forked from huggingface/diffusers
-
Notifications
You must be signed in to change notification settings - Fork 0
Expand file tree
/
Copy pathuvit_2d.py
More file actions
420 lines (358 loc) 路 14.9 KB
/
Copy pathuvit_2d.py
File metadata and controls
420 lines (358 loc) 路 14.9 KB
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
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
# coding=utf-8
# Copyright 2025 The HuggingFace Inc. team.
#
# 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 torch
import torch.nn.functional as F
from torch import nn
from torch.utils.checkpoint import checkpoint
from ...configuration_utils import ConfigMixin, register_to_config
from ...loaders import PeftAdapterMixin
from ...utils import apply_lora_scale
from ..attention import AttentionMixin, BasicTransformerBlock, SkipFFTransformerBlock
from ..attention_processor import (
ADDED_KV_ATTENTION_PROCESSORS,
CROSS_ATTENTION_PROCESSORS,
AttnAddedKVProcessor,
AttnProcessor,
)
from ..embeddings import TimestepEmbedding, get_timestep_embedding
from ..modeling_utils import ModelMixin
from ..normalization import GlobalResponseNorm, RMSNorm
from ..resnet import Downsample2D, Upsample2D
class UVit2DModel(ModelMixin, AttentionMixin, ConfigMixin, PeftAdapterMixin):
_supports_gradient_checkpointing = True
@register_to_config
def __init__(
self,
# global config
hidden_size: int = 1024,
use_bias: bool = False,
hidden_dropout: float = 0.0,
# conditioning dimensions
cond_embed_dim: int = 768,
micro_cond_encode_dim: int = 256,
micro_cond_embed_dim: int = 1280,
encoder_hidden_size: int = 768,
# num tokens
vocab_size: int = 8256, # codebook_size + 1 (for the mask token) rounded
codebook_size: int = 8192,
# `UVit2DConvEmbed`
in_channels: int = 768,
block_out_channels: int = 768,
num_res_blocks: int = 3,
downsample: bool = False,
upsample: bool = False,
block_num_heads: int = 12,
# `TransformerLayer`
num_hidden_layers: int = 22,
num_attention_heads: int = 16,
# `Attention`
attention_dropout: float = 0.0,
# `FeedForward`
intermediate_size: int = 2816,
# `Norm`
layer_norm_eps: float = 1e-6,
ln_elementwise_affine: bool = True,
sample_size: int = 64,
):
super().__init__()
self.encoder_proj = nn.Linear(encoder_hidden_size, hidden_size, bias=use_bias)
self.encoder_proj_layer_norm = RMSNorm(hidden_size, layer_norm_eps, ln_elementwise_affine)
self.embed = UVit2DConvEmbed(
in_channels, block_out_channels, vocab_size, ln_elementwise_affine, layer_norm_eps, use_bias
)
self.cond_embed = TimestepEmbedding(
micro_cond_embed_dim + cond_embed_dim, hidden_size, sample_proj_bias=use_bias
)
self.down_block = UVitBlock(
block_out_channels,
num_res_blocks,
hidden_size,
hidden_dropout,
ln_elementwise_affine,
layer_norm_eps,
use_bias,
block_num_heads,
attention_dropout,
downsample,
False,
)
self.project_to_hidden_norm = RMSNorm(block_out_channels, layer_norm_eps, ln_elementwise_affine)
self.project_to_hidden = nn.Linear(block_out_channels, hidden_size, bias=use_bias)
self.transformer_layers = nn.ModuleList(
[
BasicTransformerBlock(
dim=hidden_size,
num_attention_heads=num_attention_heads,
attention_head_dim=hidden_size // num_attention_heads,
dropout=hidden_dropout,
cross_attention_dim=hidden_size,
attention_bias=use_bias,
norm_type="ada_norm_continuous",
ada_norm_continous_conditioning_embedding_dim=hidden_size,
norm_elementwise_affine=ln_elementwise_affine,
norm_eps=layer_norm_eps,
ada_norm_bias=use_bias,
ff_inner_dim=intermediate_size,
ff_bias=use_bias,
attention_out_bias=use_bias,
)
for _ in range(num_hidden_layers)
]
)
self.project_from_hidden_norm = RMSNorm(hidden_size, layer_norm_eps, ln_elementwise_affine)
self.project_from_hidden = nn.Linear(hidden_size, block_out_channels, bias=use_bias)
self.up_block = UVitBlock(
block_out_channels,
num_res_blocks,
hidden_size,
hidden_dropout,
ln_elementwise_affine,
layer_norm_eps,
use_bias,
block_num_heads,
attention_dropout,
downsample=False,
upsample=upsample,
)
self.mlm_layer = ConvMlmLayer(
block_out_channels, in_channels, use_bias, ln_elementwise_affine, layer_norm_eps, codebook_size
)
self.gradient_checkpointing = False
@apply_lora_scale("cross_attention_kwargs")
def forward(self, input_ids, encoder_hidden_states, pooled_text_emb, micro_conds, cross_attention_kwargs=None):
r"""
Args:
input_ids (`torch.LongTensor`):
Token ids of the masked latent image tokens, with shape `(batch_size, height, width)`.
