File size: 15,972 Bytes
0324143
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
# Copyright 2024 The HuggingFace Team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

import flax.linen as nn
import jax.numpy as jnp

from ..attention_flax import FlaxTransformer2DModel
from ..resnet_flax import FlaxDownsample2D, FlaxResnetBlock2D, FlaxUpsample2D


class FlaxCrossAttnDownBlock2D(nn.Module):
    r"""

    Cross Attention 2D Downsizing block - original architecture from Unet transformers:

    https://arxiv.org/abs/2103.06104



    Parameters:

        in_channels (:obj:`int`):

            Input channels

        out_channels (:obj:`int`):

            Output channels

        dropout (:obj:`float`, *optional*, defaults to 0.0):

            Dropout rate

        num_layers (:obj:`int`, *optional*, defaults to 1):

            Number of attention blocks layers

        num_attention_heads (:obj:`int`, *optional*, defaults to 1):

            Number of attention heads of each spatial transformer block

        add_downsample (:obj:`bool`, *optional*, defaults to `True`):

            Whether to add downsampling layer before each final output

        use_memory_efficient_attention (`bool`, *optional*, defaults to `False`):

            enable memory efficient attention https://arxiv.org/abs/2112.05682

        split_head_dim (`bool`, *optional*, defaults to `False`):

            Whether to split the head dimension into a new axis for the self-attention computation. In most cases,

            enabling this flag should speed up the computation for Stable Diffusion 2.x and Stable Diffusion XL.

        dtype (:obj:`jnp.dtype`, *optional*, defaults to jnp.float32):

            Parameters `dtype`

    """

    in_channels: int
    out_channels: int
    dropout: float = 0.0
    num_layers: int = 1
    num_attention_heads: int = 1
    add_downsample: bool = True
    use_linear_projection: bool = False
    only_cross_attention: bool = False
    use_memory_efficient_attention: bool = False
    split_head_dim: bool = False
    dtype: jnp.dtype = jnp.float32
    transformer_layers_per_block: int = 1

    def setup(self):
        resnets = []
        attentions = []

        for i in range(self.num_layers):
            in_channels = self.in_channels if i == 0 else self.out_channels

            res_block = FlaxResnetBlock2D(
                in_channels=in_channels,
                out_channels=self.out_channels,
                dropout_prob=self.dropout,
                dtype=self.dtype,
            )
            resnets.append(res_block)

            attn_block = FlaxTransformer2DModel(
                in_channels=self.out_channels,
                n_heads=self.num_attention_heads,
                d_head=self.out_channels // self.num_attention_heads,
                depth=self.transformer_layers_per_block,
                use_linear_projection=self.use_linear_projection,
                only_cross_attention=self.only_cross_attention,
                use_memory_efficient_attention=self.use_memory_efficient_attention,
                split_head_dim=self.split_head_dim,
                dtype=self.dtype,
            )
            attentions.append(attn_block)

        self.resnets = resnets
        self.attentions = attentions

        if self.add_downsample:
            self.downsamplers_0 = FlaxDownsample2D(self.out_channels, dtype=self.dtype)

    def __call__(self, hidden_states, temb, encoder_hidden_states, deterministic=True):
        output_states = ()

        for resnet, attn in zip(self.resnets, self.attentions):
            hidden_states = resnet(hidden_states, temb, deterministic=deterministic)
            hidden_states = attn(hidden_states, encoder_hidden_states, deterministic=deterministic)
            output_states += (hidden_states,)

        if self.add_downsample:
            hidden_states = self.downsamplers_0(hidden_states)
            output_states += (hidden_states,)

        return hidden_states, output_states


class FlaxDownBlock2D(nn.Module):
    r"""

    Flax 2D downsizing block



    Parameters:

        in_channels (:obj:`int`):

            Input channels

        out_channels (:obj:`int`):

            Output channels

        dropout (:obj:`float`, *optional*, defaults to 0.0):

            Dropout rate

        num_layers (:obj:`int`, *optional*, defaults to 1):

            Number of attention blocks layers

        add_downsample (:obj:`bool`, *optional*, defaults to `True`):

            Whether to add downsampling layer before each final output

        dtype (:obj:`jnp.dtype`, *optional*, defaults to jnp.float32):

