-
Notifications
You must be signed in to change notification settings - Fork 14
/
utils.py
364 lines (315 loc) · 14.4 KB
/
utils.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
import os
import json
import torch
import math
from torch import nn
from typing import List
from transformers import BertTokenizer
from urllib.parse import urlparse
from timm.models.hub import download_cached_file
from .vit import interpolate_pos_embed
from .swin_transformer import interpolate_relative_pos_embed
from pathlib import Path
CONFIG_PATH=(Path(__file__).resolve().parents[1])
def read_json(rpath):
with open(rpath, 'r') as f:
return json.load(f)
def tie_encoder_decoder_weights(encoder: nn.Module, decoder: nn.Module,
base_model_prefix: str, skip_key: str):
uninitialized_encoder_weights: List[str] = []
if decoder.__class__ != encoder.__class__:
logger.info(
f"{decoder.__class__} and {encoder.__class__} are not equal. In this case make sure that all encoder weights are correctly initialized."
)
def tie_encoder_to_decoder_recursively(
decoder_pointer: nn.Module,
encoder_pointer: nn.Module,
module_name: str,
uninitialized_encoder_weights: List[str],
skip_key: str,
depth=0,
):
assert isinstance(decoder_pointer, nn.Module) and isinstance(
encoder_pointer, nn.Module
), f"{decoder_pointer} and {encoder_pointer} have to be of type torch.nn.Module"
if hasattr(decoder_pointer, "weight") and skip_key not in module_name:
assert hasattr(encoder_pointer, "weight")
encoder_pointer.weight = decoder_pointer.weight
if hasattr(decoder_pointer, "bias"):
assert hasattr(encoder_pointer, "bias")
encoder_pointer.bias = decoder_pointer.bias
print(module_name + ' is tied')
return
encoder_modules = encoder_pointer._modules
decoder_modules = decoder_pointer._modules
if len(decoder_modules) > 0:
assert (
len(encoder_modules) > 0
), f"Encoder module {encoder_pointer} does not match decoder module {decoder_pointer}"
all_encoder_weights = set([
module_name + "/" + sub_name
for sub_name in encoder_modules.keys()
])
encoder_layer_pos = 0
for name, module in decoder_modules.items():
if name.isdigit():
encoder_name = str(int(name) + encoder_layer_pos)
decoder_name = name
if not isinstance(
decoder_modules[decoder_name],
type(encoder_modules[encoder_name])) and len(
encoder_modules) != len(decoder_modules):
# this can happen if the name corresponds to the position in a list module list of layers
# in this case the decoder has added a cross-attention that the encoder does not have
# thus skip this step and subtract one layer pos from encoder
encoder_layer_pos -= 1
continue
elif name not in encoder_modules:
continue
elif depth > 500:
raise ValueError(
"Max depth of recursive function `tie_encoder_to_decoder` reached. It seems that there is a circular dependency between two or more `nn.Modules` of your model."
)
else:
decoder_name = encoder_name = name
tie_encoder_to_decoder_recursively(
decoder_modules[decoder_name],
encoder_modules[encoder_name],
module_name + "/" + name,
uninitialized_encoder_weights,
skip_key,
depth=depth + 1,
)
all_encoder_weights.remove(module_name + "/" + encoder_name)
uninitialized_encoder_weights += list(all_encoder_weights)
# tie weights recursively
tie_encoder_to_decoder_recursively(decoder, encoder, base_model_prefix,
uninitialized_encoder_weights, skip_key)
class GroupWiseLinear(nn.Module):
# could be changed to:
# output = torch.einsum('ijk,zjk->ij', x, self.W)
# or output = torch.einsum('ijk,jk->ij', x, self.W[0])
def __init__(self, num_class, hidden_dim, bias=True):
super().__init__()
self.num_class = num_class
self.hidden_dim = hidden_dim
self.bias = bias
self.W = nn.Parameter(torch.Tensor(1, num_class, hidden_dim))
if bias:
self.b = nn.Parameter(torch.Tensor(1, num_class))
self.reset_parameters()
def reset_parameters(self):
stdv = 1. / math.sqrt(self.W.size(2))