encoder_hidden_states (`torch.Tensor`):
Conditional embeddings (embeddings computed from the input conditions such as prompts) to use.
pooled_text_emb (`torch.Tensor`):
Pooled text embeddings used for additional conditioning.
micro_conds (`torch.Tensor`):
Micro-conditioning values that are embedded and combined with `pooled_text_emb`.
cross_attention_kwargs (`dict`, *optional*):
A kwargs dictionary that if specified is passed along to the `AttentionProcessor`.
"""
encoder_hidden_states = self.encoder_proj(encoder_hidden_states)
encoder_hidden_states = self.encoder_proj_layer_norm(encoder_hidden_states)
micro_cond_embeds = get_timestep_embedding(
micro_conds.flatten(), self.config.micro_cond_encode_dim, flip_sin_to_cos=True, downscale_freq_shift=0
)
micro_cond_embeds = micro_cond_embeds.reshape((input_ids.shape[0], -1))
pooled_text_emb = torch.cat([pooled_text_emb, micro_cond_embeds], dim=1)
pooled_text_emb = pooled_text_emb.to(dtype=self.dtype)
pooled_text_emb = self.cond_embed(pooled_text_emb).to(encoder_hidden_states.dtype)
hidden_states = self.embed(input_ids)
hidden_states = self.down_block(
hidden_states,
pooled_text_emb=pooled_text_emb,
encoder_hidden_states=encoder_hidden_states,
cross_attention_kwargs=cross_attention_kwargs,
)
batch_size, channels, height, width = hidden_states.shape
hidden_states = hidden_states.permute(0, 2, 3, 1).reshape(batch_size, height * width, channels)
hidden_states = self.project_to_hidden_norm(hidden_states)
hidden_states = self.project_to_hidden(hidden_states)
for layer in self.transformer_layers:
if torch.is_grad_enabled() and self.gradient_checkpointing:
def layer_(*args):
return checkpoint(layer, *args)
else:
layer_ = layer
hidden_states = layer_(
hidden_states,
encoder_hidden_states=encoder_hidden_states,
cross_attention_kwargs=cross_attention_kwargs,
added_cond_kwargs={"pooled_text_emb": pooled_text_emb},
)
hidden_states = self.project_from_hidden_norm(hidden_states)
hidden_states = self.project_from_hidden(hidden_states)
hidden_states = hidden_states.reshape(batch_size, height, width, channels).permute(0, 3, 1, 2)
hidden_states = self.up_block(
hidden_states,
pooled_text_emb=pooled_text_emb,
encoder_hidden_states=encoder_hidden_states,
cross_attention_kwargs=cross_attention_kwargs,
)
logits = self.mlm_layer(hidden_states)
return logits
# Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_default_attn_processor
def set_default_attn_processor(self):
"""
Disables custom attention processors and sets the default attention implementation.
