-
Notifications
You must be signed in to change notification settings - Fork 37
/
transfer_weights.py
executable file
·210 lines (179 loc) · 7.25 KB
/
transfer_weights.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
from timm.models.layers import config
from glasses.models.AutoTransform import AutoTransform
import logging
from argparse import ArgumentParser
from dataclasses import dataclass
from functools import partial
from io import BytesIO
from pathlib import Path
from typing import Dict
from glasses.utils.storage import LocalStorage, HuggingFaceStorage
import pretrainedmodels
import timm
import torch
from torch import Tensor, nn
from torchvision.models import (
densenet121,
densenet161,
densenet169,
densenet201,
resnet18,
resnet50,
resnet101,
resnet152,
resnext50_32x4d,
resnext101_32x8d,
vgg11,
vgg13,
vgg16,
vgg19,
wide_resnet50_2,
wide_resnet101_2,
mobilenetv2,
)
from tqdm.autonotebook import tqdm
from glasses.models.AutoModel import AutoModel
from glasses.models import *
from glasses.utils.ModuleTransfer import ModuleTransfer
from glasses.models.classification.vit import ViTTokens
from glasses.models.classification.deit import DeiTTokens
def vit_clone(key: str):
src = timm.create_model(key, pretrained="True")
dst = AutoModel.from_name(key)
cfg = AutoTransform.from_name(key)
dst = clone_model(
src,
dst,
torch.randn((1, 3, cfg.input_size, cfg.input_size)),
dest_skip=[ViTTokens],
)
dst.embedding.positions.data.copy_(src.pos_embed.data.squeeze(0))
dst.embedding.tokens.cls.data.copy_(src.cls_token.data)
return dst
def deit_clone(key: str):
k_split = key.split("_")
hub_key = "_".join(k_split[:2]) + "_distilled_" + "_".join(k_split[2:])
src = torch.hub.load("facebookresearch/deit:main", hub_key, pretrained=True)
dst = AutoModel.from_name(key)
cfg = AutoTransform.from_name(f"vit_{'_'.join(key.split('_')[1:])}")
dst = clone_model(
src,
dst,
torch.randn((1, 3, cfg.input_size, cfg.input_size)),
dest_skip=[DeiTTokens],
)
dst.embedding.positions.data.copy_(src.pos_embed.data.squeeze(0))
dst.embedding.tokens.cls.data.copy_(src.cls_token.data)
dst.embedding.tokens.dist.data.copy_(src.dist_token.data)
return dst
zoo_source = {
# "resnet18": partial(resnet18, pretrained=True),
# "resnet26": partial(timm.create_model, "resnet26", pretrained=True),
# "resnet26d": partial(timm.create_model, "resnet26d", pretrained=True),
# "resnet34": partial(timm.create_model, "resnet34", pretrained=True),
# "resnet34d": partial(timm.create_model, "resnet34d", pretrained=True),
# "resnet50": partial(resnet50, pretrained=True),
# "resnet50d": partial(timm.create_model, "resnet50d", pretrained=True),
# "resnet101": partial(resnet101, pretrained=True),
# "resnet152": partial(resnet152, pretrained=True),
# "se_resnet50": partial(timm.create_model, "seresnet50", pretrained=True),
# "resnext50_32x4d": partial(resnext50_32x4d, pretrained=True),
# "resnext101_32x8d": partial(resnext101_32x8d, pretrained=True),
# "wide_resnet50_2": partial(wide_resnet50_2, pretrained=True),
# "wide_resnet101_2": partial(wide_resnet101_2, pretrained=True),
# "eca_resnet26t": partial(timm.create_model, "ecaresnet26t", pretrained=True),
# "eca_resnet50t": partial(timm.create_model, "ecaresnet50t", pretrained=True),
# "eca_resnet50d": partial(timm.create_model, "ecaresnet50d", pretrained=True),
# "eca_resnet101d": partial(timm.create_model, "ecaresnet101d", pretrained=True),
# "regnetx_002": None,
# "regnetx_004": None,
# "regnetx_006": None,
# "regnetx_008": None,
# "regnetx_016": None,
# "regnetx_032": None,
# "regnetx_040": None,
# "regnetx_064": None,
# "regnety_002": None,
# "regnety_004": None,
# "regnety_006": None,
# "regnety_008": None,
# "regnety_016": None,
"regnety_032": None,
