will Berman committed
40
    UNetMidBlock2DSimpleCrossAttn,
41
42
43
44
    UpBlock2D,
    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
53
54
55
56
57
58
59
60
@dataclass
class UNet2DConditionOutput(BaseOutput):
    """
    Args:
        sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`):
            Hidden states conditioned on `encoder_hidden_states` input. Output of last layer of model.
    """

    sample: torch.FloatTensor


61
class UNet2DConditionModel(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin):
Kashif Rasul's avatar
Kashif Rasul committed
62
63
64
65
66
    r"""
    UNet2DConditionModel is a conditional 2D UNet model that takes in a noisy sample, conditional state, and a timestep
    and returns sample shaped output.

    This model inherits from [`ModelMixin`]. Check the superclass documentation for the generic methods the library
67
    implements for all the models (such as downloading or saving, etc.)
Kashif Rasul's avatar
Kashif Rasul committed
68
69

    Parameters:
70
71
        sample_size (`int` or `Tuple[int, int]`, *optional*, defaults to `None`):
            Height and width of input/output sample.
Kashif Rasul's avatar
Kashif Rasul committed
72
73
74
        in_channels (`int`, *optional*, defaults to 4): The number of channels in the input sample.
        out_channels (`int`, *optional*, defaults to 4): The number of channels in the output.
        center_input_sample (`bool`, *optional*, defaults to `False`): Whether to center the input sample.
Suraj Patil's avatar
Suraj Patil committed
75
        flip_sin_to_cos (`bool`, *optional*, defaults to `False`):
Kashif Rasul's avatar
Kashif Rasul committed
76
77
78
79
            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
80
        mid_block_type (`str`, *optional*, defaults to `"UNetMidBlock2DCrossAttn"`):
81
82
            The mid block type. Choose from `UNetMidBlock2DCrossAttn` or `UNetMidBlock2DSimpleCrossAttn`, will skip the
            mid block layer if `None`.
Kashif Rasul's avatar
Kashif Rasul committed
83
84
        up_block_types (`Tuple[str]`, *optional*, defaults to `("UpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D",)`):
            The tuple of upsample blocks to use.
85
86
87
        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
88
89
90
91
92
93
94
        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.
95
            If `None`, it will skip the normalization and activation layers in post-processing
Kashif Rasul's avatar
Kashif Rasul committed
96
        norm_eps (`float`, *optional*, defaults to 1e-5): The epsilon to use for the normalization.
Sanchit Gandhi's avatar
Sanchit Gandhi committed
97
98
        cross_attention_dim (`int` or `Tuple[int]`, *optional*, defaults to 1280):
            The dimension of the cross attention features.
William Berman's avatar
William Berman committed
99
        encoder_hid_dim (`int`, *optional*, defaults to None):
YiYi Xu's avatar
YiYi Xu committed
100
101
102
103
104
            If `encoder_hid_dim_type` is defined, `encoder_hidden_states` will be projected from `encoder_hid_dim`
            dimension to `cross_attention_dim`.
        encoder_hid_dim_type (`str`, *optional*, defaults to None):
            If given, the `encoder_hidden_states` and potentially other embeddings will be down-projected to text
            embeddings of dimension `cross_attention` according to `encoder_hid_dim_type`.
Kashif Rasul's avatar
Kashif Rasul committed
105
        attention_head_dim (`int`, *optional*, defaults to 8): The dimension of the attention heads.
106
107
        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
108
109
        resnet_time_scale_shift (`str`, *optional*, defaults to `"default"`): Time scale shift config
            for resnet blocks, see [`~models.resnet.ResnetBlock2D`]. Choose from `default` or `scale_shift`.
110
111
        class_embed_type (`str`, *optional*, defaults to None):
            The type of class embedding to use which is ultimately summed with the time embeddings. Choose from `None`,
Sanchit Gandhi's avatar
Sanchit Gandhi committed
112
            `"timestep"`, `"identity"`, `"projection"`, or `"simple_projection"`.
Patrick von Platen's avatar
Patrick von Platen committed
113
114
115
        addition_embed_type (`str`, *optional*, defaults to None):
            Configures an optional embedding which will be summed with the time embeddings. Choose from `None` or
            "text". "text" will use the `TextTimeEmbedding` layer.
116
117
118
        num_class_embeds (`int`, *optional*, defaults to None):
            Input dimension of the learnable embedding matrix to be projected to `time_embed_dim`, when performing
            class conditioning with `class_embed_type` equal to `None`.
119
120
        time_embedding_type (`str`, *optional*, default to `positional`):
            The type of position embedding to use for timesteps. Choose from `positional` or `fourier`.
Patrick von Platen's avatar
Patrick von Platen committed
121
122
        time_embedding_dim (`int`, *optional*, default to `None`):
            An optional override for the dimension of the projected time embedding.
123
124
125
        time_embedding_act_fn (`str`, *optional*, default to `None`):
            Optional activation function to use on the time embeddings only one time before they as passed to the rest
            of the unet. Choose from `silu`, `mish`, `gelu`, and `swish`.
126
127
128
129
130
        timestep_post_act (`str, *optional*, default to `None`):
            The second activation function to use in timestep embedding. Choose from `silu`, `mish` and `gelu`.
        time_cond_proj_dim (`int`, *optional*, default to `None`):
            The dimension of `cond_proj` layer in timestep embedding.
        conv_in_kernel (`int`, *optional*, default to `3`): The kernel size of `conv_in` layer.
Will Berman's avatar
Will Berman committed
131
132
133
        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
            using the "projection" `class_embed_type`. Required when using the "projection" `class_embed_type`.
Sanchit Gandhi's avatar
Sanchit Gandhi committed
134
        class_embeddings_concat (`bool`, *optional*, defaults to `False`): Whether to concatenate the time
135
136
137
138
139
140
            embeddings with the class embeddings.
        mid_block_only_cross_attention (`bool`, *optional*, defaults to `None`):
            Whether to use cross attention with the mid block when using the `UNetMidBlock2DSimpleCrossAttn`. If
            `only_cross_attention` is given as a single boolean and `mid_block_only_cross_attention` is None, the
            `only_cross_attention` value will be used as the value for `mid_block_only_cross_attention`. Else, it will
            default to `False`.
Kashif Rasul's avatar
Kashif Rasul committed
141
142
    """

