-
Notifications
You must be signed in to change notification settings - Fork 2.1k
/
pretrained_models.py
233 lines (207 loc) · 8.1 KB
/
pretrained_models.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
import os
import hashlib
import shutil
import urllib
_OPENCLIP_S3_BUCKET = 'https://clip-as-service.s3.us-east-2.amazonaws.com/models/torch'
_OPENCLIP_MODELS = {
'RN50::openai': ('RN50.pt', '9140964eaaf9f68c95aa8df6ca13777c'),
'RN50::yfcc15m': ('RN50-yfcc15m.pt', 'e9c564f91ae7dc754d9043fdcd2a9f22'),
'RN50::cc12m': ('RN50-cc12m.pt', '37cb01eb52bb6efe7666b1ff2d7311b5'),
'RN101::openai': ('RN101.pt', 'fa9d5f64ebf152bc56a18db245071014'),
'RN101::yfcc15m': ('RN101-yfcc15m.pt', '48f7448879ce25e355804f6bb7928cb8'),
'RN50x4::openai': ('RN50x4.pt', '03830990bc768e82f7fb684cde7e5654'),
'RN50x16::openai': ('RN50x16.pt', '83d63878a818c65d0fb417e5fab1e8fe'),
'RN50x64::openai': ('RN50x64.pt', 'a6631a0de003c4075d286140fc6dd637'),
'ViT-B-32::openai': ('ViT-B-32.pt', '3ba34e387b24dfe590eeb1ae6a8a122b'),
'ViT-B-32::laion2b_e16': (
'ViT-B-32-laion2b_e16.pt',
'df08de3d9f2dc53c71ea26e184633902',
),
'ViT-B-32::laion400m_e31': (
'ViT-B-32-laion400m_e31.pt',
'ca8015f98ab0f8780510710681d7b73e',
),
'ViT-B-32::laion400m_e32': (
'ViT-B-32-laion400m_e32.pt',
'359e0dba4a419f175599ee0c63a110d8',
),
'ViT-B-32::laion2b-s34b-b79k': (
'ViT-B-32-laion2b-s34b-b79k.bin',
'2fc036aea9cd7306f5ce7ce6abb8d0bf',
),
'ViT-B-16::openai': ('ViT-B-16.pt', '44c3d804ecac03d9545ac1a3adbca3a6'),
'ViT-B-16::laion400m_e31': (
'ViT-B-16-laion400m_e31.pt',
'31306a44224cc46fec1bc3b82fd0c4e6',
),
'ViT-B-16::laion400m_e32': (
'ViT-B-16-laion400m_e32.pt',
'07283adc5c17899f2ed22d82b563c54b',
),
'ViT-B-16-plus-240::laion400m_e31': (
'ViT-B-16-plus-240-laion400m_e31.pt',
'c88f453644a998ecb094d878a2f0738d',
),
'ViT-B-16-plus-240::laion400m_e32': (
'ViT-B-16-plus-240-laion400m_e32.pt',
'e573af3cef888441241e35022f30cc95',
),
'ViT-L-14::openai': ('ViT-L-14.pt', '096db1af569b284eb76b3881534822d9'),
'ViT-L-14::laion400m_e31': (
'ViT-L-14-laion400m_e31.pt',
'09d223a6d41d2c5c201a9da618d833aa',
),
'ViT-L-14::laion400m_e32': (
'ViT-L-14-laion400m_e32.pt',
'a76cde1bc744ca38c6036b920c847a89',
),
'ViT-L-14::laion2b-s32b-b82k': (
'ViT-L-14-laion2b-s32b-b82k.bin',
'4d2275fc7b2d7ee9db174f9b57ddecbd',
),
'ViT-L-14-336::openai': ('ViT-L-14-336px.pt', 'b311058cae50cb10fbfa2a44231c9473'),
'ViT-H-14::laion2b-s32b-b79k': (
'ViT-H-14-laion2b-s32b-b79k.bin',
'2aa6c46521b165a0daeb8cdc6668c7d3',
),
'ViT-g-14::laion2b-s12b-b42k': (
'ViT-g-14-laion2b-s12b-b42k.bin',
'3bf99353f6f1829faac0bb155be4382a',
),
'roberta-ViT-B-32::laion2b-s12b-b32k': (
'roberta-ViT-B-32-laion2b-s12b-b32k.bin',
'76d4c9d13774cc15fa0e2b1b94a8402c',
),
'xlm-roberta-base-ViT-B-32::laion5b-s13b-b90k': (
'xlm-roberta-base-ViT-B-32-laion5b-s13b-b90k.bin',
'f68abc07ef349720f1f880180803142d',
),
'xlm-roberta-large-ViT-H-14::frozen_laion5b_s13b_b90k': (
'xlm-roberta-large-ViT-H-14-frozen_laion5b_s13b_b90k.bin',
'b49991239a419d704fdba59c42d5536d',
),
# older version name format
'RN50': ('RN50.pt', '9140964eaaf9f68c95aa8df6ca13777c'),
'RN101': ('RN101.pt', 'fa9d5f64ebf152bc56a18db245071014'),
'RN50x4': ('RN50x4.pt', '03830990bc768e82f7fb684cde7e5654'),
'RN50x16': ('RN50x16.pt', '83d63878a818c65d0fb417e5fab1e8fe'),
'RN50x64': ('RN50x64.pt', 'a6631a0de003c4075d286140fc6dd637'),
'ViT-B/32': ('ViT-B-32.pt', '3ba34e387b24dfe590eeb1ae6a8a122b'),
'ViT-B/16': ('ViT-B-16.pt', '44c3d804ecac03d9545ac1a3adbca3a6'),
