-
Notifications
You must be signed in to change notification settings - Fork 10
/
Copy pathmain_blizzard.py
229 lines (202 loc) · 8.36 KB
/
main_blizzard.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
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
import torch
import torch.nn as nn
from torch.autograd import Variable
import timeit
import argparse
import numpy as np
import os
import random
import load
from models import rnn, wavenet, swavenet
import math;
from blizzard_data import Blizzard_tbptt
def save_checkpoint(state, filename='checkpoint.pth.tar'):
torch.save(state, filename)
def adjust_lr(optimizer, epoch, total_epoch, init_lr, end_lr):
assert init_lr > end_lr;
lr = end_lr + (init_lr - end_lr) * (0.5 * (1+math.cos(math.pi * float(epoch) / total_epoch)));
#print(lr);
for param_group in optimizer.param_groups:
param_group['lr'] = lr
def adjust_kd(epoch, total_epoch, init_kd, end_kd):
if (epoch > total_epoch):
return 1.;
return end_kd + (init_kd - end_kd) * ((math.cos(0.5 * math.pi * float(epoch) / total_epoch)));
def evaluate(dataset, model, args, split='valid'):
model.eval()
loss_sum = 0
cnt = 0;
length = 40
#print(dataset)
for x, y, x_mask in dataset:
#print(x, y, x_mask)
x = Variable(torch.from_numpy(x), volatile=True).float().cuda()
y = Variable(torch.from_numpy(y), volatile=True).float().cuda()
x_mask = Variable(torch.from_numpy(x_mask), volatile=True).float().cuda()
if (args.kld == 'True'):
loss, kld_loss = model([x,y,x_mask]);
total_loss = loss - kld_loss;
total_loss = total_loss.data[0];
else:
all_loss = model([x,y,x_mask]);
total_loss = all_loss.data[0]
loss_sum += total_loss;
cnt += 1;
return -loss_sum/cnt;
parser = argparse.ArgumentParser(description='PyTorch VAE for sequence')
parser.add_argument('--expname', type=str, default='tiny')
parser.add_argument('--seed', type=int, default=1234)
parser.add_argument('--num_epochs', type=int, default=40)
parser.add_argument('--data', type=str, default='/usr1/glai1/datasets/')
parser.add_argument('--file_name', type=str, default='blizzard_tbptt')
parser.add_argument('--lr', type=float, default=1e-4)
parser.add_argument('--end_lr', type=float, default=0.)
parser.add_argument('--kld', type=str, default='True')
parser.add_argument('--model_name', type=str, default='swavenet')
parser.add_argument('--batch_size', type=int, default=128)
parser.add_argument('--gpu', type=int, default=None)
parser.add_argument('--embed_size', type=int, default=1024)
parser.add_argument('--z_size', type=int, default=512)
parser.add_argument('--resume', type=str, default=None)
args = parser.parse_args()
print(args);
seed = args.seed; expname = args.expname; num_epochs = args.num_epochs; data = args.data; lr = args.lr;
model_name = args.model_name;batch_size = args.batch_size;
torch.cuda.set_device(args.gpu)
rng = np.random.RandomState(seed)
torch.manual_seed(seed)
torch.cuda.manual_seed(seed)
torch.cuda.manual_seed_all(seed);
log_interval = 500
model_id = 'blizzard_seed{}'.format(seed)
if not os.path.exists(expname):
os.makedirs(expname)
log_file_name = os.path.join(expname, model_id + '.txt')
model_file_name = os.path.join(expname, model_id + '.pt')
log_file = open(log_file_name, 'w')
print('Loading data..')
#X_mean = -0.35828044227
#X_std = 3117.59272379
bsz = 128
#file_name = 'blizzard_unseg_tbptt'
normal_path = "os.path.join(data, args.file_name + '_normal.npz')"
normal_params = np.load(os.path.join(data, args.file_name + '_normal.npz'))
X_mean = normal_params['X_mean']
X_std = normal_params['X_std']
train_data = Blizzard_tbptt(name='train',
path=args.data,
frame_size=200,
file_name=args.file_name,
X_mean=X_mean,
X_std=X_std)
valid_data = Blizzard_tbptt(name='valid',
path=args.data,
frame_size=200,
file_name=args.file_name,
X_mean=X_mean,
X_std=X_std)
test_data = Blizzard_tbptt(name='test',
path=args.data,
frame_size=200,
file_name=args.file_name,
X_mean=X_mean,
X_std=X_std)
print('X_mean: %f, X_std: %f', (X_mean, X_std))
assert bsz == 128
train_data = load.BlizzardIterator(train_data, bsz, start=0, end=2040064)
valid_data = load.BlizzardIterator(valid_data, bsz, start=2040064, end=2152704)
# Use complete batch only.
test_data = load.BlizzardIterator(test_data, bsz, start=2152704, end=2267008-128)
print('Done.')