            Parameters `dtype`

    """

    in_channels: int
    out_channels: int
    dropout: float = 0.0
    num_layers: int = 1
    add_downsample: bool = True
    dtype: jnp.dtype = jnp.float32

    def setup(self):
        resnets = []

        for i in range(self.num_layers):
            in_channels = self.in_channels if i == 0 else self.out_channels

            res_block = FlaxResnetBlock2D(
                in_channels=in_channels,
                out_channels=self.out_channels,
                dropout_prob=self.dropout,
                dtype=self.dtype,
            )
            resnets.append(res_block)
        self.resnets = resnets

        if self.add_downsample:
            self.downsamplers_0 = FlaxDownsample2D(self.out_channels, dtype=self.dtype)

    def __call__(self, hidden_states, temb, deterministic=True):
        output_states = ()

        for resnet in self.resnets:
            hidden_states = resnet(hidden_states, temb, deterministic=deterministic)
            output_states += (hidden_states,)

        if self.add_downsample:
            hidden_states = self.downsamplers_0(hidden_states)
            output_states += (hidden_states,)

        return hidden_states, output_states


class FlaxCrossAttnUpBlock2D(nn.Module):
    r"""

    Cross Attention 2D Upsampling block - original architecture from Unet transformers:

    https://arxiv.org/abs/2103.06104



    Parameters:

        in_channels (:obj:`int`):

            Input channels

        out_channels (:obj:`int`):

            Output channels

        dropout (:obj:`float`, *optional*, defaults to 0.0):

            Dropout rate

        num_layers (:obj:`int`, *optional*, defaults to 1):

            Number of attention blocks layers

        num_attention_heads (:obj:`int`, *optional*, defaults to 1):

            Number of attention heads of each spatial transformer block

        add_upsample (:obj:`bool`, *optional*, defaults to `True`):

            Whether to add upsampling layer before each final output

        use_memory_efficient_attention (`bool`, *optional*, defaults to `False`):

            enable memory efficient attention https://arxiv.org/abs/2112.05682

        split_head_dim (`bool`, *optional*, defaults to `False`):

            Whether to split the head dimension into a new axis for the self-attention computation. In most cases,

            enabling this flag should speed up the computation for Stable Diffusion 2.x and Stable Diffusion XL.

        dtype (:obj:`jnp.dtype`, *optional*, defaults to jnp.float32):

            Parameters `dtype`

    """

    in_channels: int
    out_channels: int
    prev_output_channel: int
    dropout: float = 0.0
    num_layers: int = 1
    num_attention_heads: int = 1
    add_upsample: bool = True
    use_linear_projection: bool = False
    only_cross_attention: bool = False
    use_memory_efficient_attention: bool = False
    split_head_dim: bool = False
    dtype: jnp.dtype = jnp.float32
    transformer_layers_per_block: int = 1

    def setup(self):
        resnets = []
        attentions = []

        for i in range(self.num_layers):
            res_skip_channels = self.in_channels if (i == self.num_layers - 1) else self.out_channels
            resnet_in_channels = self.prev_output_channel if i == 0 else self.out_channels

            res_block = FlaxResnetBlock2D(
                in_channels=resnet_in_channels + res_skip_channels,
                out_channels=self.out_channels,
                dropout_prob=self.dropout,
                dtype=self.dtype,
            )
            resnets.append(res_block)

            attn_block = FlaxTransformer2DModel(
                in_channels=self.out_channels,
                n_heads=self.num_attention_heads,
                d_head=self.out_channels // self.num_attention_heads,
                depth=self.transformer_layers_per_block,
                use_linear_projection=self.use_linear_projection,
                only_cross_attention=self.only_cross_attention,
                use_memory_efficient_attention=self.use_memory_efficient_attention,
                split_head_dim=self.split_head_dim,
                dtype=self.dtype,
            )
            attentions.append(attn_block)

        self.resnets = resnets
        self.attentions = attentions

        if self.add_upsample:
            self.upsamplers_0 = FlaxUpsample2D(self.out_channels, dtype=self.dtype)

    def __call__(self, hidden_states, res_hidden_states_tuple, temb, encoder_hidden_states, deterministic=True):
        for resnet, attn in zip(self.resnets, self.attentions):
            # pop res hidden states
            res_hidden_states = res_hidden_states_tuple[-1]
            res_hidden_states_tuple = res_hidden_states_tuple[:-1]
            hidden_states = jnp.concatenate((hidden_states, res_hidden_states), axis=-1)

            hidden_states = resnet(hidden_states, temb, deterministic=deterministic)
            hidden_states = attn(hidden_states, encoder_hidden_states, deterministic=deterministic)

        if self.add_upsample:
            hidden_states = self.upsamplers_0(hidden_states)

        return hidden_states


class FlaxUpBlock2D(nn.Module):
    r"""

    Flax 2D upsampling block



    Parameters:

        in_channels (:obj:`int`):

            Input channels

        out_channels (:obj:`int`):

            Output channels

        prev_output_channel (:obj:`int`):

            Output channels from the previous block

        dropout (:obj:`float`, *optional*, defaults to 0.0):

