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
|
import argparse
import os
from dataclasses import dataclass
from pathlib import Path
from typing import Iterable, Callable
import sys
import torch
import torch.distributed as dist
import yaml
from torch.utils.data import Dataset
from torchvision.datasets import CIFAR10, CIFAR100, ImageNet
from torchvision.transforms import transforms
path = str(Path(Path(__file__).parent.absolute()).parent.absolute())
sys.path.insert(0, path)
from libs.criteria import InfoNCELoss
from libs.datautils import color_distortion, Clip, RandomGaussianBlur, TwinTransform
from libs.optimizers import LARS
from libs.utils import Trainer, BaseConfig, elastic_launch
from libs.logging import BaseBatchLogRecord, Loggers
from simclr.models import CIFARSimCLRResNet50, ImageNetSimCLRResNet50
def parse_args_and_config():
parser = argparse.ArgumentParser(
description='Contrastive baseline SimCLR',
formatter_class=argparse.ArgumentDefaultsHelpFormatter
)
parser.add_argument('--codename', default='cifar10-simclr-128-lars-warmup',
type=str, help="Model descriptor")
parser.add_argument('--log-dir', default='logs', type=str,
help="Path to log directory")
parser.add_argument('--checkpoint-dir', default='checkpoints', type=str,
help="Path to checkpoints directory")
parser.add_argument('--seed', default=None, type=int,
help='Random seed for reproducibility')
parser.add_argument('--num-iters', default=23438, type=int,
help='Number of iters (default is 50 epochs equiv., '
'around dataset_size * epochs / batch_size)')
parser.add_argument('--config', type=argparse.FileType(mode='r'),
help='Path to config file (optional)')
# TODO: Add model hyperparams dataclass
parser.add_argument('--hid-dim', default=2048, type=int,
help='Number of dimension of embedding')
parser.add_argument('--out-dim', default=128, type=int,
help='Number of dimension after projection')
parser.add_argument('--temp', default=0.5, type=float,
help='Temperature in InfoNCE loss')
dataset_group = parser.add_argument_group('Dataset parameters')
dataset_group.add_argument('--dataset-dir', default='dataset', type=str,
help="Path to dataset directory")
dataset_group.add_argument('--dataset', default='cifar10', type=str,
choices=('cifar10, cifar100', 'imagenet'),
help="Name of dataset")
dataset_group.add_argument('--crop-size', default=32, type=int,
help='Random crop size after resize')
dataset_group.add_argument('--crop-scale-range', nargs=2, default=(0.8, 1),
type=float, help='Random resize scale range',
metavar=('start', 'stop'))
dataset_group.add_argument('--hflip-prob', default=0.5, type=float,
help='Random horizontal flip probability')
dataset_group.add_argument('--distort-strength', default=0.5, type=float,
help='Distortion strength')
dataset_group.add_argument('--gauss-ker-scale', default=10, type=float,
help='Gaussian kernel scale factor '
'(s = img_size / ker_size)')
dataset_group.add_argument('--gauss-sigma-range', nargs=2, default=(0.1, 2),
type=float, help='Random gaussian blur sigma range',
metavar=('start', 'stop'))
dataset_group.add_argument('--gauss-prob', default=0.5, type=float,
help='Random gaussian blur probability')
dataloader_group = parser.add_argument_group('Dataloader parameters')
dataloader_group.add_argument('--batch-size', default=128, type=int,
help='Batch size')
dataloader_group.add_argument('--num-workers', default=2, type=int,
help='Number of dataloader processes')
optim_group = parser.add_argument_group('Optimizer parameters')
optim_group.add_argument('--optim', default='lars', type=str,
choices=('adam', 'sgd', 'lars'),