"""
if all(proc.__class__ in ADDED_KV_ATTENTION_PROCESSORS for proc in self.attn_processors.values()):
processor = AttnAddedKVProcessor()
elif all(proc.__class__ in CROSS_ATTENTION_PROCESSORS for proc in self.attn_processors.values()):
processor = AttnProcessor()
else:
raise ValueError(
f"Cannot call `set_default_attn_processor` when attention processors are of type {next(iter(self.attn_processors.values()))}"
)
self.set_attn_processor(processor)
class UVit2DConvEmbed(nn.Module):
def __init__(self, in_channels, block_out_channels, vocab_size, elementwise_affine, eps, bias):
super().__init__()
self.embeddings = nn.Embedding(vocab_size, in_channels)
self.layer_norm = RMSNorm(in_channels, eps, elementwise_affine)
self.conv = nn.Conv2d(in_channels, block_out_channels, kernel_size=1, bias=bias)
def forward(self, input_ids):
embeddings = self.embeddings(input_ids)
embeddings = self.layer_norm(embeddings)
embeddings = embeddings.permute(0, 3, 1, 2)
embeddings = self.conv(embeddings)
return embeddings
class UVitBlock(nn.Module):
def __init__(
self,
channels,
num_res_blocks: int,
hidden_size,
hidden_dropout,
ln_elementwise_affine,
layer_norm_eps,
use_bias,
block_num_heads,
attention_dropout,
downsample: bool,
upsample: bool,
):
super().__init__()
if downsample:
self.downsample = Downsample2D(
channels,
use_conv=True,
padding=0,
name="Conv2d_0",
kernel_size=2,
norm_type="rms_norm",
eps=layer_norm_eps,
elementwise_affine=ln_elementwise_affine,
bias=use_bias,
)
else:
self.downsample = None
self.res_blocks = nn.ModuleList(
[
ConvNextBlock(
channels,
layer_norm_eps,
ln_elementwise_affine,
use_bias,
hidden_dropout,
hidden_size,
)
for i in range(num_res_blocks)
]
)
self.attention_blocks = nn.ModuleList(
[
SkipFFTransformerBlock(
channels,
block_num_heads,
channels // block_num_heads,
hidden_size,
use_bias,
attention_dropout,
channels,
attention_bias=use_bias,
attention_out_bias=use_bias,
)
for _ in range(num_res_blocks)
]
)
if upsample:
self.upsample = Upsample2D(
channels,
use_conv_transpose=True,
kernel_size=2,
padding=0,
name="conv",
norm_type="rms_norm",
eps=layer_norm_eps,
elementwise_affine=ln_elementwise_affine,
bias=use_bias,
interpolate=False,
)
else:
self.upsample = None
def forward(self, x, pooled_text_emb, encoder_hidden_states, cross_attention_kwargs):
if self.downsample is not None:
x = self.downsample(x)
for res_block, attention_block in zip(self.res_blocks, self.attention_blocks):
x = res_block(x, pooled_text_emb)
batch_size, channels, height, width = x.shape
x = x.view(batch_size, channels, height * width).permute(0, 2, 1)
x = attention_block(
x, encoder_hidden_states=encoder_hidden_states, cross_attention_kwargs=cross_attention_kwargs
)
x = x.permute(0, 2, 1).view(batch_size, channels, height, width)
if self.upsample is not None:
x = self.upsample(x)
return x
class ConvNextBlock(nn.Module):
def __init__(
self, channels, layer_norm_eps, ln_elementwise_affine, use_bias, hidden_dropout, hidden_size, res_ffn_factor=4
):
super().__init__()
self.depthwise = nn.Conv2d(
channels,
channels,
kernel_size=3,
padding=1,
groups=channels,
bias=use_bias,
)
self.norm = RMSNorm(channels, layer_norm_eps, ln_elementwise_affine)
self.channelwise_linear_1 = nn.Linear(channels, int(channels * res_ffn_factor), bias=use_bias)
self.channelwise_act = nn.GELU()
self.channelwise_norm = GlobalResponseNorm(int(channels * res_ffn_factor))
self.channelwise_linear_2 = nn.Linear(int(channels * res_ffn_factor), channels, bias=use_bias)
self.channelwise_dropout = nn.Dropout(hidden_dropout)
self.cond_embeds_mapper = nn.Linear(hidden_size, channels * 2, use_bias)
def forward(self, x, cond_embeds):
x_res = x
x = self.depthwise(x)
x = x.permute(0, 2, 3, 1)
x = self.norm(x)
x = self.channelwise_linear_1(x)
x = self.channelwise_act(x)
x = self.channelwise_norm(x)
x = self.channelwise_linear_2(x)
x = self.channelwise_dropout(x)
x = x.permute(0, 3, 1, 2)
x = x + x_res
scale, shift = self.cond_embeds_mapper(F.silu(cond_embeds)).chunk(2, dim=1)
x = x * (1 + scale[:, :, None, None]) + shift[:, :, None, None]
return x
class ConvMlmLayer(nn.Module):
def __init__(
self,
block_out_channels: int,
in_channels: int,
use_bias: bool,
ln_elementwise_affine: bool,
layer_norm_eps: float,
codebook_size: int,
):
super().__init__()
self.conv1 = nn.Conv2d(block_out_channels, in_channels, kernel_size=1, bias=use_bias)
self.layer_norm = RMSNorm(in_channels, layer_norm_eps, ln_elementwise_affine)
self.conv2 = nn.Conv2d(in_channels, codebook_size, kernel_size=1, bias=use_bias)
def forward(self, hidden_states):
hidden_states = self.conv1(hidden_states)
hidden_states = self.layer_norm(hidden_states.permute(0, 2, 3, 1)).permute(0, 3, 1, 2)
logits = self.conv2(hidden_states)
return logits