143
144
    _supports_gradient_checkpointing = True

Patrick von Platen's avatar
Patrick von Platen committed
145
146
147
    @register_to_config
    def __init__(
        self,
Sid Sahai's avatar
Sid Sahai committed
148
149
150
151
152
153
154
155
156
157
158
159
        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",
        ),
160
        mid_block_type: Optional[str] = "UNetMidBlock2DCrossAttn",
Sid Sahai's avatar
Sid Sahai committed
161
        up_block_types: Tuple[str] = ("UpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D"),
162
        only_cross_attention: Union[bool, Tuple[bool]] = False,
Sid Sahai's avatar
Sid Sahai committed
163
        block_out_channels: Tuple[int] = (320, 640, 1280, 1280),
164
        layers_per_block: Union[int, Tuple[int]] = 2,
Sid Sahai's avatar
Sid Sahai committed
165
166
167
        downsample_padding: int = 1,
        mid_block_scale_factor: float = 1,
        act_fn: str = "silu",
168
        norm_num_groups: Optional[int] = 32,
Sid Sahai's avatar
Sid Sahai committed
169
        norm_eps: float = 1e-5,
Sanchit Gandhi's avatar
Sanchit Gandhi committed
170
        cross_attention_dim: Union[int, Tuple[int]] = 1280,
William Berman's avatar
William Berman committed
171
        encoder_hid_dim: Optional[int] = None,
YiYi Xu's avatar
YiYi Xu committed
172
        encoder_hid_dim_type: Optional[str] = None,
Suraj Patil's avatar
Suraj Patil committed
173
        attention_head_dim: Union[int, Tuple[int]] = 8,
174
        num_attention_heads: Optional[Union[int, Tuple[int]]] = None,
175
        dual_cross_attention: bool = False,
Suraj Patil's avatar
Suraj Patil committed
176
        use_linear_projection: bool = False,
Will Berman's avatar
Will Berman committed
177
        class_embed_type: Optional[str] = None,
Patrick von Platen's avatar
Patrick von Platen committed
178
        addition_embed_type: Optional[str] = None,
179
        num_class_embeds: Optional[int] = None,
180
        upcast_attention: bool = False,
Will Berman's avatar
Will Berman committed
181
        resnet_time_scale_shift: str = "default",
182
183
        resnet_skip_time_act: bool = False,
        resnet_out_scale_factor: int = 1.0,
184
        time_embedding_type: str = "positional",
Patrick von Platen's avatar
Patrick von Platen committed
185
        time_embedding_dim: Optional[int] = None,
186
        time_embedding_act_fn: Optional[str] = None,
187
188
189
190
        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
191
        projection_class_embeddings_input_dim: Optional[int] = None,
Sanchit Gandhi's avatar
Sanchit Gandhi committed
192
        class_embeddings_concat: bool = False,
193
        mid_block_only_cross_attention: Optional[bool] = None,
194
        cross_attention_norm: Optional[str] = None,
Patrick von Platen's avatar
Patrick von Platen committed
195
        addition_embed_type_num_heads=64,
Patrick von Platen's avatar
Patrick von Platen committed
196
197
198
199
200
    ):
        super().__init__()

