File size: 9,243 Bytes
37b3db0
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
# Copyright (c) Meta Platforms, Inc. and affiliates

import argparse


def get_default_params(model_name):
    # Params from paper (https://arxiv.org/pdf/2103.00020.pdf)
    model_name = model_name.lower()
    if "vit" in model_name:
        return {"lr": 5.0e-4, "beta1": 0.9, "beta2": 0.98, "eps": 1.0e-6}
    else:
        return {"lr": 5.0e-4, "beta1": 0.9, "beta2": 0.999, "eps": 1.0e-8}


def parse_args():
    parser = argparse.ArgumentParser()
    parser.add_argument(
        "--train-data",
        type=str,
        default=None,
        help="Path to csv filewith training data",
    )
    parser.add_argument(
        "--val-data",
        type=str,
        default=None,
        help="Path to csv file with validation data",
    )
    parser.add_argument(
        "--train-num-samples",
        type=int,
        default=None,
        help="Number of samples in dataset. Required for webdataset if not available in info file.",
    )
    parser.add_argument(
        "--val-num-samples",
        type=int,
        default=None,
        help="Number of samples in dataset. Useful for webdataset if not available in info file.",
    )
    parser.add_argument(
        "--dataset-type",
        choices=["webdataset", "csv", "auto"],
        default="auto",
        help="Which type of dataset to process."
    )
    parser.add_argument(
        "--dataset-resampled",
        default=False,
        action="store_true",
        help="Whether to use sampling with replacement for webdataset shard selection."
    )
    parser.add_argument(
        "--csv-separator",
        type=str,
        default="\t",
        help="For csv-like datasets, which separator to use."
    )
    parser.add_argument(
        "--csv-img-key",
        type=str,
        default="filepath",
        help="For csv-like datasets, the name of the key for the image paths."
    )
    parser.add_argument(
        "--csv-caption-key",
        type=str,
        default="title",
        help="For csv-like datasets, the name of the key for the captions."
    )
    parser.add_argument(
        "--imagenet-val",
        type=str,
        default=None,
        help="Path to imagenet val set for conducting zero shot evaluation.",
    )
    parser.add_argument(
        "--imagenet-v2",
        type=str,
        default=None,
        help="Path to imagenet v2 for conducting zero shot evaluation.",
    )
    parser.add_argument(
        "--logs",
        type=str,
        default="./logs/",
        help="Where to store tensorboard logs. Use None to avoid storing logs.",
    )
    parser.add_argument(
        "--log-local",
        action="store_true",
        default=False,
        help="log files on local master, otherwise global master only.",
    )
    parser.add_argument(
        "--name",
        type=str,
        default=None,
        help="Optional identifier for the experiment when storing logs. Otherwise use current time.",
    )
    parser.add_argument(
        "--workers", type=int, default=1, help="Number of dataloader workers per GPU."
    )
    parser.add_argument(
        "--batch-size", type=int, default=64, help="Batch size per GPU."
    )
    parser.add_argument(
        "--epochs", type=int, default=32, help="Number of epochs to train for."
    )
    parser.add_argument("--lr", type=float, default=None, help="Learning rate.")
    parser.add_argument("--beta1", type=float, default=None, help="Adam beta 1.")
    parser.add_argument("--beta2", type=float, default=None, help="Adam beta 2.")
    parser.add_argument("--eps", type=float, default=None, help="Adam epsilon.")
    parser.add_argument("--wd", type=float, default=0.2, help="Weight decay.")
    parser.add_argument(
        "--warmup", type=int, default=10000, help="Number of steps to warmup for."
    )
    parser.add_argument(
        "--use-bn-sync",
        default=False,
        action="store_true",
        help="Whether to use batch norm sync.")
    parser.add_argument(
        "--skip-scheduler",
        action="store_true",
        default=False,
        help="Use this flag to skip the learning rate decay.",
    )
    parser.add_argument(
        "--save-frequency", type=int, default=1, help="How often to save checkpoints."
    )
    parser.add_argument(
        "--save-most-recent",
        action="store_true",
        default=False,
        help="Always save the most recent model trained to epoch_latest.pt.",
    )
    parser.add_argument(
        "--zeroshot-frequency", type=int, default=2, help="How often to run zero shot."
    )
    parser.add_argument(
