-
Notifications
You must be signed in to change notification settings - Fork 129
/
model.py
483 lines (428 loc) · 18.2 KB
/
model.py
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
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
# Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.
#
# This source code is licensed under the BSD-style license found in the
# LICENSE file in the root directory of this source tree.
from typing import Callable, NamedTuple, Optional, Tuple, Union
import torch
from torch import nn, Tensor
from torchmultimodal.models.masked_auto_encoder.position_embeddings import (
get_2d_sin_cos_embeddings,
)
from torchmultimodal.models.masked_auto_encoder.swin_decoder import SwinTransformer
from torchmultimodal.modules.encoders.vision_transformer import (
VisionTransformer,
vit_b_16,
vit_l_16,
)
from torchmultimodal.modules.layers.patch_embedding import PatchEmbeddings
from torchmultimodal.modules.layers.transformer import (
TransformerEncoder,
TransformerOutput,
)
MAE_MODEL_MAPPING = {
"vit_b16_image": "https://download.pytorch.org/models/multimodal/mae/mae_pretrained_vit_base.pth",
"vit_l16_image": "https://download.pytorch.org/models/multimodal/mae/mae_pretrained_vit_large.pth",
"vit_b16_audio": "https://download.pytorch.org/models/multimodal/audio_mae/audio_mae_pretrained_vit_base.pth",
}
class MAEOutput(NamedTuple):
encoder_output: Union[TransformerOutput, Tensor]
decoder_pred: Optional[Tensor] = None
label_patches: Optional[Tensor] = None
mask: Optional[Tensor] = None
class MaskedAutoEncoder(nn.Module):
"""
MAE (https://arxiv.org/abs/2111.06377) is a pretraining technique to mask out patches of the input
before passing through the encoder and then using a decoder to predict the masked patches
The code has been adapted from the original implementation https://github.com/facebookresearch/mae
Args:
encoder_transformer (nn.Module): instance of encoder transformer
decoder_transformer (nn.Module): instance of decoder transformer
input_size (Union[int, Tuple[int,int]): size of the input. if tuple, the format should be height,width.
If an int, a square input is assumed. Default: 224
patch_size (int): size of the patches. Default: 16
num_channels (int): number of input channels. Default: 3
embed_dim (int): embedding dim of input to the encoder transformer (or output dim of patch embedding). Default: 768
masking_ratio (float): ratio of patches to mask. Default: 0.75
decoder_embed_dim (int): embedding dim of the input to the decoder transformer. Default: 512
"""
def __init__(
self,
encoder_transformer: nn.Module,
decoder_transformer: nn.Module,
input_size: Union[int, Tuple[int, int]] = 224,
patch_size: int = 16,
num_channels: int = 3,
embed_dim: int = 768,
masking_ratio: float = 0.75,
decoder_embed_dim: int = 512,
use_cls_in_decoder: bool = True,
):
super().__init__()
self.patch_size = patch_size
self.embeddings = PatchEmbeddings(
image_size=input_size,
patch_size=patch_size,
num_channels=num_channels,
hidden_size=embed_dim,
patch_drop_rate=masking_ratio,
)
self.embeddings.position_embeddings.requires_grad = False
self.encoder = encoder_transformer
self.decoder_embed = DecoderEmbeddings(
encoder_embed_dim=embed_dim,
decoder_embed_dim=decoder_embed_dim,
image_size=input_size,
patch_size=patch_size,
)
self.decoder_embed.position_embeddings.requires_grad = False
self.decoder_transformer = decoder_transformer
self.decoder_pred = nn.Linear(decoder_embed_dim, patch_size**2 * num_channels)
self.use_cls_in_decoder = use_cls_in_decoder
self._initialize_weights(input_size, embed_dim, decoder_embed_dim)
def _initialize_weights(
self,
input_size: Union[int, Tuple[int, int]],
encoder_embed_dim: int,
decoder_embed_dim: int,
) -> None:
if isinstance(input_size, int):
input_h = input_w = input_size
else:
input_h, input_w = input_size
num_patches_h = input_h // self.patch_size
num_patches_w = input_w // self.patch_size
self.embeddings.position_embeddings.data = get_2d_sin_cos_embeddings(
encoder_embed_dim, (num_patches_w, num_patches_h)
)
self.decoder_embed.position_embeddings.data = get_2d_sin_cos_embeddings(
decoder_embed_dim, (num_patches_w, num_patches_h)
)
# initialize embeddings like nn.Linear (instead of nn.Conv2d)
w = self.embeddings.conv_projection.weight.data
torch.nn.init.xavier_uniform_(w.view([w.shape[0], -1]))
torch.nn.init.normal_(self.embeddings.cls_token, std=0.02)