        self.sample_size = sample_size

201
202
203
204
205
206
207
208
        # 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
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
        # 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}."
            )

225
226
227
228
229
        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
230
231
232
233
234
        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
235
236
237
238
239
        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}."
            )

240
241
242
243
244
        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
245
        # input
246
247
248
249
        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
250
251

        # time
252
        if time_embedding_type == "fourier":
Patrick von Platen's avatar
Patrick von Platen committed
253
            time_embed_dim = time_embedding_dim or block_out_channels[0] * 2
254
255
256
257
258
259
260
            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
261
            time_embed_dim = time_embedding_dim or block_out_channels[0] * 4
262
263
264
265
266

            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
267
                f"{time_embedding_type} does not exist. Please make sure to use one of `fourier` or `positional`."
268
            )
Patrick von Platen's avatar
Patrick von Platen committed
269

270
271
272
273
274
275
276
        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
277

YiYi Xu's avatar
YiYi Xu committed
278
279
        if encoder_hid_dim_type is None and encoder_hid_dim is not None:
            encoder_hid_dim_type = "text_proj"
280
            self.register_to_config(encoder_hid_dim_type=encoder_hid_dim_type)
YiYi Xu's avatar
YiYi Xu committed
281
282
283
284
285
286
287
288
            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
289
            self.encoder_hid_proj = nn.Linear(encoder_hid_dim, cross_attention_dim)
YiYi Xu's avatar
YiYi Xu committed
290
291
292
293
294
295
296
297
298
299
300
301
302
303
        elif encoder_hid_dim_type == "text_image_proj":
            # image_embed_dim DOESN'T have to be `cross_attention_dim`. To not clutter the __init__ too much
            # they are set to `cross_attention_dim` here as this is exactly the required dimension for the currently only use
            # case when `addition_embed_type == "text_image_proj"` (Kadinsky 2.1)`
            self.encoder_hid_proj = TextImageProjection(
                text_embed_dim=encoder_hid_dim,
                image_embed_dim=cross_attention_dim,
                cross_attention_dim=cross_attention_dim,
            )

        elif encoder_hid_dim_type 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
304
305
306
        else:
            self.encoder_hid_proj = None

307
        # class embedding
Will Berman's avatar
Will Berman committed
308
        if class_embed_type is None and num_class_embeds is not None:
309
            self.class_embedding = nn.Embedding(num_class_embeds, time_embed_dim)
Will Berman's avatar
Will Berman committed
310
        elif class_embed_type == "timestep":
311
            self.class_embedding = TimestepEmbedding(timestep_input_dim, time_embed_dim, act_fn=act_fn)
Will Berman's avatar
Will Berman committed
312
313
        elif class_embed_type == "identity":
            self.class_embedding = nn.Identity(time_embed_dim, time_embed_dim)
Will Berman's avatar
Will Berman committed
314
315
316
317
318
319
320
321
322
323
324
325
326
        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
327
328
329
330
331
332
        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
333
334
        else:
            self.class_embedding = None
335

Patrick von Platen's avatar
Patrick von Platen committed
336
337
338
339
340
341
342
343
344
        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
345
346
347
348
349
350
351
        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
            )
Patrick von Platen's avatar
Patrick von Platen committed
352
        elif addition_embed_type is not None:
YiYi Xu's avatar
YiYi Xu committed
353
            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
354

355
356
357
        if time_embedding_act_fn is None:
            self.time_embed_act = None
        else:
358
            self.time_embed_act = get_activation(time_embedding_act_fn)
359

