forked from zhongshaoyy/Audio2Face
-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathmain.py
166 lines (134 loc) · 6.2 KB
/
main.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
'''
virtualenv: 63server pytorch
author: Yachun Li ([email protected])
'''
import torch
import torch.autograd as autograd
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
import os
import shutil
import time
from datetime import datetime
from dataset import BlendshapeDataset
from models import NvidiaNet, LSTMNvidiaNet, FullyLSTM
from models_testae import *
# gpu setting
gpu_id = 1
os.environ['CUDA_VISIBLE_DEVICES'] = str(gpu_id)
# hyper-parameters
n_blendshape = 51
learning_rate = 0.0001
batch_size = 100
epochs = 500
print_freq = 20
best_loss = 10000000
# data path
dataroot = '/home/liyachun/data/audio2bs'
# data_path = os.path.join(dataroot, audio2bs)
data_path = dataroot
checkpoint_path = './checkpoint-lstmae-2distconcat_kl001/'
if not os.path.isdir(checkpoint_path): os.mkdir(checkpoint_path)
# Reconstruction + KL divergence losses summed over all elements and batch
def loss_function(recon_x, x, mu, logvar):
# BCE = F.binary_cross_entropy(recon_x, x, size_average=False)
MSE = F.mse_loss(recon_x, x)
# see Appendix B from VAE paper:
# Kingma and Welling. Auto-Encoding Variational Bayes. ICLR, 2014
# https://arxiv.org/abs/1312.6114
# 0.5 * sum(1 + log(sigma^2) - mu^2 - sigma^2)
KLD = -0.5 * torch.sum(1 + logvar - mu.pow(2) - logvar.exp())
# print('loss percent: MSE %.4f(%.6f), KLD %.4f, total %.4f'
# % (MSE.data[0], MSE.data[0]/(MSE.data[0]+KLD.data[0]), KLD.data[0], MSE.data[0]+KLD.data[0]))
return MSE + 0.01*KLD
def main():
global best_loss
model = LSTMAE2dist(is_concat=True)
print(model)
# model = nn.DataParallel(model)
# get data
train_loader = torch.utils.data.DataLoader(
BlendshapeDataset(feature_file=os.path.join(data_path, 'feature/0201train-mfcc39.npy'),
target_file=os.path.join(data_path, 'blendshape/0201train.txt')),
batch_size=batch_size, shuffle=True, num_workers=2
)
val_loader = torch.utils.data.DataLoader(
BlendshapeDataset(feature_file=os.path.join(data_path, 'test/feature/0201_06-1min-mfcc39.npy'),
target_file=os.path.join(data_path, 'test/blendshape/0201_06-1min.txt')),
batch_size=batch_size, shuffle=False, num_workers=2
)
# define loss and optimiser
criterion = nn.MSELoss() #??.cuda()
optimizer = optim.Adam(model.parameters(), lr=learning_rate)
if torch.cuda.is_available():
model = model.cuda()
# training
print('------------\n Training begin at %s' % datetime.now())
for epoch in range(epochs):
start_time = time.time()
model.train()
train_loss = 0.
for i, (input, target) in enumerate(train_loader):
target = target.cuda(async=True)
input_var = autograd.Variable(input.float()).cuda()
target_var = autograd.Variable(target.float())
# compute model output
# audio_z, bs_z, output = model(input_var, target_var)
# loss = criterion(output, target_var)
audio_z, bs_z, output, mu, logvar = model(input_var, target_var) # method2: loss change
loss = loss_function(output, target_var, mu, logvar)
train_loss += loss.data[0]
# compute gradient and do the backpropagate
optimizer.zero_grad()
loss.backward()
optimizer.step()
# if i % print_freq == 0:
# print('Training -- epoch: {} | iteration: {}/{} | loss: {:.6f} \r'
# .format(epoch+1, i, len(train_loader), loss.data[0]))
train_loss /= len(train_loader)
print('Glance at training z: max/min of hidden audio(%.4f/%.4f), blendshape(%.4f/%.4f)'
% (max(audio_z.data[0]), min(audio_z.data[0]), max(bs_z.data[0]), min(bs_z.data[0])))
model.eval()
eval_loss = 0.
for input, target in val_loader:
target = target.cuda(async=True)
input_var = autograd.Variable(input.float(), volatile=True).cuda()
target_var = autograd.Variable(target.float(), volatile=True)
# compute output temporal?!!
# audio_z, bs_z, output = model(input_var, target_var)
# loss = criterion(output, target_var)
audio_z, bs_z, output, mu, logvar = model(input_var, target_var) # method2: loss change
loss = loss_function(output, target_var, mu, logvar)
eval_loss += loss.data[0]
eval_loss /= len(val_loader)
# count time of 1 epoch
past_time = time.time() - start_time
print('Glance at validating z: max/min of hidden audio(%.4f/%.4f), blendshape(%.4f/%.4f)'
% (max(audio_z.data[0]), min(audio_z.data[0]), max(bs_z.data[0]), min(bs_z.data[0])))
# print('Evaluating -- epoch: {} | loss: {:.6f} \r'.format(epoch+1, eval_loss/len(val_loader)))
print('epoch: {:03} | train_loss: {:.6f} | eval_loss: {:.6f} | {:.4f} sec/epoch \r'
.format(epoch+1, train_loss, eval_loss, past_time))
# save best model on val
is_best = eval_loss < best_loss
best_loss = min(eval_loss, best_loss)
if is_best:
torch.save({
'epoch': epoch + 1,
'state_dict': model.state_dict(),
'eval_loss': best_loss,
}, checkpoint_path+'model_best.pth.tar')
# save models every 100 epoch
if (epoch+1) % 100 == 0:
torch.save({
'epoch': epoch + 1,
'state_dict': model.state_dict(),
'eval_loss': eval_loss,
}, checkpoint_path+'checkpoint-epoch'+str(epoch+1)+'.pth.tar')
print('Training finished at %s' % datetime.now())
# def save_checkpoint(state, is_best, filename='checkpoint.pth.tar'):
# torch.save(state, checkpoint_path+filename)
# if is_best:
# shutil.copyfile(checkpoint_path+filename, checkpoint_path+'model_best.pth.tar')
if __name__ == '__main__':
main()