-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathtrain.py
187 lines (163 loc) · 8.1 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
import argparse
import os
import time
import numpy as np
import torch
import torch.nn as nn
import torch.backends.cudnn as cudnn
import torch.nn.functional as F
import torchvision.transforms as trn
import torchvision.datasets as dset
from models.wrn import WideResNet, Identity
def cosine_annealing(step, total_steps, lr_max, lr_min):
return lr_min + (lr_max - lr_min) * 0.5 * (
1 + np.cos(step / total_steps * np.pi))
def train(model, train_loader, scheduler, optimizer, state, device):
model.train()
loss_avg = 0.0
for i, (tr_data, tr_target) in enumerate(train_loader):
tr_data, tr_target = tr_data.to(device), tr_target.long().squeeze().to(device)
output = model(tr_data)
# backward
optimizer.zero_grad()
loss = F.cross_entropy(output, tr_target)
loss.backward()
optimizer.step()
scheduler.step()
loss_avg += float(loss)
loss_avg /= (i + 1)
state["train_loss"] = loss_avg
model.eval()
def test(model, test_loader, state, device):
model.eval()
loss_avg = 0.0
correct = 0
with torch.no_grad():
for data, target in test_loader:
data, target = data.to(device), target.to(device)
output = model(data)
loss = F.cross_entropy(output, target)
pred = output.data.max(1)[1]
correct += pred.eq(target.data).sum().item()
loss_avg += float(loss.data)
state['test_loss'] = loss_avg / len(test_loader)
state['test_accuracy'] = 1.0 * correct / len(test_loader.dataset)
def main(args):
torch.manual_seed(args.seed)
np.random.seed(args.seed)
state = {k: v for k, v in args._get_kwargs()}
print(state)
exp_dir = os.path.join(args.snapshots_dir, "run{:02d}".format(args.run))
# Make save directory
if not os.path.exists(exp_dir):
os.makedirs(exp_dir)
if not os.path.isdir(exp_dir):
raise Exception("%s is not a dir" % args.snapshots_dir)
with open(os.path.join(exp_dir, "experiment_info.txt"), "a") as f:
f.write("{}\n".format(args))
device = "cpu" if args.ngpu == 0 else "cuda"
test_transform = trn.Compose([trn.ToTensor()])
if args.dataset == 'cifar10':
size = 32
train_transform = trn.Compose([trn.RandomHorizontalFlip(), trn.RandomCrop(size, padding=4), trn.ToTensor()])
train_data = dset.CIFAR10(args.data_dir, train=True, transform=train_transform, download=True)
test_data = dset.CIFAR10(args.data_dir, train=False, transform=test_transform, download=True)
num_classes = 10
print("loaded CIFAR-10")
elif args.dataset == 'cifar100':
size = 32
train_transform = trn.Compose([trn.RandomHorizontalFlip(), trn.RandomCrop(size, padding=4), trn.ToTensor()])
train_data = dset.CIFAR100(args.data_dir, train=True, transform=train_transform, download=True)
test_data = dset.CIFAR100(args.data_dir, train=False, transform=test_transform, download=True)
num_classes = 100
print("loaded CIFAR-100")
else:
assert args.dataset == "tinyimagenet"
size = 64
train_transform = trn.Compose([trn.RandomHorizontalFlip(), trn.RandomCrop(size, padding=4), trn.ToTensor()])
train_data = dset.ImageFolder(os.path.join(args.data_dir, "tiny-imagenet-200", "train"), transform=train_transform)
test_data = dset.ImageFolder(os.path.join(args.data_dir, "tiny-imagenet-200", "val"), transform=test_transform)
num_classes = 200
print("loaded TIN")
train_loader = torch.utils.data.DataLoader(
train_data, batch_size=args.batch_size, shuffle=True,
num_workers=args.prefetch, pin_memory=True)
test_loader = torch.utils.data.DataLoader(
test_data, batch_size=args.test_bs, shuffle=False,
num_workers=args.prefetch, pin_memory=True)
is_tinyimagenet = args.dataset == "tinyimagenet"
norm = {"batch_norm": nn.BatchNorm2d, "group_norm": nn.GroupNorm, "instance_norm": nn.InstanceNorm2d,