Patrick von Platen's avatar
Patrick von Platen committed
360
361
362
        self.down_blocks = nn.ModuleList([])
        self.up_blocks = nn.ModuleList([])

363
        if isinstance(only_cross_attention, bool):
364
365
366
            if mid_block_only_cross_attention is None:
                mid_block_only_cross_attention = only_cross_attention

367
368
            only_cross_attention = [only_cross_attention] * len(down_block_types)

369
370
371
        if mid_block_only_cross_attention is None:
            mid_block_only_cross_attention = False

372
373
374
        if isinstance(num_attention_heads, int):
            num_attention_heads = (num_attention_heads,) * len(down_block_types)

Suraj Patil's avatar
Suraj Patil committed
375
376
377
        if isinstance(attention_head_dim, int):
            attention_head_dim = (attention_head_dim,) * len(down_block_types)

Sanchit Gandhi's avatar
Sanchit Gandhi committed
378
379
380
        if isinstance(cross_attention_dim, int):
            cross_attention_dim = (cross_attention_dim,) * len(down_block_types)

381
382
383
        if isinstance(layers_per_block, int):
            layers_per_block = [layers_per_block] * len(down_block_types)

Sanchit Gandhi's avatar
Sanchit Gandhi committed
384
385
386
387
388
389
390
391
        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
392
393
394
395
396
397
398
399
400
        # 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,
401
                num_layers=layers_per_block[i],
Patrick von Platen's avatar
Patrick von Platen committed
402
403
                in_channels=input_channel,
                out_channels=output_channel,
Sanchit Gandhi's avatar
Sanchit Gandhi committed
404
                temb_channels=blocks_time_embed_dim,
Patrick von Platen's avatar
Patrick von Platen committed
405
406
407
                add_downsample=not is_final_block,
                resnet_eps=norm_eps,
                resnet_act_fn=act_fn,
408
                resnet_groups=norm_num_groups,
Sanchit Gandhi's avatar
Sanchit Gandhi committed
409
                cross_attention_dim=cross_attention_dim[i],
410
                num_attention_heads=num_attention_heads[i],
Patrick von Platen's avatar
Patrick von Platen committed
411
                downsample_padding=downsample_padding,
412
                dual_cross_attention=dual_cross_attention,
Suraj Patil's avatar
Suraj Patil committed
413
                use_linear_projection=use_linear_projection,
414
                only_cross_attention=only_cross_attention[i],
415
                upcast_attention=upcast_attention,
Will Berman's avatar
Will Berman committed
416
                resnet_time_scale_shift=resnet_time_scale_shift,
417
418
                resnet_skip_time_act=resnet_skip_time_act,
                resnet_out_scale_factor=resnet_out_scale_factor,
419
                cross_attention_norm=cross_attention_norm,
420
                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
421
422
423
424
            )
            self.down_blocks.append(down_block)

        # mid
Will Berman's avatar
Will Berman committed
425
426
427
        if mid_block_type == "UNetMidBlock2DCrossAttn":
            self.mid_block = UNetMidBlock2DCrossAttn(
                in_channels=block_out_channels[-1],
Sanchit Gandhi's avatar
Sanchit Gandhi committed
428
                temb_channels=blocks_time_embed_dim,
Will Berman's avatar
Will Berman committed
429
430
431
432
                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
433
                cross_attention_dim=cross_attention_dim[-1],
434
                num_attention_heads=num_attention_heads[-1],
Will Berman's avatar
Will Berman committed
435
436
437
438
439
440
441
442
                resnet_groups=norm_num_groups,
                dual_cross_attention=dual_cross_attention,
                use_linear_projection=use_linear_projection,
                upcast_attention=upcast_attention,
            )
        elif mid_block_type == "UNetMidBlock2DSimpleCrossAttn":
            self.mid_block = UNetMidBlock2DSimpleCrossAttn(
                in_channels=block_out_channels[-1],
Sanchit Gandhi's avatar
Sanchit Gandhi committed
443
                temb_channels=blocks_time_embed_dim,
Will Berman's avatar
Will Berman committed
444
445
446
                resnet_eps=norm_eps,
                resnet_act_fn=act_fn,
                output_scale_factor=mid_block_scale_factor,
Sanchit Gandhi's avatar
Sanchit Gandhi committed
447
                cross_attention_dim=cross_attention_dim[-1],
448
                attention_head_dim=attention_head_dim[-1],
Will Berman's avatar
Will Berman committed
449
450
                resnet_groups=norm_num_groups,
                resnet_time_scale_shift=resnet_time_scale_shift,
451
                skip_time_act=resnet_skip_time_act,
452
                only_cross_attention=mid_block_only_cross_attention,
453
                cross_attention_norm=cross_attention_norm,
Will Berman's avatar
Will Berman committed
454
            )
455
456
        elif mid_block_type is None:
            self.mid_block = None
Will Berman's avatar
Will Berman committed
457
458
        else:
            raise ValueError(f"unknown mid_block_type : {mid_block_type}")
Patrick von Platen's avatar
Patrick von Platen committed
459

