-
Notifications
You must be signed in to change notification settings - Fork 20
/
train.py
182 lines (152 loc) · 6.92 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
"""Train Glow on CIFAR-10.
Train script adapted from: https://github.com/kuangliu/pytorch-cifar/
"""
import argparse
import numpy as np
import os
import random
import torch
import torch.optim as optim
import torch.optim.lr_scheduler as sched
import torch.backends.cudnn as cudnn
import torch.utils.data as data
import torchvision
import torchvision.transforms as transforms
import util
from models import Glow
from tqdm import tqdm
def main(args):
# Set up main device and scale batch size
device = 'cuda' if torch.cuda.is_available() and args.gpu_ids else 'cpu'
args.batch_size *= max(1, len(args.gpu_ids))
# Set random seeds
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
torch.cuda.manual_seed_all(args.seed)
# No normalization applied, since Glow expects inputs in (0, 1)
transform_train = transforms.Compose([
transforms.RandomHorizontalFlip(),
transforms.ToTensor()
])
transform_test = transforms.Compose([
transforms.ToTensor()
])
trainset = torchvision.datasets.CIFAR10(root='data', train=True, download=True, transform=transform_train)
trainloader = data.DataLoader(trainset, batch_size=args.batch_size, shuffle=True, num_workers=args.num_workers)
testset = torchvision.datasets.CIFAR10(root='data', train=False, download=True, transform=transform_test)
testloader = data.DataLoader(testset, batch_size=args.batch_size, shuffle=False, num_workers=args.num_workers)
# Model
print('Building model..')
net = Glow(num_channels=args.num_channels,
num_levels=args.num_levels,
num_steps=args.num_steps)
net = net.to(device)
if device == 'cuda':
net = torch.nn.DataParallel(net, args.gpu_ids)
cudnn.benchmark = args.benchmark
start_epoch = 0
if args.resume:
# Load checkpoint.
print('Resuming from checkpoint at ckpts/best.pth.tar...')
assert os.path.isdir('ckpts'), 'Error: no checkpoint directory found!'
checkpoint = torch.load('ckpts/best.pth.tar')
net.load_state_dict(checkpoint['net'])
global best_loss
global global_step
best_loss = checkpoint['test_loss']
start_epoch = checkpoint['epoch']
global_step = start_epoch * len(trainset)
loss_fn = util.NLLLoss().to(device)
optimizer = optim.Adam(net.parameters(), lr=args.lr)
scheduler = sched.LambdaLR(optimizer, lambda s: min(1., s / args.warm_up))
for epoch in range(start_epoch, start_epoch + args.num_epochs):
train(epoch, net, trainloader, device, optimizer, scheduler,
loss_fn, args.max_grad_norm)
test(epoch, net, testloader, device, loss_fn, args.num_samples)
@torch.enable_grad()
def train(epoch, net, trainloader, device, optimizer, scheduler, loss_fn, max_grad_norm):
global global_step
print('\nEpoch: %d' % epoch)
net.train()
loss_meter = util.AverageMeter()
with tqdm(total=len(trainloader.dataset)) as progress_bar:
for x, _ in trainloader:
x = x.to(device)
optimizer.zero_grad()
z, sldj = net(x, reverse=False)
loss = loss_fn(z, sldj)
loss_meter.update(loss.item(), x.size(0))
loss.backward()
if max_grad_norm > 0:
util.clip_grad_norm(optimizer, max_grad_norm)
optimizer.step()
scheduler.step(global_step)
progress_bar.set_postfix(nll=loss_meter.avg,
bpd=util.bits_per_dim(x, loss_meter.avg),
lr=optimizer.param_groups[0]['lr'])
progress_bar.update(x.size(0))
global_step += x.size(0)
@torch.no_grad()
def sample(net, batch_size, device):
"""Sample from RealNVP model.
Args:
net (torch.nn.DataParallel): The RealNVP model wrapped in DataParallel.
batch_size (int): Number of samples to generate.
device (torch.device): Device to use.
"""
z = torch.randn((batch_size, 3, 32, 32), dtype=torch.float32, device=device)
x, _ = net(z, reverse=True)
x = torch.sigmoid(x)
return x
@torch.no_grad()
def test(epoch, net, testloader, device, loss_fn, num_samples):
global best_loss
net.eval()
loss_meter = util.AverageMeter()
with tqdm(total=len(testloader.dataset)) as progress_bar:
for x, _ in testloader:
x = x.to(device)
z, sldj = net(x, reverse=False)
loss = loss_fn(z, sldj)
loss_meter.update(loss.item(), x.size(0))
progress_bar.set_postfix(nll=loss_meter.avg,
bpd=util.bits_per_dim(x, loss_meter.avg))
progress_bar.update(x.size(0))
# Save checkpoint
if loss_meter.avg < best_loss:
print('Saving...')
state = {
'net': net.state_dict(),
'test_loss': loss_meter.avg,
'epoch': epoch,
}
os.makedirs('ckpts', exist_ok=True)
torch.save(state, 'ckpts/best.pth.tar')
best_loss = loss_meter.avg
# Save samples and data
images = sample(net, num_samples, device)
os.makedirs('samples', exist_ok=True)
images_concat = torchvision.utils.make_grid(images, nrow=int(num_samples ** 0.5), padding=2, pad_value=255)
torchvision.utils.save_image(images_concat, 'samples/epoch_{}.png'.format(epoch))
if __name__ == '__main__':
parser = argparse.ArgumentParser(description='Glow on CIFAR-10')
def str2bool(s):
return s.lower().startswith('t')
parser.add_argument('--batch_size', default=64, type=int, help='Batch size per GPU')
parser.add_argument('--benchmark', type=str2bool, default=True, help='Turn on CUDNN benchmarking')
parser.add_argument('--gpu_ids', default=[0], type=eval, help='IDs of GPUs to use')
parser.add_argument('--lr', default=1e-3, type=float, help='Learning rate')
parser.add_argument('--max_grad_norm', type=float, default=-1., help='Max gradient norm for clipping')
parser.add_argument('--num_channels', '-C', default=512, type=int, help='Number of channels in hidden layers')
parser.add_argument('--num_levels', '-L', default=3, type=int, help='Number of levels in the Glow model')
parser.add_argument('--num_steps', '-K', default=32, type=int, help='Number of steps of flow in each level')
parser.add_argument('--num_epochs', default=100, type=int, help='Number of epochs to train')
parser.add_argument('--num_samples', default=64, type=int, help='Number of samples at test time')
parser.add_argument('--num_workers', default=8, type=int, help='Number of data loader threads')
parser.add_argument('--resume', type=str2bool, default=False, help='Resume from checkpoint')
parser.add_argument('--seed', type=int, default=0, help='Random seed for reproducibility')
parser.add_argument('--warm_up', default=500000, type=int, help='Number of steps for lr warm-up')
best_loss = 0
global_step = 0
main(parser.parse_args())