-
Notifications
You must be signed in to change notification settings - Fork 17
/
train.py
171 lines (135 loc) · 6.54 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
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
from torch.optim import lr_scheduler
from torchvision import datasets, transforms, utils
import numpy as np
import pdb
import argparse
import time
from invertible_layers import *
from utils import *
# ------------------------------------------------------------------------------
parser = argparse.ArgumentParser()
# training
parser.add_argument('--batch_size', type=int, default=64)
parser.add_argument('--depth', type=int, default=32)
parser.add_argument('--n_levels', type=int, default=3)
parser.add_argument('--norm', type=str, default='actnorm')
parser.add_argument('--permutation', type=str, default='conv')
parser.add_argument('--coupling', type=str, default='affine')
parser.add_argument('--n_bits_x', type=int, default=8)
parser.add_argument('--n_epochs', type=int, default=2000)
parser.add_argument('--learntop', action='store_true')
parser.add_argument('--n_warmup', type=int, default=20, help='number of warmup epochs')
parser.add_argument('--lr', type=float, default=1e-3)
# logging
parser.add_argument('--print_every', type=int, default=500, help='print NLL every _ minibatches')
parser.add_argument('--test_every', type=int, default=5, help='test on valid every _ epochs')
parser.add_argument('--save_every', type=int, default=5, help='save model every _ epochs')
parser.add_argument('--data_dir', type=str, default='../pixelcnn-pp')
parser.add_argument('--save_dir', type=str, default='exps', help='directory for log / saving')
parser.add_argument('--load_dir', type=str, default=None, help='directory from which to load existing model')
args = parser.parse_args()
args.n_bins = 2 ** args.n_bits_x
# reproducibility is good
np.random.seed(0)
torch.manual_seed(0)
torch.cuda.manual_seed_all(0)
# loading / dataset preprocessing
tf = transforms.Compose([transforms.ToTensor(),
lambda x: x + torch.zeros_like(x).uniform_(0., 1./args.n_bins)])
train_loader = torch.utils.data.DataLoader(datasets.CIFAR10(args.data_dir, train=True,
download=True, transform=tf), batch_size=args.batch_size, shuffle=True, num_workers=10, drop_last=True)
test_loader = torch.utils.data.DataLoader(datasets.CIFAR10(args.data_dir, train=False,
transform=tf), batch_size=args.batch_size, shuffle=False, num_workers=10, drop_last=True)
# construct model and ship to GPU
model = Glow_((args.batch_size, 3, 32, 32), args).cuda()
print(model)
print("number of model parameters:", sum([np.prod(p.size()) for p in model.parameters()]))
# set up the optimizer
optim = optim.Adam(model.parameters(), lr=1e-3)
scheduler = torch.optim.lr_scheduler.StepLR(optim, step_size=45, gamma=0.1)
# data dependant init
init_loader = torch.utils.data.DataLoader(datasets.CIFAR10(args.data_dir, train=True,
download=True, transform=tf), batch_size=512, shuffle=True, num_workers=1)
with torch.no_grad():
model.eval()
for (img, _) in init_loader:
img = img.cuda()
objective = torch.zeros_like(img[:, 0, 0, 0])
_ = model(img, objective)
break
# once init is done, we leverage Data Parallel
model = nn.DataParallel(model).cuda()
start_epoch = 0
# load trained model if necessary (must be done after DataParallel)
if args.load_dir is not None:
model, optim, start_epoch = load_session(model, optim, args)
# training loop
# ------------------------------------------------------------------------------
for epoch in range(start_epoch, args.n_epochs):
print('epoch %s' % epoch)
model.train()
avg_train_bits_x = 0.
num_batches = len(train_loader)
for i, (img, label) in enumerate(train_loader):
# if i > 10 : break
t = time.time()
img = img.cuda()
objective = torch.zeros_like(img[:, 0, 0, 0])
# discretizing cost
objective += float(-np.log(args.n_bins) * np.prod(img.shape[1:]))
# log_det_jacobian cost (and some prior from Split OP)
z, objective = model(img, objective)
nll = (-objective) / float(np.log(2.) * np.prod(img.shape[1:]))
# Generative loss
nobj = torch.mean(nll)
optim.zero_grad()
nobj.backward()
torch.nn.utils.clip_grad_value_(model.parameters(), 5)
torch.nn.utils.clip_grad_norm_(model.parameters(), 100)
optim.step()
avg_train_bits_x += nobj.item()
# update learning rate
new_lr = float(args.lr * min(1., (i + epoch * num_batches) / (args.n_warmup * num_batches)))
for pg in optim.param_groups: pg['lr'] = new_lr
if (i + 1) % args.print_every == 0:
print('avg train bits per pixel {:.4f}'.format(avg_train_bits_x / args.print_every))
avg_train_bits_x = 0.
sample = model.module.sample()
grid = utils.make_grid(sample)
utils.save_image(grid, '../glow/samples/cifar_Test_{}_{}.png'.format(epoch, i // args.print_every))
print('iteration took {:.4f}'.format(time.time() - t))
# test loop
# --------------------------------------------------------------------------
if (epoch + 1) % args.test_every == 0:
model.eval()
avg_test_bits_x = 0.
with torch.no_grad():
for i, (img, label) in enumerate(test_loader):
# if i > 10 : break
img = img.cuda()
objective = torch.zeros_like(img[:, 0, 0, 0])
# discretizing cost
objective += float(-np.log(args.n_bins) * np.prod(img.shape[1:]))
# log_det_jacobian cost (and some prior from Split OP)
z, objective = model(img, objective)
last_img = img
nll = (-objective) / float(np.log(2.) * np.prod(img.shape[1:]))
# Generative loss
nobj = torch.mean(nll)
avg_test_bits_x += nobj
print('avg test bits per pixel {:.4f}'.format(avg_test_bits_x.item() / i))
sample = model.module.sample()
grid = utils.make_grid(sample)
utils.save_image(grid, '../glow/samples/cifar_Test_{}.png'.format(epoch))
# reconstruct
x_hat = model.module.reverse_(z, objective)[0]
grid = utils.make_grid(x_hat)
utils.save_image(grid, '../glow/samples/cifar_Test_Recon{}.png'.format(epoch))
grid = utils.make_grid(last_img)
utils.save_image(grid, '../glow/samples/cifar_Test_Target.png')
if (epoch + 1) % args.save_every == 0:
save_session(model, optim, args, epoch)