460
461
462
        # count how many layers upsample the images
        self.num_upsamplers = 0

Patrick von Platen's avatar
Patrick von Platen committed
463
464
        # up
        reversed_block_out_channels = list(reversed(block_out_channels))
465
        reversed_num_attention_heads = list(reversed(num_attention_heads))
466
        reversed_layers_per_block = list(reversed(layers_per_block))
Sanchit Gandhi's avatar
Sanchit Gandhi committed
467
        reversed_cross_attention_dim = list(reversed(cross_attention_dim))
468
        only_cross_attention = list(reversed(only_cross_attention))
469

Patrick von Platen's avatar
Patrick von Platen committed
470
471
        output_channel = reversed_block_out_channels[0]
        for i, up_block_type in enumerate(up_block_types):
472
473
            is_final_block = i == len(block_out_channels) - 1

Patrick von Platen's avatar
Patrick von Platen committed
474
475
476
477
            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)]

478
479
480
481
482
483
            # 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
484
485
486

            up_block = get_up_block(
                up_block_type,
487
                num_layers=reversed_layers_per_block[i] + 1,
Patrick von Platen's avatar
Patrick von Platen committed
488
489
490
                in_channels=input_channel,
                out_channels=output_channel,
                prev_output_channel=prev_output_channel,
Sanchit Gandhi's avatar
Sanchit Gandhi committed
491
                temb_channels=blocks_time_embed_dim,
492
                add_upsample=add_upsample,
Patrick von Platen's avatar
Patrick von Platen committed
493
494
                resnet_eps=norm_eps,
                resnet_act_fn=act_fn,
495
                resnet_groups=norm_num_groups,
Sanchit Gandhi's avatar
Sanchit Gandhi committed
496
                cross_attention_dim=reversed_cross_attention_dim[i],
497
                num_attention_heads=reversed_num_attention_heads[i],
498
                dual_cross_attention=dual_cross_attention,
Suraj Patil's avatar
Suraj Patil committed
499
                use_linear_projection=use_linear_projection,
500
                only_cross_attention=only_cross_attention[i],
501
                upcast_attention=upcast_attention,
Will Berman's avatar
Will Berman committed
502
                resnet_time_scale_shift=resnet_time_scale_shift,
503
504
                resnet_skip_time_act=resnet_skip_time_act,
                resnet_out_scale_factor=resnet_out_scale_factor,
505
                cross_attention_norm=cross_attention_norm,
506
                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
507
508
509
510
511
            )
            self.up_blocks.append(up_block)
            prev_output_channel = output_channel

        # out
512
513
514
515
        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
            )
516

517
            self.conv_act = get_activation(act_fn)
518

519
520
521
522
523
524
525
526
        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
527

528
    @property
Patrick von Platen's avatar
Patrick von Platen committed
529
    def attn_processors(self) -> Dict[str, AttentionProcessor]:
530
531
532
533
534
        r"""
        Returns:
            `dict` of attention processors: A dictionary containing all attention processors used in the model with
            indexed by its weight name.
        """
535
        # set recursively
536
537
        processors = {}

Patrick von Platen's avatar
Patrick von Platen committed
538
        def fn_recursive_add_processors(name: str, module: torch.nn.Module, processors: Dict[str, AttentionProcessor]):
539
540
541
542
543
544
545
546
547
548
549
550
551
            if hasattr(module, "set_processor"):
                processors[f"{name}.processor"] = module.processor