torch.nn.init.normal_(self.decoder_embed.mask_token, std=0.02)
self.apply(self._init_weights)
def _init_weights(self, module: nn.Module) -> None:
if isinstance(module, nn.Linear):
torch.nn.init.xavier_uniform_(module.weight)
if module.bias is not None:
nn.init.constant_(module.bias, 0)
elif isinstance(module, nn.LayerNorm):
nn.init.constant_(module.bias, 0)
nn.init.constant_(module.weight, 1.0)
def _patchify_input(self, x: Tensor) -> Tensor:
# patchify the input tensor with the output shape = bsz x num_patch x (patch_area * channels)
bsz, channels, height, width = x.shape
num_patches_h = height // self.patch_size
num_patches_w = width // self.patch_size
label_patches = x.reshape(
bsz,
channels,
num_patches_h,
self.patch_size,
num_patches_w,
self.patch_size,
)
label_patches = torch.einsum("nchpwq->nhwpqc", label_patches)
label_patches = label_patches.reshape(
bsz, num_patches_h * num_patches_w, self.patch_size**2 * channels
)
return label_patches
def forward(self, x: Tensor) -> MAEOutput:
"""
Args:
x (Tensor): input tensor with shape bsz x channels x height x width
Returns:
output of MAEOutput type where encoder_output gives the output from the encoder,
decoder_pred gives prediction from the decoder followed by linear head,
mask indicates the masked out patches i.e. 1 refers to masked patches and 0 refers to unmasked patches
label_patches indicates the patchified ground truth pixels
"""
embedding_out = self.embeddings(x)
encoder_out = self.encoder(embedding_out.embeddings)
if not self.training:
# TODO: check if error should be raised is masking ratio != 0 here
return MAEOutput(encoder_out)
decoder_embedding = self.decoder_embed(
encoder_out.last_hidden_state, embedding_out.ids_restore
)
decoder_input = decoder_embedding
if not self.use_cls_in_decoder:
decoder_input = decoder_input[:, 1:, :]
decoder_out = self.decoder_transformer(decoder_input)
pred = self.decoder_pred(decoder_out.last_hidden_state)
if self.use_cls_in_decoder:
pred = pred[:, 1:, :]
label_patches = self._patchify_input(x)
return MAEOutput(
encoder_output=encoder_out,
decoder_pred=pred,
label_patches=label_patches,
mask=embedding_out.random_mask,
)
class DecoderEmbeddings(nn.Module):
"""
Construct the decoder embeddings from encoder embeddings.
Args:
encoder_embed_dim (int): Input dim for decoder embedding i.e. output dim of the encoder.
decoder_embed_dim (int): output dim for decoder embedding.
image_size (Union[int, Tuple[int, int]]): Size of the original input image. If set to an int, we assume a square input.
Defaults to 224.
patch_size (int): Patch size for the decoder.
"""
def __init__(
self,
encoder_embed_dim: int,
decoder_embed_dim: int,
image_size: Union[int, Tuple[int, int]] = 224,
patch_size: int = 16,
) -> None:
super().__init__()
self.decoder_embed = nn.Linear(encoder_embed_dim, decoder_embed_dim, bias=True)
self.mask_token = nn.Parameter(torch.zeros(1, 1, decoder_embed_dim))
if isinstance(image_size, int):
image_size = (image_size, image_size)
num_patches = (image_size[0] // patch_size) * (image_size[1] // patch_size)
self.position_embeddings = nn.Parameter(
torch.zeros(1, num_patches + 1, decoder_embed_dim)
)
def forward(
self,
x: Tensor,
ids_restore: Tensor,
) -> Tensor:
x = self.decoder_embed(x)
mask_tokens = self.mask_token.repeat(
x.shape[0], ids_restore.shape[1] + 1 - x.shape[1], 1
)
x_ = torch.cat([x[:, 1:, :], mask_tokens], dim=1) # no cls token
x_ = torch.gather(
x_, dim=1, index=ids_restore.unsqueeze(-1).repeat(1, 1, x.shape[2])
) # unshuffle
x = torch.cat([x[:, :1, :], x_], dim=1) # append cls token
# adding positional embeddings
x = x + self.position_embeddings
return x
def image_mae(
*,
# patch embedding
image_size: int = 224,
patch_size: int = 16,
masking_ratio: float = 0.75,
# encoder
encoder_layers: int = 12,
encoder_hidden_dim: int = 768,
encoder_heads: int = 12,
encoder_dim_feedforward: int = 3072,
encoder_layer_norm_eps: float = 1e-6,
encoder_activation: Callable = nn.GELU,
encoder_final_layer_norm_eps: float = 1e-6,
# decoder
decoder_layers: int = 8,
decoder_hidden_dim: int = 512,
decoder_heads: int = 16,
decoder_dim_feedforward: int = 2048,
decoder_layer_norm_eps: float = 1e-6,
decoder_activation: Callable = nn.GELU,
decoder_final_layer_norm_eps: float = 1e-6,
) -> MaskedAutoEncoder:
"""
Helper function to build the image mae model instantiation as described in the paper
with the encoder and decoder transformer similar to vision transformer.
