forked from Kyubyong/tacotron
-
Notifications
You must be signed in to change notification settings - Fork 3
/
train_multi_gpus.py
131 lines (105 loc) · 6.15 KB
/
train_multi_gpus.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
# -*- coding: utf-8 -*-
#/usr/bin/python2
'''
By kyubyong park. [email protected].
https://www.github.com/kyubyong/tacotron
'''
from __future__ import print_function
import tensorflow as tf
import numpy as np
import librosa
import os
from tqdm import tqdm
from hyperparams import Hyperparams as hp
from prepro import *
from networks import encode, decode1, decode2
from modules import *
from data_load import get_batch
from utils import shift_by_one
class Graph:
def __init__(self, is_training=True):
self.graph = tf.Graph()
with self.graph.as_default():
if is_training:
self.x, self.y, self.z, self.num_batch = get_batch()
self.decoder_inputs = shift_by_one(self.y)
# Make sure that batch size was multiplied by # gpus.
# Now we split the mini-batch data by # gpus.
self.x = tf.split(self.x, hp.num_gpus, 0)
self.y = tf.split(self.y, hp.num_gpus, 0)
self.z = tf.split(self.z, hp.num_gpus, 0)
self.decoder_inputs = tf.split(self.decoder_inputs, hp.num_gpus, 0)
# optimizer
self.optimizer = tf.train.AdamOptimizer(learning_rate=hp.lr)
self.losses, self.grads_and_vars_list = [], []
for i in range(hp.num_gpus):
with tf.variable_scope('net', reuse=bool(i)):
with tf.device('/gpu:{}'.format(i)):
with tf.name_scope('gpu_{}'.format(i)):
# Encoder
self.memory = encode(self.x[i], is_training=is_training) # (N, T, E)
# Decoder
self.outputs1 = decode1(self.decoder_inputs[i], self.memory, is_training=is_training) # (N, T', hp.n_mels*hp.r)
self.outputs2 = decode2(self.outputs1, is_training=is_training) # (N, T', (1+hp.n_fft//2)*hp.r)
# Loss
if hp.loss_type=="l1": # L1 loss
self.loss1 = tf.abs(self.outputs1 - self.y[i])
self.loss2 = tf.abs(self.outputs2 - self.z[i])
else: # L2 loss
self.loss1 = tf.squared_difference(self.outputs1, self.y[i])
self.loss2 = tf.squared_difference(self.outputs2, self.z[i])
# Target masking
if hp.target_zeros_masking:
self.loss1 *= tf.to_float(tf.not_equal(self.y[i], 0.))
self.loss2 *= tf.to_float(tf.not_equal(self.z[i], 0.))
self.mean_loss1 = tf.reduce_mean(self.loss1)
self.mean_loss2 = tf.reduce_mean(self.loss2)
self.mean_loss = self.mean_loss1 + self.mean_loss2
self.losses.append(self.mean_loss)
self.grads_and_vars = self.optimizer.compute_gradients(self.mean_loss)
self.grads_and_vars_list.append(self.grads_and_vars)
with tf.device('/cpu:0'):
# Aggregate losses, then calculate average loss.
self.loss = tf.add_n(self.losses) / len(self.losses)
#Aggregate gradients, then calculate average gradients.
self.mean_grads_and_vars = []
for grads_and_vars in zip(*self.grads_and_vars_list):
grads = []
for grad, var in grads_and_vars:
grads.append(tf.expand_dims(grad, 0))
mean_grad = tf.reduce_mean(tf.concat(grads, 0), 0) #()
self.mean_grads_and_vars.append((mean_grad, var))
# Training Scheme
self.global_step = tf.Variable(0, name='global_step', trainable=False)
self.train_op = self.optimizer.apply_gradients(self.mean_grads_and_vars, self.global_step)
# Summmary
tf.summary.scalar('loss', self.loss)
self.merged = tf.summary.merge_all()
else: # Evaluation
self.x = tf.placeholder(tf.int32, shape=(None, None))
self.decoder_inputs = tf.placeholder(tf.float32, shape=(None, None, hp.n_mels*hp.r))
with tf.variable_scope('net'):
# Encoder
self.memory = encode(self.x, is_training=is_training) # (N, T, E)
# Decoder
self.outputs1 = decode1(self.decoder_inputs, self.memory, is_training=is_training) # (N, T', hp.n_mels*hp.r)
self.outputs2 = decode2(self.outputs1, is_training=is_training) # (N, T', (1+hp.n_fft//2)*hp.r)
def main():
g = Graph(); print("Training Graph loaded")
with g.graph.as_default():
# Load vocabulary
char2idx, idx2char = load_vocab()
# Training
sv = tf.train.Supervisor(logdir=hp.logdir,
save_model_secs=0)
with sv.managed_session(config=tf.ConfigProto(allow_soft_placement=True)) as sess:
for epoch in range(1, hp.num_epochs+1):
if sv.should_stop(): break
for step in tqdm(range(g.num_batch), total=g.num_batch, ncols=70, leave=False, unit='b'):
sess.run(g.train_op)
# Write checkpoint files at every epoch
gs = sess.run(g.global_step)
sv.saver.save(sess, hp.logdir + '/model_epoch_%02d_gs_%d' % (epoch, gs))
if __name__ == '__main__':
main()
print("Done")