            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
552
    def set_attn_processor(self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]]):
553
554
        r"""
        Parameters:
Patrick von Platen's avatar
Patrick von Platen committed
555
            `processor (`dict` of `AttentionProcessor` or `AttentionProcessor`):
556
                The instantiated processor class or a dictionary of processor classes that will be set as the processor
Patrick von Platen's avatar
Patrick von Platen committed
557
                of **all** `Attention` layers.
Alexander Pivovarov's avatar
Alexander Pivovarov committed
558
            In case `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.:
559
560
561
562
563
564
565
566
567
568
569

        """
        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):
570
            if hasattr(module, "set_processor"):
571
572
573
574
                if not isinstance(processor, dict):
                    module.set_processor(processor)
                else:
                    module.set_processor(processor.pop(f"{name}.processor"))
575

576
577
            for sub_name, child in module.named_children():
                fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)
578

579
580
        for name, module in self.named_children():
            fn_recursive_attn_processor(name, module, processor)
581

582
583
584
585
586
587
    def set_default_attn_processor(self):
        """
        Disables custom attention processors and sets the default attention implementation.
        """
        self.set_attn_processor(AttnProcessor())

588
    def set_attention_slice(self, slice_size):
589
590
        r"""
        Enable sliced attention computation.
591

592
593
        When this option is enabled, the attention module will split the input tensor in slices, to compute attention
        in several steps. This is useful to save some memory in exchange for a small speed decrease.
594

595
596
597
        Args:
            slice_size (`str` or `int` or `list(int)`, *optional*, defaults to `"auto"`):
                When `"auto"`, halves the input to the attention heads, so attention will be computed in two steps. If
Alexander Pivovarov's avatar
Alexander Pivovarov committed
598
                `"max"`, maximum amount of memory will be saved by running only one slice at a time. If a number is
599
600
                provided, uses as many slices as `num_attention_heads // slice_size`. In this case,
                `num_attention_heads` must be a multiple of `slice_size`.
601
602
603
        """
        sliceable_head_dims = []

Alexander Pivovarov's avatar
Alexander Pivovarov committed
604
        def fn_recursive_retrieve_sliceable_dims(module: torch.nn.Module):
605
606
607
608
            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
609
                fn_recursive_retrieve_sliceable_dims(child)
610
611
612

        # retrieve number of attention layers
        for module in self.children():
Alexander Pivovarov's avatar
Alexander Pivovarov committed
613
            fn_recursive_retrieve_sliceable_dims(module)
614

Alexander Pivovarov's avatar
Alexander Pivovarov committed
615
        num_sliceable_layers = len(sliceable_head_dims)
616
617
618
619
620
621
622

        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
623
            slice_size = num_sliceable_layers * [1]
624

Alexander Pivovarov's avatar
Alexander Pivovarov committed
625
        slice_size = num_sliceable_layers * [slice_size] if not isinstance(slice_size, list) else slice_size
626
627
628
629
630
631

        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)}."
            )
632

633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
        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)
652

653
654
655
656
    def _set_gradient_checkpointing(self, module, value=False):
        if isinstance(module, (CrossAttnDownBlock2D, DownBlock2D, CrossAttnUpBlock2D, UpBlock2D)):
            module.gradient_checkpointing = value

Patrick von Platen's avatar
Patrick von Platen committed
657
658
659
660
661
    def forward(
        self,
        sample: torch.FloatTensor,
        timestep: Union[torch.Tensor, float, int],
        encoder_hidden_states: torch.Tensor,
662
        class_labels: Optional[torch.Tensor] = None,
663
        timestep_cond: Optional[torch.Tensor] = None,
Will Berman's avatar
Will Berman committed
664
        attention_mask: Optional[torch.Tensor] = None,
665
        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
YiYi Xu's avatar
YiYi Xu committed
666
        added_cond_kwargs: Optional[Dict[str, torch.Tensor]] = None,
667
668
        down_block_additional_residuals: Optional[Tuple[torch.Tensor]] = None,
        mid_block_additional_residual: Optional[torch.Tensor] = None,
669
        encoder_attention_mask: Optional[torch.Tensor] = None,
670
671
        return_dict: bool = True,
    ) -> Union[UNet2DConditionOutput, Tuple]:
672
        r"""
Kashif Rasul's avatar
Kashif Rasul committed
673
674
        Args:
            sample (`torch.FloatTensor`): (batch, channel, height, width) noisy inputs tensor
675
            timestep (`torch.FloatTensor` or `float` or `int`): (batch) timesteps
676
            encoder_hidden_states (`torch.FloatTensor`): (batch, sequence_length, feature_dim) encoder hidden states
677
678
679
680
            encoder_attention_mask (`torch.Tensor`):
                (batch, sequence_length) cross-attention mask, applied to encoder_hidden_states. True = keep, False =
                discard. Mask will be converted into a bias, which adds large negative values to attention scores
                corresponding to "discard" tokens.
Kashif Rasul's avatar
Kashif Rasul committed
681
682
            return_dict (`bool`, *optional*, defaults to `True`):
                Whether or not to return a [`models.unet_2d_condition.UNet2DConditionOutput`] instead of a plain tuple.
683
            cross_attention_kwargs (`dict`, *optional*):
Patrick von Platen's avatar
Patrick von Platen committed
684
                A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
685
686
                `self.processor` in
                [diffusers.cross_attention](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/cross_attention.py).
YiYi Xu's avatar
YiYi Xu committed
687
688
689
690
            added_cond_kwargs (`dict`, *optional*):
                A kwargs dictionary that if specified includes additonal conditions that can be used for additonal time
                embeddings or encoder hidden states projections. See the configurations `encoder_hid_dim_type` and
                `addition_embed_type` for more information.
Kashif Rasul's avatar
Kashif Rasul committed
691
692
693
694
695
696