"regnety_040": None,
"regnety_064": None,
# "densenet121": partial(densenet121, pretrained=True),
# "densenet169": partial(densenet169, pretrained=True),
# "densenet201": partial(densenet201, pretrained=True),
# "densenet161": partial(densenet161, pretrained=True),
# "vgg11": partial(vgg11, pretrained=True),
# "vgg13": partial(vgg13, pretrained=True),
# "vgg16": partial(vgg16, pretrained=True),
# "vgg19": partial(vgg19, pretrained=True),
# "vgg11_bn": pretrainedmodels.__dict__["vgg11_bn"],
# "vgg13_bn": pretrainedmodels.__dict__["vgg13_bn"],
# "vgg16_bn": pretrainedmodels.__dict__["vgg16_bn"],
# "vgg19_bn": pretrainedmodels.__dict__["vgg19_bn"],
# "efficientnet_b0": partial(timm.create_model, "efficientnet_b0", pretrained=True),
# "efficientnet_b1": partial(timm.create_model, "efficientnet_b1", pretrained=True),
# "efficientnet_b2": partial(timm.create_model, "efficientnet_b2", pretrained=True),
# "efficientnet_b3": partial(timm.create_model, "efficientnet_b3", pretrained=True),
# "efficientnet_lite0": partial(
# timm.create_model, "efficientnet_lite0", pretrained=True
# ),
# # "mobilenet_v2": partial(mobilenetv2, pretrained=True),
# "vit_base_patch16_224": (vit_clone, True),
# "vit_base_patch16_384": (vit_clone, True),
# "vit_base_patch32_384": (vit_clone, True),
# "vit_huge_patch16_224": (vit_clone, True),
# "vit_huge_patch32_384": (vit_clone, True),
# "vit_large_patch16_224": (vit_clone, True),
# "vit_large_patch16_384": (vit_clone, True),
# "vit_large_patch32_384": (vit_clone, True),
# "deit_tiny_patch16_224": (deit_clone, True),
# "deit_small_patch16_224": (deit_clone, True),
# "deit_base_patch16_224": (deit_clone, True),
# "deit_base_patch16_384": (deit_clone, True),
}
def clone_model(
src: nn.Module, dst: nn.Module, x: Tensor = torch.rand((1, 3, 224, 224)), **kwargs
) -> nn.Module:
src = src.eval()
dst = dst.eval()
a = src(x)
b = dst(x)
ModuleTransfer(src, dst, **kwargs)(x)
return dst
if __name__ == "__main__":
parser = ArgumentParser()
parser.add_argument("--storage", type=str, choices=["local", "hf"], default="hf")
parser.add_argument("-o", type=Path)
args = parser.parse_args()
logging.basicConfig(format="%(levelname)s:%(message)s", level=logging.INFO)
logging.info(f"Using {args.storage} storage 💾")
# store the pretrained names
with open("pretrained_models.txt", "w") as f:
f.write(",".join(list(zoo_source.keys())))
if args.o is not None:
save_dir = args.o
save_dir.mkdir(exist_ok=True)
storages = {"local": LocalStorage, "hf": HuggingFaceStorage}
storage = storages[args.storage]()
if args.storage == "local":
logging.info(f"Store root={storage.root}")
override = True
bar = tqdm(zoo_source.items())
uploading_bar = tqdm()
for key, src_def in bar:
bar.set_description(key)
if src_def is None:
# it means I was lazy and I meant to use timm
src_def = partial(timm.create_model, key, pretrained=True)
if key not in storage or override:
if type(src_def) is tuple:
# I have a custom clone func -> not the most elegant way, but it works!
clone_func, flag = src_def
cloned = clone_func(key)
else:
src, dst = src_def(), AutoModel.from_name(key)
cloned = clone_model(src, dst)
storage.put(key, cloned)