-
Notifications
You must be signed in to change notification settings - Fork 0
/
dist_train_g.py
213 lines (164 loc) · 9.35 KB
/
dist_train_g.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
import argparse
import datetime
import math
import os
import stat
from itertools import cycle
import torch
import torch.distributed as dist
import torch.optim as optim
import torch.optim.lr_scheduler as lr_scheduler
import torch.utils.data as Data
from torch import nn
from tqdm import tqdm
from configs.transforms import transforms_dict
from datasets.folder import ImageFolderG, ImageFolderUnsupG
from utils.gan import get_model, load_weights, update
def train(args):
device = torch.device(args.device if torch.cuda.is_available() else "cpu")
num_workers = min([os.cpu_count(), args.batch_size if args.batch_size > 1 else 0, 8])
torch.cuda.set_device(args.local_rank)
dist.init_process_group(backend="nccl", init_method="env://")
to_save = False if args.do_not_save or args.local_rank != 0 else True
saving_path = args.saving_root
if to_save and not os.path.exists(saving_path):
os.makedirs(saving_path)
os.chmod(saving_path, stat.S_IRWXU | stat.S_IRWXG | stat.S_IRWXO)
sp = "" if args.g_weights == "" else " "
train_path = os.path.join(args.data_root, args.dataset, args.sub, "train_few_shot_tgt" if args.fs else "train_tgt")
val_path = os.path.join(args.data_root, args.dataset, "val_tgt" if args.fs else "val_tgt")
if args.local_rank == 0:
print(f"device: {device}\n"
f"Using {num_workers} dataloader workers every process\n\n"
f"{sp} train path: {train_path}\n"
f"{sp} val path: {val_path}\n" +
(f"g-weights path: {args.g_weights}\n" if args.g_weights != "" else "") +
(f"{sp}saving path: {saving_path}\n\n" if to_save else "\n") +
f"dataset: {args.dataset} (src{len(os.listdir(train_path.replace('train_few_shot_tgt', 'train_src')))}-tgt{len(os.listdir(train_path))})" +
(f" | using unsup.\n" if args.unsup or not args.fs else "\n") +
(f"remarks: {args.remarks}\n\n" if args.remarks is not None else "\n"))
if args.fs:
train_dataset = ImageFolderG(root=train_path, transform=transforms_dict[args.dataset]["train"])
else:
train_dataset = ImageFolderUnsupG(root=train_path, transform=transforms_dict[args.dataset]["train"])
val_dataset = ImageFolderG(root=val_path, transform=transforms_dict[args.dataset]["val"])
train_sampler = Data.distributed.DistributedSampler(dataset=train_dataset, shuffle=False)
val_sampler = Data.distributed.DistributedSampler(dataset=val_dataset, shuffle=False)
train_loader = Data.DataLoader(dataset=train_dataset,
sampler=train_sampler,
batch_size=args.batch_size,
pin_memory=True,
num_workers=num_workers,
collate_fn=ImageFolderG.collate_fn)
val_loader = Data.DataLoader(dataset=val_dataset,
sampler=val_sampler,
batch_size=args.batch_size,
pin_memory=True,
num_workers=num_workers,
collate_fn=ImageFolderG.collate_fn)
if args.unsup:
train_path_unsup = os.path.join(args.data_root, args.dataset, args.sub, "train_tgt")
train_dataset_unsup = ImageFolderUnsupG(root=train_path_unsup, transform=transforms_dict[args.dataset]["train"])
train_sampler_unsup = Data.distributed.DistributedSampler(dataset=train_dataset_unsup, shuffle=False)
train_loader_unsup = Data.DataLoader(dataset=train_dataset_unsup,
sampler=train_sampler_unsup,
batch_size=args.batch_size,
pin_memory=True,
num_workers=num_workers,
collate_fn=ImageFolderG.collate_fn)
train_loader_unsup = cycle(train_loader_unsup)
generator = get_model(dataset=args.dataset, mode='G', device=device)
load_weights(model=generator, weights=args.g_weights, device=device)
if args.local_rank == 0:
print("\n")
generator = nn.parallel.DistributedDataParallel(generator, device_ids=[args.local_rank])
criterion = nn.BCELoss()
optim_g = optim.Adam(generator.parameters(), lr=args.lr)
lr_lambda = lambda x: ((1 + math.cos(x * math.pi / args.epochs)) / 2) * (1 - args.lrf) + args.lrf # cosine