        Returns:
            [`~models.unet_2d_condition.UNet2DConditionOutput`] or `tuple`:
            [`~models.unet_2d_condition.UNet2DConditionOutput`] if `return_dict` is True, otherwise a `tuple`. When
            returning a tuple, the first element is the sample tensor.
        """
697
        # By default samples have to be AT least a multiple of the overall upsampling factor.
Alexander Pivovarov's avatar
Alexander Pivovarov committed
698
        # The overall upsampling factor is equal to 2 ** (# num of upsampling layers).
699
700
701
702
703
704
705
706
707
708
709
710
        # 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

711
712
713
714
715
716
717
718
        # 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
719
        if attention_mask is not None:
720
721
722
723
            # 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
724
725
726
            attention_mask = (1 - attention_mask.to(sample.dtype)) * -10000.0
            attention_mask = attention_mask.unsqueeze(1)

727
728
729
730
731
        # 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
732
733
734
735
736
737
738
        # 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):
739
            # TODO: this requires sync between CPU and GPU. So try to pass timesteps as tensors if you can
740
741
            # 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
742
            if isinstance(timestep, float):
743
744
745
746
747
                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:
748
            timesteps = timesteps[None].to(sample.device)
Patrick von Platen's avatar
Patrick von Platen committed
749

750
        # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
751
        timesteps = timesteps.expand(sample.shape[0])
752

Patrick von Platen's avatar
Patrick von Platen committed
753
        t_emb = self.time_proj(timesteps)
754

755
        # `Timesteps` does not contain any weights and will always return f32 tensors
756
757
        # but time_embedding might actually be running in fp16. so we need to cast here.
        # there might be better ways to encapsulate this.
758
        t_emb = t_emb.to(dtype=sample.dtype)
759
760

        emb = self.time_embedding(t_emb, timestep_cond)
Patrick von Platen's avatar
Patrick von Platen committed
761

Will Berman's avatar
Will Berman committed
762
        if self.class_embedding is not None:
763
764
            if class_labels is None:
                raise ValueError("class_labels should be provided when num_class_embeds > 0")
Will Berman's avatar
Will Berman committed
765
766
767
768

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

769
770
771
772
                # `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)

773
            class_emb = self.class_embedding(class_labels).to(dtype=sample.dtype)
Sanchit Gandhi's avatar
Sanchit Gandhi committed
774
775
776
777
778

            if self.config.class_embeddings_concat:
                emb = torch.cat([emb, class_emb], dim=-1)
            else:
                emb = emb + class_emb
779

Patrick von Platen's avatar
Patrick von Platen committed
780
781
782
        if self.config.addition_embed_type == "text":
            aug_emb = self.add_embedding(encoder_hidden_states)
            emb = emb + aug_emb
YiYi Xu's avatar
YiYi Xu committed
783
784
785
786
787
788
789
790
791
792
793
794
        elif self.config.addition_embed_type == "text_image":
            # Kadinsky 2.1 - style
            if "image_embeds" not in added_cond_kwargs:
                raise ValueError(
                    f"{self.__class__} has the config param `addition_embed_type` set to 'text_image' which requires the keyword argument `image_embeds` to be passed in `added_cond_kwargs`"
                )

            image_embs = added_cond_kwargs.get("image_embeds")
            text_embs = added_cond_kwargs.get("text_embeds", encoder_hidden_states)

            aug_emb = self.add_embedding(text_embs, image_embs)
            emb = emb + aug_emb
Patrick von Platen's avatar
Patrick von Platen committed
795