            Dropout rate

        num_layers (:obj:`int`, *optional*, defaults to 1):

            Number of attention blocks layers

        add_downsample (:obj:`bool`, *optional*, defaults to `True`):

            Whether to add downsampling layer before each final output

        dtype (:obj:`jnp.dtype`, *optional*, defaults to jnp.float32):

            Parameters `dtype`

    """

    in_channels: int
    out_channels: int
    prev_output_channel: int
    dropout: float = 0.0
    num_layers: int = 1
    add_upsample: bool = True
    dtype: jnp.dtype = jnp.float32

    def setup(self):
        resnets = []

        for i in range(self.num_layers):
            res_skip_channels = self.in_channels if (i == self.num_layers - 1) else self.out_channels
            resnet_in_channels = self.prev_output_channel if i == 0 else self.out_channels

            res_block = FlaxResnetBlock2D(
                in_channels=resnet_in_channels + res_skip_channels,
                out_channels=self.out_channels,
                dropout_prob=self.dropout,
                dtype=self.dtype,
            )
            resnets.append(res_block)

        self.resnets = resnets

        if self.add_upsample:
            self.upsamplers_0 = FlaxUpsample2D(self.out_channels, dtype=self.dtype)

    def __call__(self, hidden_states, res_hidden_states_tuple, temb, deterministic=True):
        for resnet in self.resnets:
            # pop res hidden states
            res_hidden_states = res_hidden_states_tuple[-1]
            res_hidden_states_tuple = res_hidden_states_tuple[:-1]
            hidden_states = jnp.concatenate((hidden_states, res_hidden_states), axis=-1)

            hidden_states = resnet(hidden_states, temb, deterministic=deterministic)

        if self.add_upsample:
            hidden_states = self.upsamplers_0(hidden_states)

        return hidden_states


class FlaxUNetMidBlock2DCrossAttn(nn.Module):
    r"""

    Cross Attention 2D Mid-level block - original architecture from Unet transformers: https://arxiv.org/abs/2103.06104



    Parameters:

        in_channels (:obj:`int`):

            Input channels

        dropout (:obj:`float`, *optional*, defaults to 0.0):

            Dropout rate

        num_layers (:obj:`int`, *optional*, defaults to 1):

            Number of attention blocks layers

        num_attention_heads (:obj:`int`, *optional*, defaults to 1):

            Number of attention heads of each spatial transformer block

        use_memory_efficient_attention (`bool`, *optional*, defaults to `False`):

            enable memory efficient attention https://arxiv.org/abs/2112.05682

        split_head_dim (`bool`, *optional*, defaults to `False`):

            Whether to split the head dimension into a new axis for the self-attention computation. In most cases,

            enabling this flag should speed up the computation for Stable Diffusion 2.x and Stable Diffusion XL.

        dtype (:obj:`jnp.dtype`, *optional*, defaults to jnp.float32):

            Parameters `dtype`

    """

    in_channels: int
    dropout: float = 0.0
    num_layers: int = 1
    num_attention_heads: int = 1
    use_linear_projection: bool = False
    use_memory_efficient_attention: bool = False
    split_head_dim: bool = False
    dtype: jnp.dtype = jnp.float32
    transformer_layers_per_block: int = 1

    def setup(self):
        # there is always at least one resnet
        resnets = [
            FlaxResnetBlock2D(
                in_channels=self.in_channels,
                out_channels=self.in_channels,
                dropout_prob=self.dropout,
                dtype=self.dtype,
            )
        ]

        attentions = []

        for _ in range(self.num_layers):
            attn_block = FlaxTransformer2DModel(
                in_channels=self.in_channels,
                n_heads=self.num_attention_heads,
                d_head=self.in_channels // self.num_attention_heads,
                depth=self.transformer_layers_per_block,
                use_linear_projection=self.use_linear_projection,
                use_memory_efficient_attention=self.use_memory_efficient_attention,
                split_head_dim=self.split_head_dim,
                dtype=self.dtype,
            )
            attentions.append(attn_block)

            res_block = FlaxResnetBlock2D(
                in_channels=self.in_channels,
                out_channels=self.in_channels,
                dropout_prob=self.dropout,
                dtype=self.dtype,
            )
            resnets.append(res_block)

        self.resnets = resnets
        self.attentions = attentions

    def __call__(self, hidden_states, temb, encoder_hidden_states, deterministic=True):
        hidden_states = self.resnets[0](hidden_states, temb)
        for attn, resnet in zip(self.attentions, self.resnets[1:]):
            hidden_states = attn(hidden_states, encoder_hidden_states, deterministic=deterministic)
            hidden_states = resnet(hidden_states, temb, deterministic=deterministic)

        return hidden_states