# Build model
nbatches = train_data.nbatch
total_step = num_epochs * nbatches;
if args.model_name[:10] == 'wavenetvae':
model = eval(args.model_name).Model(input_dim=200, embed_dim=args.embed_size, z_dim=args.z_size, data=None)
else:
model = eval(args.model_name).Model(input_dim=200, embed_dim=512, output_dim=1024, data=None)
model.cuda()
opt = torch.optim.Adam(model.parameters(), lr=lr, eps=1e-5)
kld_step = 0.00005
if (args.resume == None):
step = 0
kld_weight = kld_step;
start_epoch = 0;
else:
if os.path.isfile(args.resume):
print("=> loading checkpoint '{}'".format(args.resume))
checkpoint = torch.load(args.resume)
start_epoch = checkpoint['epoch']
model.load_state_dict(checkpoint['state_dict'])
opt.load_state_dict(checkpoint['optimizer'])
print("=> loaded checkpoint '{}' (epoch {})"
.format(args.resume, checkpoint['epoch']))
step = (start_epoch) * nbatches;
kld_weight = adjust_kd(step, 20 * nbatches, kld_step, 1.);
else:
print("=> no checkpoint found at '{}'".format(args.resume))
nParams = sum([p.nelement() for p in model.parameters()])
print('* number of parameters: %d' % nParams)
t = timeit.default_timer()
loss_sum = 0;
kld_loss_sum = 0;
logp_loss_sum = 0;
flag = True;
#print('test', evaluate(test_data, model, args, split='test'))
for epoch in range(start_epoch, num_epochs):
old_valid_loss = np.inf
model.train()
print('Epoch {}: ({})'.format(epoch, model_id.upper()))
for x, y, x_mask in train_data:
opt.zero_grad()
x = Variable(torch.from_numpy(x)).float().cuda()
y = Variable(torch.from_numpy(y)).float().cuda()
x_mask = Variable(torch.from_numpy(x_mask)).float().cuda()
if (args.kld == 'True'):
loss, kld_loss = model([x,y,x_mask]);
total_loss = loss - kld_loss * kld_weight;
if np.isnan(total_loss.data[0]) or np.isinf(total_loss.data[0]):
print("NaN") # Useful to see if training is stuck.
flag = False;
break
total_loss.backward();
total_loss = total_loss.data[0];
kld_loss_sum += kld_loss.data[0];
logp_loss_sum += loss.data[0];
else:
all_loss = model([x,y,x_mask]);
if np.isnan(all_loss.data[0]) or np.isinf(all_loss.data[0]):
continue
all_loss.backward()
total_loss = all_loss.data[0]
torch.nn.utils.clip_grad_norm(model.parameters(), 0.1, 'inf')
opt.step()
loss_sum += total_loss;
step += 1;
if step % log_interval == 0:
s = timeit.default_timer()
log_line = 'total time: [%f], epoch: [%d/%d], step: [%d/%d], loss: %f, logp_loss:%f, kld_loss: %f, actual_loss:%f' % (
s-t, epoch, num_epochs, step % nbatches, nbatches,
-loss_sum / log_interval, -logp_loss_sum/log_interval, -kld_loss_sum/log_interval, -(logp_loss_sum-kld_loss_sum)/log_interval)
print(log_line)
log_file.write(log_line + '\n')
log_file.flush()
loss_sum = 0;
kld_loss_sum = 0;
logp_loss_sum = 0;
kld_weight = adjust_kd(step, 20 * nbatches, kld_step, 1.);
adjust_lr(opt, step, total_step, args.lr, args.end_lr);
if flag == False:
break;
print(float(step)/total_step);
# evaluate per epoch
print('--- Epoch finished ----')
val_loss = evaluate(valid_data, model, args)
log_line = 'validation -- epoch: %s, nll: %f' % (epoch, val_loss)
print(log_line)
log_file.write(log_line + '\n')
test_loss = evaluate(test_data, model, args, split='test')
log_line = 'test -- epoch: %s, nll: %f' % (epoch, test_loss)
print(log_line)
log_file.write(log_line + '\n')
log_file.flush()