help="Name of optimizer")
optim_group.add_argument('--lr', default=1., type=float,
help='Learning rate')
optim_group.add_argument('--betas', nargs=2, default=(0.9, 0.999), type=float,
help='Adam betas', metavar=('beta1', 'beta2'))
optim_group.add_argument('--momentum', default=0.9, type=float,
help='SDG momentum')
optim_group.add_argument('--weight-decay', default=1e-6, type=float,
help='Weight decay (l2 regularization)')
sched_group = parser.add_argument_group('Scheduler parameters')
sched_group.add_argument('--sched', default='warmup-anneal', type=str,
choices=('const', None, 'linear', 'warmup-anneal'),
help="Name of scheduler")
sched_group.add_argument('--warmup-iters', default=2344, type=int,
help='Epochs for warmup (`warmup-anneal` scheduler only)')
args = parser.parse_args()
if args.config:
config = yaml.safe_load(args.config)
args.__dict__ |= {
k: tuple(v) if isinstance(v, list) else v
for k, v in config.items()
}
args.checkpoint_dir = os.path.join(args.checkpoint_dir, args.codename)
args.log_dir = os.path.join(args.log_dir, args.codename)
return args
@dataclass
class SimCLRConfig(BaseConfig):
@dataclass
class DatasetConfig(BaseConfig.DatasetConfig):
dataset_dir: str
crop_size: int
crop_scale_range: tuple[float, float]
hflip_prob: float
distort_strength: float
gauss_ker_scale: float
gauss_sigma_range: tuple[float, float]
gauss_prob: float
@dataclass
class OptimConfig(BaseConfig.OptimConfig):
momentum: float
betas: tuple[float, float]
weight_decay: float
class SimCLRTrainer(Trainer):
def __init__(self, hid_dim, out_dim, **kwargs):
self.hid_dim = hid_dim
self.out_dim = out_dim
super(SimCLRTrainer, self).__init__(**kwargs)
@dataclass
class BatchLogRecord(BaseBatchLogRecord):
lr: float | None
train_loss: float | None
train_accuracy: float | None
eval_loss: float | None
eval_accuracy: float | None
@staticmethod
def _prepare_dataset(dataset_config: SimCLRConfig.DatasetConfig) -> tuple[Dataset, Dataset]:
basic_augmentation = transforms.Compose([
transforms.RandomResizedCrop(
dataset_config.crop_size,
scale=dataset_config.crop_scale_range,
interpolation=transforms.InterpolationMode.BICUBIC
),
transforms.RandomHorizontalFlip(dataset_config.hflip_prob),
color_distortion(dataset_config.distort_strength),
])
if dataset_config.dataset in {'cifar10', 'cifar100', 'cifar'}:
transform = transforms.Compose([
basic_augmentation,
transforms.ToTensor(),
Clip(),
])
if dataset_config.dataset in {'cifar10', 'cifar'}:
train_set = CIFAR10(dataset_config.dataset_dir, train=True,
transform=TwinTransform(transform),
download=True)
test_set = CIFAR10(dataset_config.dataset_dir, train=False,
transform=TwinTransform(transform))
else: # CIFAR-100
train_set = CIFAR100(dataset_config.dataset_dir, train=True,
transform=TwinTransform(transform),
download=True)
test_set = CIFAR100(dataset_config.dataset_dir, train=False,
transform=TwinTransform(transform))
elif dataset_config.dataset in {'imagenet1k', 'imagenet'}:
random_gaussian_blur = RandomGaussianBlur(
kernel_size=dataset_config.crop_size // dataset_config.gauss_ker_scale,
sigma_range=dataset_config.gauss_sigma_range,
p=dataset_config.gauss_prob
),
transform = transforms.Compose([
basic_augmentation,
random_gaussian_blur,
transforms.ToTensor(),
Clip()
])
train_set = ImageNet(dataset_config.dataset_dir, 'train',
transform=TwinTransform(transform))
test_set = ImageNet(dataset_config.dataset_dir, 'val',
transform=TwinTransform(transform))
else:
raise NotImplementedError(f"Unimplemented dataset: '{dataset_config.dataset}")
return train_set, test_set
def _init_models(self, dataset: str) -> Iterable[tuple[str, torch.nn.Module]]:
if dataset in {'cifar10', 'cifar100', 'cifar'}:
model = CIFARSimCLRResNet50(self.hid_dim, self.out_dim)
elif dataset in {'imagenet1k', 'imagenet'}:
model = ImageNetSimCLRResNet50(self.hid_dim, self.out_dim)
else:
raise NotImplementedError(f"Unimplemented dataset: '{dataset}")
yield 'model', model
@staticmethod
def _configure_optimizers(
models: Iterable[tuple[str, torch.nn.Module]],
optim_config: SimCLRConfig.OptimConfig,
) -> Iterable[tuple[str, torch.optim.Optimizer]]:
def exclude_from_wd_and_adaptation(name):
if 'bn' in name:
return True
if optim_config.optim == 'lars' and 'bias' in name:
return True
for model_name, model in models:
param_groups = [
{
'params': [p for name, p in model.named_parameters()
if not exclude_from_wd_and_adaptation(name)],
'weight_decay': optim_config.weight_decay,
'layer_adaptation': True,
},
{
'params': [p for name, p in model.named_parameters()
if exclude_from_wd_and_adaptation(name)],
'weight_decay': 0.,
'layer_adaptation': False,
},
]
if optim_config.optim == 'adam':
optimizer = torch.optim.Adam(
param_groups,
lr=optim_config.lr,
betas=optim_config.betas,
)
elif optim_config.optim in {'sgd', 'lars'}:
optimizer = torch.optim.SGD(
param_groups,
lr=optim_config.lr,
momentum=optim_config.momentum,
)
else:
raise NotImplementedError(f"Unimplemented optimizer: '{optim_config.optim}'")
yield f"{model_name}_optim", optimizer
def _custom_init_fn(self, config: SimCLRConfig):
self.optims = {n: LARS(o) if config.optim_config.optim == 'lars' else o
for n, o in self.optims.items()}
def train(self, num_iters: int, loss_fn: Callable, logger: Loggers, device: int):
model = self.models['model']
optim = self.optims['model_optim']
sched = self.scheds['model_optim_sched']
train_loader = iter(self.train_loader)
model.train()
for iter_ in range(self.restore_iter, num_iters):
input_, _ = next(train_loader)
input_ = torch.cat(input_).to(device)
model.zero_grad()
output = model(input_)
train_loss, train_accuracy = loss_fn(output)
train_loss.backward()
optim.step()
if logger is not None:
self.log(logger, self.BatchLogRecord(
iter_, num_iters, iter_, iter_, num_iters,
optim.param_groups[0]['lr'],
train_loss.item(), train_accuracy.item(),
eval_loss=None, eval_accuracy=None,
))
dist.barrier()
if (iter_ + 1) % (num_iters // 100) == 0:
# TODO Gather results from other workers
metrics = torch.Tensor(list(self.eval(loss_fn, device)))
if logger is not None:
metrics_mean = metrics.mean(0)
eval_loss = metrics_mean[0].item()
eval_accuracy = metrics_mean[1].item()
eval_log = self.BatchLogRecord(
iter_, num_iters, iter_, iter_, num_iters,
lr=None, train_loss=None, train_accuracy=None,
eval_loss=eval_loss, eval_accuracy=eval_accuracy,
)
self.log(logger, eval_log)
self.save_checkpoint(eval_log)
model.train()
dist.barrier()
if sched is not None:
sched.step()
def eval(self, loss_fn: Callable, device: int):
model = self.models['model']
model.eval()
with torch.no_grad():
for input_, _ in self.test_loader:
input_ = torch.cat(input_).to(device)
output = model(input_)
loss, accuracy = loss_fn(output)
yield loss.item(), accuracy.item()
def main(local_rank, global_rank):
args = parse_args_and_config()
config = SimCLRConfig.from_args(args)
trainer = SimCLRTrainer(
seed=args.seed,
checkpoint_dir=args.checkpoint_dir,
device=local_rank,
inf_mode=True,
num_iters=args.num_iters,
config=config,
hid_dim=args.hid_dim,
out_dim=args.out_dim,
)
loggers = None
if global_rank == 0:
loggers = trainer.init_logger(args.log_dir)
trainer.train(args.num_iters, InfoNCELoss(args.temp), loggers, local_rank)
if __name__ == '__main__':
elastic_launch(main)
|