'ViT-L/14': ('ViT-L-14.pt', '096db1af569b284eb76b3881534822d9'),
'ViT-L/14@336px': ('ViT-L-14-336px.pt', 'b311058cae50cb10fbfa2a44231c9473'),
}
_MULTILINGUALCLIP_MODELS = {
'M-CLIP/XLM-Roberta-Large-Vit-B-32': (),
'M-CLIP/XLM-Roberta-Large-Vit-L-14': (),
'M-CLIP/XLM-Roberta-Large-Vit-B-16Plus': (),
'M-CLIP/LABSE-Vit-L-14': (),
}
_VISUAL_MODEL_IMAGE_SIZE = {
'RN50': 224,
'RN101': 224,
'RN50x4': 288,
'RN50x16': 384,
'RN50x64': 448,
'ViT-B-32': 224,
'roberta-ViT-B-32': 224,
'xlm-roberta-base-ViT-B-32': 224,
'ViT-B-16': 224,
'Vit-B-16Plus': 240,
'ViT-B-16-plus-240': 240,
'ViT-L-14': 224,
'ViT-L-14-336': 336,
'ViT-H-14': 224,
'xlm-roberta-large-ViT-H-14': 224,
'ViT-g-14': 224,
}
def md5file(filename: str):
hash_md5 = hashlib.md5()
with open(filename, 'rb') as f:
for chunk in iter(lambda: f.read(4096), b""):
hash_md5.update(chunk)
return hash_md5.hexdigest()
def get_model_url_md5(name: str):
model_pretrained = _OPENCLIP_MODELS[name]
if len(model_pretrained) == 0: # not on s3
return None, None
else:
return (_OPENCLIP_S3_BUCKET + '/' + model_pretrained[0], model_pretrained[1])
def download_model(
url: str,
target_folder: str = os.path.expanduser("~/.cache/clip"),
md5sum: str = None,
with_resume: bool = True,
max_attempts: int = 3,
) -> str:
os.makedirs(target_folder, exist_ok=True)
filename = os.path.basename(url)
download_target = os.path.join(target_folder, filename)
if os.path.exists(download_target):
if not os.path.isfile(download_target):
raise FileExistsError(f'{download_target} exists and is not a regular file')
actual_md5sum = md5file(download_target)
if (not md5sum) or actual_md5sum == md5sum:
return download_target
from rich.progress import (
DownloadColumn,
Progress,
TextColumn,
TimeRemainingColumn,
TransferSpeedColumn,
)
progress = Progress(
" \n", # divide this bar from Flow's bar
TextColumn("[bold blue]{task.fields[filename]}", justify="right"),
"[progress.percentage]{task.percentage:>3.1f}%",
"•",
DownloadColumn(),
"•",
TransferSpeedColumn(),
"•",
TimeRemainingColumn(),
)
with progress:
task = progress.add_task('download', filename=filename, start=False)
for _ in range(max_attempts):
tmp_file_path = download_target + '.part'
resume_byte_pos = (
os.path.getsize(tmp_file_path) if os.path.exists(tmp_file_path) else 0
)
try:
# resolve the 403 error by passing a valid user-agent
req = urllib.request.Request(url, headers={'User-Agent': 'Mozilla/5.0'})
total_bytes = int(
urllib.request.urlopen(req).info().get('Content-Length', -1)
)
mode = 'ab' if (with_resume and resume_byte_pos) else 'wb'
with open(tmp_file_path, mode) as output:
progress.update(task, total=total_bytes)
progress.start_task(task)
if resume_byte_pos and with_resume:
progress.update(task, advance=resume_byte_pos)
req.headers['Range'] = f'bytes={resume_byte_pos}-'
with urllib.request.urlopen(req) as source:
while True:
buffer = source.read(8192)
if not buffer:
break
output.write(buffer)
progress.update(task, advance=len(buffer))
actual_md5 = md5file(tmp_file_path)
if (md5sum and actual_md5 == md5sum) or (not md5sum):
shutil.move(tmp_file_path, download_target)
return download_target
else:
os.remove(tmp_file_path)
raise RuntimeError(
f'MD5 mismatch: expected {md5sum}, got {actual_md5}'
)
except Exception as ex:
progress.console.print(
f'Failed to download {url} with {ex!r} at the {_}th attempt'
)
progress.reset(task)
raise RuntimeError(
f'Failed to download {url} within retry limit {max_attempts}'
)