        "--val-frequency", type=int, default=1, help="How often to run evaluation with val data."
    )
    parser.add_argument(
        "--resume",
        default=None,
        type=str,
        help="path to latest checkpoint (default: none)",
    )
    parser.add_argument(
        "--precision",
        choices=["amp", "fp16", "fp32"],
        default="amp",
        help="Floating point precision."
    )
    parser.add_argument(
        "--model",
        type=str,
        default="RN50",
        help="Name of the vision backbone to use.",
    )
    parser.add_argument(
        "--pretrained",
        default='',
        type=str,
        help="Use a pretrained CLIP model weights with the specified tag or file path.",
    )
    parser.add_argument(
        "--pretrained-image",
        default=False,
        action='store_true',
        help="Load imagenet pretrained weights for image tower backbone if available.",
    )
    parser.add_argument(
        "--lock-image",
        default=False,
        action='store_true',
        help="Lock full image tower by disabling gradients.",
    )
    parser.add_argument(
        "--lock-image-unlocked-groups",
        type=int,
        default=0,
        help="Leave last n image tower layer groups unlocked.",
    )
    parser.add_argument(
        "--lock-image-freeze-bn-stats",
        default=False,
        action='store_true',
        help="Freeze BatchNorm running stats in image tower for any locked layers.",
    )
    parser.add_argument(
        "--grad-checkpointing",
        default=False,
        action='store_true',
        help="Enable gradient checkpointing.",
    )
    parser.add_argument(
        "--local-loss",
        default=False,
        action="store_true",
        help="calculate loss w/ local features @ global (instead of realizing full global @ global matrix)"
    )
    parser.add_argument(
        "--gather-with-grad",
        default=False,
        action="store_true",
        help="enable full distributed gradient for feature gather"
    )
    parser.add_argument(
        "--force-quick-gelu",
        default=False,
        action='store_true',
        help="Force use of QuickGELU activation for non-OpenAI transformer models.",
    )
    parser.add_argument(
        "--torchscript",
        default=False,
        action='store_true',
        help="torch.jit.script the model, also uses jit version of OpenAI models if pretrained=='openai'",
    )
    parser.add_argument(
        "--trace",
        default=False,
        action='store_true',
        help="torch.jit.trace the model for inference / eval only",
    )
    # arguments for distributed training
    parser.add_argument(
        "--dist-url",
        default="env://",
        type=str,
        help="url used to set up distributed training",
    )
    parser.add_argument(
        "--dist-backend", default="nccl", type=str, help="distributed backend"
    )
    parser.add_argument(
        "--report-to",
        default='',
        type=str,
        help="Options are ['wandb', 'tensorboard', 'wandb,tensorboard']"
    )
    parser.add_argument(
        "--wandb-notes",
        default='',
        type=str,
        help="Notes if logging with wandb"
    )
    parser.add_argument(
        "--debug",
        default=False,
        action="store_true",
        help="If true, more information is logged."
    )
    parser.add_argument(
        "--copy-codebase",
        default=False,
        action="store_true",
        help="If true, we copy the entire base on the log diretory, and execute from there."
    )
    parser.add_argument(
        "--horovod",
        default=False,
        action="store_true",
        help="Use horovod for distributed training."
    )
    parser.add_argument(
        "--ddp-static-graph",
        default=False,
        action='store_true',
        help="Enable static graph optimization for DDP in PyTorch >= 1.11.",
    )
    parser.add_argument(
        "--no-set-device-rank",
        default=False,
        action="store_true",
        help="Don't set device index from local rank (when CUDA_VISIBLE_DEVICES restricted to one per proc)."
    )
    parser.add_argument(
        "--seed", type=int, default=0, help="Default random seed."
    )
    parser.add_argument(
        "--norm_gradient_clip", type=float, default=None, help="Gradient clip."
    )
    args = parser.parse_args()

    # If some params are not passed, we use the default values based on model name.
    default_params = get_default_params(args.model)
    for name, val in default_params.items():
        if getattr(args, name) is None:
            setattr(args, name, val)

    return args