forked from pclucas14/pixel-cnn-pp
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathtrain_gaussian.py
173 lines (152 loc) · 7.37 KB
/
train_gaussian.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
import time
import os
import argparse
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
from tensorboardX import SummaryWriter
from utils import *
from model import *
from PIL import Image
parser = argparse.ArgumentParser()
# data I/O
parser.add_argument('-i', '--data_dir', type=str,
default='data', help='Location for the dataset')
parser.add_argument('-o', '--save_dir', type=str, default='models',
help='Location for parameter checkpoints and samples')
parser.add_argument('-d', '--dataset', type=str,
default='cifar', help='Can be either cifar|mnist')
parser.add_argument('-p', '--print_every', type=int, default=50,
help='how many iterations between print statements')
parser.add_argument('-t', '--save_interval', type=int, default=10,
help='Every how many epochs to write checkpoint/samples?')
parser.add_argument('-r', '--load_params', type=str, default=None,
help='Restore training from previous model checkpoint?')
# model
parser.add_argument('-q', '--nr_resnet', type=int, default=5,
help='Number of residual blocks per stage of the model')
parser.add_argument('-n', '--nr_filters', type=int, default=160,
help='Number of filters to use across the model. Higher = larger model.')
parser.add_argument('-m', '--nr_logistic_mix', type=int, default=10,
help='Number of logistic components in the mixture. Higher = more flexible model')
parser.add_argument('-l', '--lr', type=float,
default=0.0001, help='Base learning rate')
parser.add_argument('-e', '--lr_decay', type=float, default=0.999995,
help='Learning rate decay, applied every step of the optimization')
parser.add_argument('-b', '--batch_size', type=int, default=64,
help='Batch size during training per GPU')
parser.add_argument('-x', '--max_epochs', type=int,
default=5000, help='How many epochs to run in total?')
parser.add_argument('-s', '--seed', type=int, default=1,
help='Random seed to use')
args = parser.parse_args()
# reproducibility
torch.manual_seed(args.seed)
np.random.seed(args.seed)
model_name = 'pcnn_lr:{:.5f}_nr-resnet{}_nr-filters{}'.format(args.lr, args.nr_resnet, args.nr_filters)
assert not os.path.exists(os.path.join('runs', model_name)), '{} already exists!'.format(model_name)
writer = SummaryWriter(log_dir=os.path.join('runs', model_name))
sample_batch_size = 25
obs = (1, 28, 28) if 'mnist' in args.dataset else (3, 32, 32)
input_channels = obs[0]
quantize = lambda x : ((x * 255.) + torch.rand_like(x)) / 256.
rescaling = lambda x : (x - .5) * 2.
rescaling_inv = lambda x : .5 * x + .5
kwargs = {'num_workers':1, 'pin_memory':True, 'drop_last':True}
# ds_transforms = transforms.Compose([transforms.ToTensor(), transforms.RandomCrop(32, padding=4), transforms.RandomHorizontalFlip(), rescaling])
ds_transforms = transforms.Compose([transforms.ToTensor(), quantize, rescaling])
if 'mnist' in args.dataset :
train_loader = torch.utils.data.DataLoader(datasets.MNIST(args.data_dir, download=True,
train=True, transform=ds_transforms), batch_size=args.batch_size,
shuffle=True, **kwargs)
test_loader = torch.utils.data.DataLoader(datasets.MNIST(args.data_dir, train=False,
transform=ds_transforms), batch_size=args.batch_size, shuffle=True, **kwargs)
loss_op = lambda real, nn_out : gaussian_loss_1d(real, nn_out)
sample_op = lambda nn_out : sample_from_gaussian_1d(nn_out)
elif 'cifar' in args.dataset :
train_loader = torch.utils.data.DataLoader(datasets.CIFAR10(args.data_dir, train=True,
download=True, transform=ds_transforms), batch_size=args.batch_size, shuffle=True, **kwargs)
test_loader = torch.utils.data.DataLoader(datasets.CIFAR10(args.data_dir, train=False,
transform=ds_transforms), batch_size=args.batch_size, shuffle=True, **kwargs)
loss_op = lambda real, fake : discretized_mix_logistic_loss(real, fake)
sample_op = lambda x : sample_from_discretized_mix_logistic(x, args.nr_logistic_mix)
else :
raise Exception('{} dataset not in {mnist, cifar10}'.format(args.dataset))
model = PixelCNN(nr_resnet=args.nr_resnet, nr_filters=args.nr_filters,
input_channels=input_channels, nr_logistic_mix=args.nr_logistic_mix, gauss_out=True)
model = model.cuda()
if args.load_params:
load_part_of_model(model, args.load_params)
# model.load_state_dict(torch.load(args.load_params))
print('model parameters loaded')
optimizer = optim.Adam(model.parameters(), lr=args.lr)
scheduler = lr_scheduler.StepLR(optimizer, step_size=1, gamma=args.lr_decay)
def sample(model):
model.train(False)
with torch.no_grad():
data = torch.zeros(sample_batch_size, obs[0], obs[1], obs[2])
data = data.cuda()
for i in range(obs[1]):
for j in range(obs[2]):
# data_v = Variable(data, volatile=True)
data_v = data
out = model(data_v, sample=True)
out_sample = sample_op(out)
data[:, :, i, j] = out_sample.data[:, :, i, j]
return data
print('starting training')
writes = 0
for epoch in range(args.max_epochs):
model.train(True)
torch.cuda.synchronize()
train_loss = 0.
time_ = time.time()
model.train()
for batch_idx, (input,_) in enumerate(train_loader):
input = input.cuda()
# input = Variable(input)
output = model(input)
loss = loss_op(input, output)
optimizer.zero_grad()
loss.backward()
optimizer.step()
train_loss += loss.item()
if (batch_idx +1) % args.print_every == 0 :
deno = args.print_every * args.batch_size * np.prod(obs) * np.log(2.)
writer.add_scalar('train/bpd', (train_loss / deno), writes)
print('loss : {:.4f}, time : {:.4f}'.format(
(train_loss / deno),
(time.time() - time_)))
train_loss = 0.
writes += 1
time_ = time.time()
# decrease learning rate
scheduler.step()
torch.cuda.synchronize()
model.eval()
with torch.no_grad():
test_loss = 0.
for batch_idx, (input,_) in enumerate(test_loader):
input = input.cuda()
# input_var = Variable(input)
input_var = input
output = model(input_var)
loss = loss_op(input_var, output)
test_loss += loss.item()
# del loss, output
deno = batch_idx * args.batch_size * np.prod(obs) * np.log(2.)
writer.add_scalar('test/bpd', (test_loss / deno), writes)
writer.add_scalar('testBPDbyEpoch', (test_loss / deno), epoch)
print('test loss : %s' % (test_loss / deno))
if (epoch + 1) % args.save_interval == 0:
if not os.path.exists("./models"):
os.makedirs("./models")
torch.save(model.state_dict(), 'models/{}_{}.pth'.format(model_name, epoch))
print('sampling...')
sample_t = sample(model)
sample_t = rescaling_inv(sample_t)
utils.save_image(sample_t,'images/{}_{}.png'.format(model_name, epoch),
nrow=5, padding=0)