Args:
image_size (int): size of the input. Default: 224
patch_size (int): size of the patches. Default: 16
masking_ratio (float): ratio of patches to mask. Default: 0.75
encoder_layers(int): number of encoder layers. Default: 12
encoder_hidden_dim (int): hidden dim of the encoder transformer. Default: 768
encoder_heads (int): number of encoder heads. Default: 12
encoder_dim_feedforward (int): hidden dim of the encoder transformer feedforward layer. Default: 3072
encoder_activation (Callable): activation function for encoder layers. Default: nn.GELU
encoder_layer_norm_eps (float): epsilon for encoder layer normalization. Default: 1e-6
encoder_final_layer_norm_eps (float): epsilon for encoder final layer normalization. Default: 1e-6
decoder_layers(int): number of decoder layers. Default: 8
decoder_hidden_dim (int): hidden dim of the decoder transformer. Default: 512
decoder_heads (int): number of decoder heads. Default: 16
decoder_dim_feedforward (int): hidden dim of the decoder transformer feedforward layer. Default: 2048
decoder_layer_norm_eps (float): epsilon for decoder layer normalization. Default: 1e-6
decoder_activation (float): activation function for decoder layers. Default: nn.GELU
decoder_final_layer_norm_eps (float): epsilon for decoder final layer normalization. Default: 1e-6
"""
encoder_transformer = TransformerEncoder(
n_layer=encoder_layers,
d_model=encoder_hidden_dim,
n_head=encoder_heads,
dim_feedforward=encoder_dim_feedforward,
final_layer_norm_eps=encoder_final_layer_norm_eps,
layer_norm_eps=encoder_layer_norm_eps,
norm_first=True,
activation=encoder_activation,
)
decoder_transformer = TransformerEncoder(
n_layer=decoder_layers,
d_model=decoder_hidden_dim,
n_head=decoder_heads,
dim_feedforward=decoder_dim_feedforward,
layer_norm_eps=decoder_layer_norm_eps,
final_layer_norm_eps=decoder_final_layer_norm_eps,
norm_first=True,
activation=decoder_activation,
)
return MaskedAutoEncoder(
encoder_transformer=encoder_transformer,
decoder_transformer=decoder_transformer,
input_size=image_size,
patch_size=patch_size,
num_channels=3,
embed_dim=encoder_hidden_dim,
masking_ratio=masking_ratio,
decoder_embed_dim=decoder_hidden_dim,
)
def vit_l_16_image_mae() -> MaskedAutoEncoder:
return image_mae(
image_size=224,
patch_size=16,
masking_ratio=0.75,
encoder_layers=24,
encoder_hidden_dim=1024,
encoder_heads=16,
encoder_dim_feedforward=4096,
decoder_layers=8,
decoder_hidden_dim=512,
decoder_heads=16,
decoder_dim_feedforward=2048,
)
def vit_b_16_image_mae_encoder(pretrained: bool = False) -> VisionTransformer:
ckpt_path = MAE_MODEL_MAPPING["vit_b16_image"] if pretrained else None
return vit_b_16(final_layer_norm_eps=None, ckpt_path=ckpt_path)
def vit_l_16_image_mae_encoder(pretrained: bool = False) -> VisionTransformer:
ckpt_path = MAE_MODEL_MAPPING["vit_l16_image"] if pretrained else None
return vit_l_16(final_layer_norm_eps=None, ckpt_path=ckpt_path)
def audio_mae(
*,
# patch embedding
input_size: Tuple[int, int] = (1024, 128),
patch_size: int = 16,
masking_ratio: float = 0.8,
# encoder
encoder_layers: int = 12,
encoder_hidden_dim: int = 768,
encoder_heads: int = 16,
encoder_dim_feedforward: int = 3072,
encoder_layer_norm_eps: float = 1e-6,
encoder_activation: Callable = nn.GELU,
encoder_final_layer_norm_eps: float = 1e-6,
# decoder
window_size: Tuple[int, int] = (4, 4),
decoder_layers: int = 16,
decoder_hidden_dim: int = 512,
decoder_heads: int = 16,
decoder_dim_feedforward: int = 2048,
decoder_layer_norm_eps: float = 1e-6,
decoder_activation: Callable = nn.GELU,
decoder_final_layer_norm_eps: float = 1e-6,
) -> MaskedAutoEncoder:
"""
Helper function to build the standard audio mae model with the encoder similar to vision transformer\
and decoder transformer similar to swin transformer.