scheduler_g = lr_scheduler.LambdaLR(optim_g, lr_lambda=lr_lambda)
for epoch in range(1, args.epochs + 1):
generator.train()
total_g, correct_g = 0, 0
train_loader_ = tqdm(train_loader) if args.local_rank == 0 else train_loader
for i, (img_pairs, similarities) in enumerate(train_loader_):
img_pairs = img_pairs.to(device)
similarities = similarities.to(device)
g_out = generator(img_pairs) # [B, 1]
g_loss = criterion(g_out, similarities)
total_g += similarities.shape[0]
correct_g += ((g_out.detach() + 0.5).int() == similarities.int()).sum().item()
acc_g_train = correct_g / total_g
update(generator, g_loss, optim_g)
if args.local_rank == 0:
train_loader_.desc = f"epoch [{epoch}/{args.epochs}] " \
f"loss G: {g_loss:.3f} | " \
f"acc G: {acc_g_train:.3f} | " \
f"lr: {optim_g.param_groups[0]['lr']:.2e}"
if args.unsup:
img_pairs, similarities = next(train_loader_unsup)
img_pairs = img_pairs.to(device)
similarities = similarities.to(device)
g_out = generator(img_pairs) # [B, 1]
g_loss = criterion(g_out, similarities)
update(generator, g_loss, optim_g)
scheduler_g.step()
if epoch % args.eval_interval == 0:
with torch.no_grad():
generator.eval()
total, correct = 0, 0
val_loader_ = tqdm(val_loader) if args.local_rank == 0 else val_loader
for i, (img_pairs, similarities) in enumerate(val_loader_):
img_pairs = img_pairs.to(device)
similarities = similarities.to(device)
g_out = generator(img_pairs) # [B, 1]
g_loss = criterion(g_out, similarities)
total_per_batch = torch.IntTensor([similarities.shape[0]]).to(device)
dist.all_reduce(total_per_batch, op=dist.ReduceOp.SUM)
total += total_per_batch.item()
correct_per_batch = ((g_out + 0.5).int() == similarities.int()).sum()
dist.all_reduce(correct_per_batch, op=dist.ReduceOp.SUM)
correct += correct_per_batch.item()
acc_g_tgt = correct / total
if args.local_rank == 0:
val_loader_.desc = f"valid loss G: {g_loss:.3f} | acc: {acc_g_tgt:.3f}"
if to_save and epoch % args.saving_interval == 0:
if args.fs:
pth = f"pre-G-{args.dataset}-{args.sub.split('-')[0]}-shot-{epoch}.pth"
else:
pth = f"pre-G-{args.dataset}-{epoch}.pth"
pth_path = os.path.join(saving_path, pth)
torch.save(generator.state_dict(), pth_path)
os.chmod(pth_path, stat.S_IRWXU | stat.S_IRWXG | stat.S_IRWXO)
print("model saved")
if args.local_rank == 0:
print(f"remarks: {args.remarks}\n\n" if args.remarks is not None else "\n")
dist.destroy_process_group()
if __name__ == '__main__':
date_time = datetime.datetime.now()
parser = argparse.ArgumentParser()
parser.add_argument("--local_rank", type=int, default=-1)
parser.add_argument('--device', default='cuda', help='device id (i.e. 0 or 0,1 or cpu)')
parser.add_argument('--batch-size', type=int, default=32)
parser.add_argument('--epochs', type=int, default=100)
parser.add_argument('--lr', type=float, default=2e-5)
parser.add_argument('--lrf', type=float, default=0.01)
parser.add_argument('--eval-interval', type=int, default=1)
parser.add_argument('--unsup', action='store_true', help='whether to use the unsupervised mechanism')
# datasets
parser.add_argument('--data-root', type=str, default='./datasets/datasets')
parser.add_argument('--dataset', type=str, default='mini-imagenet')
parser.add_argument('--sub', type=str, default='src50-tgt50')
parser.add_argument('--fs', action='store_true', help='whether to train few-shot target')
# weights
parser.add_argument('--g-weights', type=str, help='initial g_weights path')
# save weights
parser.add_argument('--do-not-save', action="store_true")
parser.add_argument('--saving-root', type=str, default='./pre-trained')
parser.add_argument('--date-time', type=str,
default=f'{date_time.month:02d}-{date_time.day:02d}_'
f'{date_time.hour:02d}-{date_time.minute:02d}-{date_time.second:02d}')
parser.add_argument('--saving-interval', type=int, default=1)
# remarks
parser.add_argument('--remarks', type=str)
args = parser.parse_args()
train(args)