-
Notifications
You must be signed in to change notification settings - Fork 0
/
train.py
229 lines (176 loc) · 9.78 KB
/
train.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
import os
import sys
import math
import argparse
import json
from pathlib import Path
import webdataset as wds
import torch
print(torch.__version__)
from numpy import mean as npmean
import torch.backends.cudnn as cudnn
import torchvision.transforms as transforms
from torchvision.datasets import ImageFolder
from torch.utils.data import DataLoader, SequentialSampler
from torch.nn.parallel import DistributedDataParallel as DDP
from torchvision.utils import save_image
import tae
import util.misc as misc
from util.misc import NativeScalerWithGradNormCount as NativeScaler
def get_args_parser():
parser = argparse.ArgumentParser('TAE training with webdataset', add_help=False)
parser.add_argument('--batch_size_per_gpu', default=256, type=int, help='Batch size per GPU (effective batch size is batch_size_per_gpu * accum_iter * # gpus')
parser.add_argument('--accum_iter', default=1, type=int, help='Accumulate gradient iterations (for increasing the effective batch size under memory constraints)')
parser.add_argument('--save_prefix', default="", type=str, help="""prefix for saving checkpoint and log files""")
parser.add_argument('--save_freq', default=10000, type=int, help='Save checkpoint every this many iterations.')
# Model parameters
parser.add_argument('--model', default='', type=str, help='Name of model to train')
parser.add_argument('--ckpt', default='', help='resume from a checkpoint')
parser.add_argument('--input_size', default=224, type=int, help='images input size')
parser.add_argument('--compile', action='store_true', help='whether to compile the model for improved efficiency (default: false)')
parser.add_argument('--display', action='store_true', help='whether to display reconstruction at regular intervals.')
# Optimizer parameters
parser.add_argument('--weight_decay', type=float, default=0.05, help='weight decay (default: 0.05)')
parser.add_argument('--max_lr', type=float, default=0.0001, help='max learning rate')
parser.add_argument('--min_lr', type=float, default=0.00001, help='min learning rate')
parser.add_argument('--switch_it', type=float, default=900000, help='iteration at which to switch to lower lr')
parser.add_argument('--num_its', type=float, default=1000001, help='total number of iterations')
# Dataset parameters
parser.add_argument('--train_data_path', default='', type=str)
parser.add_argument('--val_data_path', default='', type=str)
parser.add_argument('--output_dir', default='./output_dir', help='path where to save, empty for no saving')
parser.add_argument('--device', default='cuda', help='device to use for training/testing')
parser.add_argument('--num_workers', default=16, type=int)
parser.add_argument('--jitter_scale', default=[0.2, 1.0], type=float, nargs="+")
parser.add_argument('--jitter_ratio', default=[3.0/4.0, 4.0/3.0], type=float, nargs="+")
# distributed training parameters
parser.add_argument('--dist_url', default='env://', help='url used to set up distributed training')
return parser
def main(args):
misc.init_distributed_mode(args)
print('job dir: {}'.format(os.path.dirname(os.path.realpath(__file__))))
print("{}".format(args).replace(', ', ',\n'))
device = torch.device(args.device)
cudnn.benchmark = True
# validation transforms
val_transform = transforms.Compose([
transforms.Resize(args.input_size + 32, interpolation=3),
transforms.CenterCrop(args.input_size),
transforms.ToTensor(),
transforms.Normalize((0.485, 0.456, 0.406), (0.229, 0.224, 0.225)),
])
# training transforms
train_transform = transforms.Compose([
transforms.RandomResizedCrop(args.input_size, scale=args.jitter_scale, ratio=args.jitter_ratio, interpolation=3),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
transforms.Normalize((0.485, 0.456, 0.406), (0.229, 0.224, 0.225)),
])
# train and val datasets and loaders
train_dataset = wds.WebDataset(args.train_data_path, resampled=True).shuffle(10000, initial=10000).decode("pil").to_tuple("jpg", "cls").map_tuple(train_transform, lambda x: x)
train_loader = wds.WebLoader(train_dataset, batch_size=args.batch_size_per_gpu, num_workers=args.num_workers)
val_dataset = ImageFolder(args.val_data_path, transform=val_transform)
val_sampler = SequentialSampler(val_dataset)
val_loader = DataLoader(val_dataset, sampler=val_sampler, batch_size=8*args.batch_size_per_gpu, num_workers=args.num_workers, pin_memory=True, drop_last=False) # note we use a larger batch size for eval
print(f"Train and val data loaded.")