for i in range(self.num_class):
self.W[0][i].data.uniform_(-stdv, stdv)
if self.bias:
for i in range(self.num_class):
self.b[0][i].data.uniform_(-stdv, stdv)
def forward(self, x):
# x: B,K,d
x = (self.W * x).sum(-1)
if self.bias:
x = x + self.b
return x
def init_tokenizer():
tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
tokenizer.add_special_tokens({'bos_token': '[DEC]'})
tokenizer.add_special_tokens({'additional_special_tokens': ['[ENC]']})
tokenizer.enc_token_id = tokenizer.additional_special_tokens_ids[0]
return tokenizer
def create_vit(vit,
image_size,
use_grad_checkpointing=False,
ckpt_layer=0,
drop_path_rate=0):
assert vit in ['base', 'large'], "vit parameter must be base or large"
if vit == 'base':
vision_width = 768
visual_encoder = VisionTransformer(
img_size=image_size,
patch_size=16,
embed_dim=vision_width,
depth=12,
num_heads=12,
use_grad_checkpointing=use_grad_checkpointing,
ckpt_layer=ckpt_layer,
drop_path_rate=0 or drop_path_rate)
elif vit == 'large':
vision_width = 1024
visual_encoder = VisionTransformer(
img_size=image_size,
patch_size=16,
embed_dim=vision_width,
depth=24,
num_heads=16,
use_grad_checkpointing=use_grad_checkpointing,
ckpt_layer=ckpt_layer,
drop_path_rate=0.1 or drop_path_rate)
return visual_encoder, vision_width
def is_url(url_or_filename):
parsed = urlparse(url_or_filename)
return parsed.scheme in ("http", "https")
def load_checkpoint(model, url_or_filename):
if is_url(url_or_filename):
cached_file = download_cached_file(url_or_filename,
check_hash=False,
progress=True)
checkpoint = torch.load(cached_file, map_location='cpu')
elif os.path.isfile(url_or_filename):
checkpoint = torch.load(url_or_filename, map_location='cpu')
else:
raise RuntimeError('checkpoint url or path is invalid')
state_dict = checkpoint['model']
state_dict['visual_encoder.pos_embed'] = interpolate_pos_embed(
state_dict['visual_encoder.pos_embed'], model.visual_encoder)
if 'visual_encoder_m.pos_embed' in model.state_dict().keys():
state_dict['visual_encoder_m.pos_embed'] = interpolate_pos_embed(
state_dict['visual_encoder_m.pos_embed'], model.visual_encoder_m)
for key in model.state_dict().keys():
if key in state_dict.keys():
if state_dict[key].shape != model.state_dict()[key].shape:
del state_dict[key]
msg = model.load_state_dict(state_dict, strict=False)
print('load checkpoint from %s' % url_or_filename)
return model, msg
# def load_checkpoint_condition(model, url_or_filename):
def load_checkpoint_swinlarge_condition(model, url_or_filename, kwargs):
if kwargs['image_size'] == 224:
vision_config_path = f'{CONFIG_PATH}/configs/swin/config_swinL_224.json'
elif kwargs['image_size'] == 384:
vision_config_path = f'{CONFIG_PATH}/configs/swin/config_swinL_384.json'
window_size = read_json(vision_config_path)['window_size']
print('--------------')
print(url_or_filename)
print('--------------')
if is_url(url_or_filename):
cached_file = download_cached_file(url_or_filename,
check_hash=False,
progress=True)
checkpoint = torch.load(cached_file, map_location='cpu')
elif os.path.isfile(url_or_filename):
checkpoint = torch.load(url_or_filename, map_location='cpu')
else:
raise RuntimeError('checkpoint url or path is invalid')
state_dict = checkpoint['params']
for k in list(state_dict.keys()):
if 'relative_position_bias_table' in k:
dst_num_pos = (2 * window_size - 1)**2
state_dict[k] = interpolate_relative_pos_embed(state_dict[k],
dst_num_pos,
param_name=k)
elif ('relative_position_index' in k) or ('attn_mask' in k):
del state_dict[k]
elif "vision_multi" in k:
state_dict[k.replace("vision_multi",
"tagging_head")] = state_dict.pop(k)
msg = model.load_state_dict(state_dict, strict=False)
print('load checkpoint from %s' % url_or_filename)
return model, msg
def load_checkpoint_swinbase(model, url_or_filename, kwargs):
if kwargs['image_size'] == 224:
vision_config_path = f'{CONFIG_PATH}/configs/swin/config_swinB_224.json'
elif kwargs['image_size'] == 384:
vision_config_path = f'{CONFIG_PATH}/configs/swin/config_swinB_384.json'
window_size = read_json(vision_config_path)['window_size']
print('--------------')
print(url_or_filename)
print('--------------')
if is_url(url_or_filename):
cached_file = download_cached_file(url_or_filename,
check_hash=False,
progress=True)
checkpoint = torch.load(cached_file, map_location='cpu')
elif os.path.isfile(url_or_filename):
checkpoint = torch.load(url_or_filename, map_location='cpu')
else:
raise RuntimeError('checkpoint url or path is invalid')
state_dict = checkpoint['model']
for k in list(state_dict.keys()):
if 'relative_position_bias_table' in k:
dst_num_pos = (2 * window_size - 1)**2
state_dict[k] = interpolate_relative_pos_embed(state_dict[k],
dst_num_pos,
param_name=k)
elif ('relative_position_index' in k) or ('attn_mask' in k):
del state_dict[k]
elif "vision_multi" in k:
state_dict[k.replace("vision_multi",
"tagging_head")] = state_dict.pop(k)
msg = model.load_state_dict(state_dict, strict=False)
print('load checkpoint from %s' % url_or_filename)
return model, msg
def load_checkpoint_swinlarge(model, url_or_filename, kwargs):
if kwargs['image_size'] == 224:
vision_config_path = f'{CONFIG_PATH}/configs/swin/config_swinL_224.json'
elif kwargs['image_size'] == 384:
vision_config_path = f'{CONFIG_PATH}/configs/swin/config_swinL_384.json'
window_size = read_json(vision_config_path)['window_size']
print('--------------')
print(url_or_filename)
print('--------------')
if is_url(url_or_filename):
cached_file = download_cached_file(url_or_filename,
check_hash=False,
progress=True)
checkpoint = torch.load(cached_file, map_location='cpu')
elif os.path.isfile(url_or_filename):
checkpoint = torch.load(url_or_filename, map_location='cpu')
else:
raise RuntimeError('checkpoint url or path is invalid')
state_dict = checkpoint['model']
for k in list(state_dict.keys()):
if 'relative_position_bias_table' in k:
dst_num_pos = (2 * window_size - 1)**2
state_dict[k] = interpolate_relative_pos_embed(state_dict[k],
dst_num_pos,
param_name=k)
elif ('relative_position_index' in k) or ('attn_mask' in k):
del state_dict[k]
elif "vision_multi" in k:
state_dict[k.replace("vision_multi",
"tagging_head")] = state_dict.pop(k)
msg = model.load_state_dict(state_dict, strict=False)
print('load checkpoint from %s' % url_or_filename)
return model, msg
# Tagging loss function
# copy from https://github.com/Alibaba-MIIL/ASL/blob/main/src/loss_functions/losses.py
class AsymmetricLoss(nn.Module):
def __init__(self, gamma_neg=4, gamma_pos=1, clip=0.05, eps=1e-8, disable_torch_grad_focal_loss=True):
super(AsymmetricLoss, self).__init__()
self.gamma_neg = gamma_neg
self.gamma_pos = gamma_pos
self.clip = clip
self.disable_torch_grad_focal_loss = disable_torch_grad_focal_loss
self.eps = eps
def forward(self, x, y):
""""
Parameters
----------
x: input logits
y: targets (multi-label binarized vector)
"""
# Calculating Probabilities
x_sigmoid = torch.sigmoid(x)
xs_pos = x_sigmoid
xs_neg = 1 - x_sigmoid
# Asymmetric Clipping
if self.clip is not None and self.clip > 0:
xs_neg = (xs_neg + self.clip).clamp(max=1)
# Basic CE calculation
los_pos = y * torch.log(xs_pos.clamp(min=self.eps))
los_neg = (1 - y) * torch.log(xs_neg.clamp(min=self.eps))
loss = los_pos + los_neg
# Asymmetric Focusing
if self.gamma_neg > 0 or self.gamma_pos > 0:
if self.disable_torch_grad_focal_loss:
torch.set_grad_enabled(False)
pt0 = xs_pos * y
pt1 = xs_neg * (1 - y) # pt = p if t > 0 else 1-p
pt = pt0 + pt1
one_sided_gamma = self.gamma_pos * y + self.gamma_neg * (1 - y)
one_sided_w = torch.pow(1 - pt, one_sided_gamma)
if self.disable_torch_grad_focal_loss:
torch.set_grad_enabled(True)
loss *= one_sided_w
return -loss.sum()