796
797
798
        if self.time_embed_act is not None:
            emb = self.time_embed_act(emb)

YiYi Xu's avatar
YiYi Xu committed
799
        if self.encoder_hid_proj is not None and self.config.encoder_hid_dim_type == "text_proj":
William Berman's avatar
William Berman committed
800
            encoder_hidden_states = self.encoder_hid_proj(encoder_hidden_states)
YiYi Xu's avatar
YiYi Xu committed
801
802
803
804
805
806
807
808
809
        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)
William Berman's avatar
William Berman committed
810

Patrick von Platen's avatar
Patrick von Platen committed
811
812
813
814
815
816
        # 2. pre-process
        sample = self.conv_in(sample)

        # 3. down
        down_block_res_samples = (sample,)
        for downsample_block in self.down_blocks:
817
            if hasattr(downsample_block, "has_cross_attention") and downsample_block.has_cross_attention:
Patrick von Platen's avatar
Patrick von Platen committed
818
                sample, res_samples = downsample_block(
819
820
821
                    hidden_states=sample,
                    temb=emb,
                    encoder_hidden_states=encoder_hidden_states,
Will Berman's avatar
Will Berman committed
822
                    attention_mask=attention_mask,
823
                    cross_attention_kwargs=cross_attention_kwargs,
824
                    encoder_attention_mask=encoder_attention_mask,
Patrick von Platen's avatar
Patrick von Platen committed
825
826
827
828
829
830
                )
            else:
                sample, res_samples = downsample_block(hidden_states=sample, temb=emb)

            down_block_res_samples += res_samples

831
832
833
834
835
836
        if down_block_additional_residuals is not None:
            new_down_block_res_samples = ()

            for down_block_res_sample, down_block_additional_residual in zip(
                down_block_res_samples, down_block_additional_residuals
            ):
837
                down_block_res_sample = down_block_res_sample + down_block_additional_residual
838
                new_down_block_res_samples = new_down_block_res_samples + (down_block_res_sample,)
839
840
841

            down_block_res_samples = new_down_block_res_samples

Patrick von Platen's avatar
Patrick von Platen committed
842
        # 4. mid
843
844
845
846
847
848
849
        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,
850
                encoder_attention_mask=encoder_attention_mask,
851
            )
Patrick von Platen's avatar
Patrick von Platen committed
852

853
        if mid_block_additional_residual is not None:
854
            sample = sample + mid_block_additional_residual
855

Patrick von Platen's avatar
Patrick von Platen committed
856
        # 5. up
857
858
859
        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
860
861
862
            res_samples = down_block_res_samples[-len(upsample_block.resnets) :]
            down_block_res_samples = down_block_res_samples[: -len(upsample_block.resnets)]

863
864
865
866
867
            # 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:]

868
            if hasattr(upsample_block, "has_cross_attention") and upsample_block.has_cross_attention:
Patrick von Platen's avatar
Patrick von Platen committed
869
870
871
872
873
                sample = upsample_block(
                    hidden_states=sample,
                    temb=emb,
                    res_hidden_states_tuple=res_samples,
                    encoder_hidden_states=encoder_hidden_states,
874
                    cross_attention_kwargs=cross_attention_kwargs,
875
                    upsample_size=upsample_size,
Will Berman's avatar
Will Berman committed
876
                    attention_mask=attention_mask,
877
                    encoder_attention_mask=encoder_attention_mask,
Patrick von Platen's avatar
Patrick von Platen committed
878
879
                )
            else:
880
881
882
                sample = upsample_block(
                    hidden_states=sample, temb=emb, res_hidden_states_tuple=res_samples, upsample_size=upsample_size
                )
883

Patrick von Platen's avatar
Patrick von Platen committed
884
        # 6. post-process
885
886
887
        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
888
889
        sample = self.conv_out(sample)

890
891
        if not return_dict:
            return (sample,)
Patrick von Platen's avatar
Patrick von Platen committed
892

893
        return UNet2DConditionOutput(sample=sample)