# define the model
model = tae.__dict__[args.model]()
model.to(device)
model_without_ddp = model
# optionally compile model
if args.compile:
model = torch.compile(model)
model = DDP(model, device_ids=[args.gpu]) # TODO: try FSDP
print(f"Model: {model_without_ddp}")
print(f"Number of params (M): {(sum(p.numel() for p in model_without_ddp.parameters() if p.requires_grad) / 1.e6)}")
# set wd as 0 for bias and norm layers
param_groups = misc.add_weight_decay(model_without_ddp, args.weight_decay, bias_wd=False)
optimizer = torch.optim.AdamW(param_groups, lr=args.max_lr, betas=(0.9, 0.95), fused=True) # setting fused True for faster updates (hopefully)
loss_scaler = NativeScaler()
misc.load_model(args.ckpt, model_without_ddp, optimizer=optimizer, loss_scaler=loss_scaler)
model.train()
metric_logger = misc.MetricLogger(delimiter=" ")
optimizer.zero_grad()
best_eval_loss = 100.0
print("Starting TAE training!")
# infinite stream for iterable webdataset
for it, (samples, _) in enumerate(train_loader):
if it == args.num_its:
break
if it % args.accum_iter == 0:
misc.adjust_learning_rate(optimizer, args.max_lr, args.min_lr, it, args.switch_it)
# optionally pick 8 examples for display and softmax estimation at regular intervals
if args.display and it % args.save_freq == 0:
samples_for_display_and_softmax = samples[:8, ...]
samples = samples.to(device, non_blocking=True)
with torch.cuda.amp.autocast():
loss, _ = model(samples)
loss_value = loss.item()
if not math.isfinite(loss_value):
print("Loss is {}, stopping training".format(loss_value))
sys.exit(1)
loss = loss / args.accum_iter
loss_scaler(loss, optimizer, parameters=model.parameters(), update_grad=(it + 1) % args.accum_iter == 0)
if (it + 1) % args.accum_iter == 0:
optimizer.zero_grad()
torch.cuda.synchronize()
metric_logger.update(loss=loss_value)
metric_logger.update(lr=optimizer.param_groups[0]["lr"])
if it != 0 and it % args.save_freq == 0:
# estimate eval loss
print(f"Iteration {it}, evaluating ...")
eval_loss = evaluate(val_loader, model_without_ddp, device)
# save checkpoint only if eval_loss decreases
if eval_loss < best_eval_loss:
print("Best eval loss improved! Saving checkpoint.")
save_dict = {
'model': model_without_ddp.state_dict(),
'optimizer': optimizer.state_dict(),
'args': args,
'iteration': it,
'scaler': loss_scaler.state_dict(),
}
misc.save_on_master(save_dict, os.path.join(args.output_dir, f"{args.save_prefix}_checkpoint.pth"))
best_eval_loss = eval_loss
# gather the stats from all processes
metric_logger.synchronize_between_processes()
train_stats = {k: meter.global_avg for k, meter in metric_logger.meters.items()}
log_stats = {**{f'train_{k}': v for k, v in train_stats.items()}, 'eval_loss': eval_loss, 'iteration': it}
# write log
if misc.is_main_process():
with (Path(args.output_dir) / (args.save_prefix + "_log.txt")).open("a") as f:
f.write(json.dumps(log_stats) + "\n")
# optionally display some reconstructions
if args.display:
with torch.no_grad():
samples_for_display_and_softmax = samples_for_display_and_softmax.to(device, non_blocking=True)
with torch.cuda.amp.autocast():
_, pred = model_without_ddp(samples_for_display_and_softmax)
pred = model_without_ddp.unpatchify(pred)
combined = torch.cat((samples_for_display_and_softmax, pred), 0)
# save original images and their reconstructions
save_image(combined, os.path.join(args.output_dir, f"{args.save_prefix}_reconstructions_iter_{it}.jpg"), nrow=8, padding=1, normalize=True, scale_each=True)
# start a fresh logger to wipe off old stats
metric_logger = misc.MetricLogger(delimiter=" ")
# switch back to train mode, not 100% sure if this is strictly necessary since we're passing the unwrapped model to eval now
model.train()
@torch.no_grad()
def evaluate(data_loader, model, device):
# switch to eval mode
model.eval()
eval_loss = []
for _, (samples, _) in enumerate(data_loader):
samples = samples.to(device, non_blocking=True)
# compute loss
with torch.cuda.amp.autocast():
loss, _ = model(samples)
loss_value = loss.item()
eval_loss.append(loss_value)
eval_loss = npmean(eval_loss)
print(f"Current eval loss: {eval_loss}")
return eval_loss
if __name__ == '__main__':
args = get_args_parser()
args = args.parse_args()
Path(args.output_dir).mkdir(parents=True, exist_ok=True)
main(args)