Args:
image_size (Tuple[int, int]): (height, width) of the input. Default: (1024, 128)
patch_size (int): size of the patches. Default: 16
masking_ratio (float): ratio of patches to mask. Default: 0.8
encoder_layers(int): number of encoder layers. Default: 12
encoder_hidden_dim (int): hidden dim of the encoder transformer. Default: 768
encoder_heads (int): number of encoder heads. Default: 16
encoder_dim_feedforward (int): hidden dim of the encoder transformer feedforward layer. Default: 3072
encoder_activation (Callable): activation function for encoder layers. Default: nn.GELU
encoder_layer_norm_eps (float): epsilon for encoder layer normalization. Default: 1e-6
encoder_final_layer_norm_eps (float): epsilon for encoder final layer normalization. Default: 1e-6
decoder_layers(int): number of decoder layers. Default: 16
decoder_hidden_dim (int): hidden dim of the decoder transformer. Default: 512
decoder_heads (int): number of decoder heads. Default: 16
decoder_dim_feedforward (int): hidden dim of the decoder transformer feedforward layer. Default: 2048
decoder_layer_norm_eps (float): epsilon for decoder layer normalization. Default: 1e-6
decoder_activation (float): activation function for decoder layers. Default: nn.GELU
decoder_final_layer_norm_eps (float): epsilon for decoder final layer normalization. Default: 1e-6
"""
encoder_transformer = TransformerEncoder(
n_layer=encoder_layers,
d_model=encoder_hidden_dim,
n_head=encoder_heads,
dim_feedforward=encoder_dim_feedforward,
final_layer_norm_eps=encoder_final_layer_norm_eps,
layer_norm_eps=encoder_layer_norm_eps,
norm_first=True,
activation=encoder_activation,
)
decoder_input_size = (input_size[0] // patch_size, input_size[1] // patch_size)
decoder_transformer = SwinTransformer(
n_layer=decoder_layers,
input_dim=decoder_hidden_dim,
feedforward_dim=decoder_dim_feedforward,
num_heads=decoder_heads,
input_size=decoder_input_size,
window_size=window_size,
)
return MaskedAutoEncoder(
encoder_transformer=encoder_transformer,
decoder_transformer=decoder_transformer,
input_size=input_size,
patch_size=patch_size,
num_channels=1,
embed_dim=encoder_hidden_dim,
masking_ratio=masking_ratio,
decoder_embed_dim=decoder_hidden_dim,
use_cls_in_decoder=False,
)
def vit_s_16_audio_mae() -> MaskedAutoEncoder:
return audio_mae(
input_size=(1024, 128),
patch_size=16,
masking_ratio=0.8,
encoder_layers=12,
encoder_hidden_dim=384,
encoder_heads=6,
encoder_dim_feedforward=1536,
decoder_layers=16,
decoder_hidden_dim=512,
decoder_heads=16,
decoder_dim_feedforward=2048,
)
def vit_b_16_audio_mae() -> MaskedAutoEncoder:
return audio_mae(
input_size=(1024, 128),
patch_size=16,
masking_ratio=0.8,
encoder_layers=12,
encoder_hidden_dim=768,
encoder_heads=12,
encoder_dim_feedforward=3072,
decoder_layers=16,
decoder_hidden_dim=512,
decoder_heads=16,
decoder_dim_feedforward=2048,
)
def vit_l_16_audio_mae() -> MaskedAutoEncoder:
return audio_mae(
input_size=(1024, 128),
patch_size=16,
masking_ratio=0.8,
encoder_layers=24,
encoder_hidden_dim=1024,
encoder_heads=16,
encoder_dim_feedforward=4096,
decoder_layers=16,
decoder_hidden_dim=512,
decoder_heads=16,
decoder_dim_feedforward=2048,
)
def vit_b_16_audio_mae_encoder(pretrained: bool = False) -> VisionTransformer:
ckpt_path = MAE_MODEL_MAPPING["vit_b16_audio"] if pretrained else None
return vit_b_16(
final_layer_norm_eps=None,
num_channels=1,
image_size=(1024, 128),
ckpt_path=ckpt_path,
)