"layer_norm": nn.LayerNorm, "identity": Identity}[args.norm]
model = WideResNet(args.layers, num_classes, widen_factor=args.widen_factor, dropRate=args.droprate,
tinyimagenet=is_tinyimagenet, norm=norm, ngroups=args.ngroups, affine=not args.no_affine)
if args.ngpu > 1:
model = nn.DataParallel(model, device_ids=list(range(args.ngpu)))
model.to(device)
if args.ngpu > 0:
torch.cuda.manual_seed(args.seed)
cudnn.benchmark = True # fire on all cylinders
optimizer = torch.optim.SGD(model.parameters(), state["learning_rate"], momentum=state["momentum"],
weight_decay=state["decay"], nesterov=True)
print("Beginning Training\n")
scheduler = torch.optim.lr_scheduler.LambdaLR(
optimizer,
lr_lambda=lambda step: cosine_annealing(
step,
args.epochs * len(train_loader),
1, # since lr_lambda computes multiplicative factor
1e-6 / args.learning_rate))
for epoch in range(args.epochs):
state["epoch"] = epoch
begin_epoch = time.time()
# train
train(model, train_loader, scheduler, optimizer, state, device)
# test
test(model, test_loader, state, device)
# Save model
save_file = os.path.join(exp_dir, "checkpoint_epoch_" + str(epoch) + ".pt")
torch.save(
{"epoch": epoch, "model_state_dict": model.state_dict(), "optimizer_state_dict": optimizer.state_dict()},
save_file)
# Let us not waste space and delete the previous model unless epoch is a multiple of 25
if not epoch % 25 == 0:
prev_path = os.path.join(exp_dir, "checkpoint_epoch_" + str(epoch - 1) + ".pt")
if os.path.exists(prev_path): os.remove(prev_path)
# Show results
with open(os.path.join(exp_dir, "training_results.csv"), "a") as f:
f.write("%03d,%05d,%0.6f,%0.5f,%0.2f\n" % (
(epoch + 1),
time.time() - begin_epoch,
state["train_loss"],
state["test_loss"],
100 - 100. * state["test_accuracy"],
))
print("Epoch {0:3d} | Time {1:5d} | Train Loss {2:.4f} | Test Loss {3:.3f} | Test Error {4:.2f}".format(
(epoch + 1),
int(time.time() - begin_epoch),
state["train_loss"],
state["test_loss"],
100 - 100. * state["test_accuracy"])
)
if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument("--learning_rate", "-lr", type=float, default=0.1, help="The initial learning rate.")
parser.add_argument("--batch_size", "-b", type=int, default=128, help="Batch size.")
parser.add_argument("--test_bs", type=int, default=256)
parser.add_argument("--momentum", type=float, default=0.9, help="Momentum.")
parser.add_argument("--decay", "-d", type=float, default=0.0005, help="Weight decay (L2 penalty).")
parser.add_argument('--epochs', '-e', type=int, default=100, help='Number of epochs to train.')
parser.add_argument("--layers", default=40, type=int, help="total number of layers")
parser.add_argument("--widen-factor", default=2, type=int, help="widen factor")
parser.add_argument("--droprate", default=0.3, type=float, help="dropout probability")
parser.add_argument("--seed", type=int, default=1, help="seed")
parser.add_argument("--prefetch", type=int, default=6, help="Pre-fetching threads.")
parser.add_argument("--dataset", type=str, default="cifar10")
parser.add_argument("--norm", type=str, default="batch_norm")
parser.add_argument("--ngroups", type=int, default=4, help="Number of groups if GroupNorm is used.")
parser.add_argument("--no_affine", action="store_true", help="Whether to not use affine parameters for normalization.")
parser.add_argument("--ngpu", type=int, default=1, help="0 = CPU.")
parser.add_argument("--snapshots_dir", type=str, default="./snapshots", required=True)
parser.add_argument("--data_dir", type=str, default="./data", required=True)
parser.add_argument("--run", type=int, default=1)
args = parser.parse_args()
main(args)