-
Notifications
You must be signed in to change notification settings - Fork 1
/
improved_model.py
168 lines (117 loc) · 6.09 KB
/
improved_model.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
from __future__ import print_function
import tensorflow as tf
import data_reader
import numpy as np
from tensorflow.contrib import rnn
# Hyper-params
learning_rate = 0.001
training_steps = 1200
batch_size = 128
display_step = 1
beta = 0.001
num_input = 6 * 2 # Prosody
timesteps = 1200 # 60 sec * 20 frames/sec = 1200
num_hidden = 30 # num units in LSTM cell
keep_prob_train = 0.75
experiments = [5,10,20,40,60]
for experiment in experiments:
tf.reset_default_graph()
num_output_units = experiment # 20 frames/sec
# Reading data
print("Reading data...")
x_train, y_train, x_test, y_test = data_reader.get_data()
y_train = y_train[:,:,0:num_output_units]
y_test = y_test[:,:,0:num_output_units]
print(x_train.shape)
# tf Graph input
X = tf.placeholder("float", [None, timesteps, num_input])
Y = tf.placeholder("float", [None, timesteps, num_output_units])
keep_prob = tf.placeholder(tf.float32) # dropout (keep probability)
# Define weights/biases
weights = {
'hidden1': tf.get_variable("w_hid1", shape=(num_input, num_input),
# initializer=tf.random_normal_initializer()),
initializer=tf.contrib.layers.xavier_initializer()),
'hidden2': tf.get_variable("w_hid2", shape=(num_input, num_input),
# initializer=tf.random_normal_initializer()),
initializer=tf.contrib.layers.xavier_initializer()),
'out': tf.get_variable("w_out", shape=[num_hidden, num_output_units],
initializer=tf.contrib.layers.xavier_initializer())
}
biases = {
'hidden1': tf.get_variable("b_hid1", shape=[num_input],
initializer=tf.contrib.layers.xavier_initializer()),
'hidden2': tf.get_variable("b_hid2", shape=[num_input],
initializer=tf.contrib.layers.xavier_initializer()),
'out': tf.get_variable("b_out", shape=[num_output_units],
initializer=tf.contrib.layers.xavier_initializer())
}
def parametric_relu(_x, name):
alpha = tf.get_variable(name, _x.get_shape()[-1],
initializer=tf.constant_initializer(0.1),
dtype=_x.dtype)
pos = tf.nn.relu(_x)
neg = alpha * (_x - abs(_x)) * 0.5
return pos + neg
def build_lstm_rnn(x, weights, biases):
# Prepare data shape to match `rnn` function requirements
# Current data input shape: (batch_size, timesteps, num_input)
# Required structure: list of size 'timesteps', where each item in the list is a tensor of shape:
# (batch_size, num_input)
# Unstack to get a list of 'timesteps' tensors of shape (batch_size, n_input)
x = tf.reshape(x, [-1, num_input])
x = tf.nn.bias_add(tf.matmul(x, weights['hidden1']), biases['hidden1']) # Linear activation
x = parametric_relu(x, "alpha_h1")
x = tf.nn.dropout(x, keep_prob)
x = tf.nn.bias_add(tf.matmul(x, weights['hidden2']), biases['hidden2']) # Linear activation
x = parametric_relu(x, "alpha_h2")
x = tf.nn.dropout(x, keep_prob)
x = tf.reshape(x, [-1, timesteps, num_input])
x = tf.unstack(x, timesteps, 1)
# Basic LSTM Cell with num_hidden units
lstm_cell = rnn.BasicLSTMCell(num_hidden, forget_bias=1.0)
# Skantze used a static structure instead of a dynamic one. Why?
# outputs is a list of tensors of shape (batch_size, num_hidden). Size of the list: timesteps
outputs, states = rnn.static_rnn(lstm_cell, x, dtype=tf.float32)
# Reshape 'outputs' to be a 2D matrix, so we can perform outputs * weights
outputs = tf.reshape(outputs, [-1, num_hidden])
sigmoid_acts = parametric_relu(tf.sigmoid(tf.add(tf.matmul(outputs, weights['out']), biases['out'])))
# Reshape the result back to (timesteps, num_examples, num_output_units)
sigmoid_acts = tf.reshape(sigmoid_acts, [timesteps, -1, num_output_units])
# The shape of the prediction must be (num_examples, timesteps, num_output_units)
sigmoid_acts = tf.transpose(sigmoid_acts, [1,0,2])
return sigmoid_acts
print("Building graph...")
net_out = build_lstm_rnn(X, weights, biases)
# Define loss and optimizer
loss_op = tf.reduce_mean(tf.squared_difference(net_out, Y))
regularizer = tf.nn.l2_loss(weights['out'])
loss_op = tf.reduce_mean(loss_op + beta * regularizer)
optimizer = tf.train.RMSPropOptimizer(learning_rate=learning_rate)
train_op = optimizer.minimize(loss_op)
# Mean Absolute Error operation to measure performance
mae_op = tf.reduce_mean(tf.abs(Y - tf.minimum(tf.maximum(net_out, 0), 1)))
# Initialize the variables (i.e. assign their default value)
init = tf.global_variables_initializer()
print("Starting session... Num output units: ", experiment)
# Add ops to save and restore all the variables.
saver = tf.train.Saver()
# Start training
with tf.Session() as sess:
# saver.restore(sess, "toyota_sigmoid_model" + str(experiment) + ".ckpt")
# Run the initializer
sess.run(init)
num_batches = int(x_train.shape[0] / batch_size)
for epoch in range(training_steps):
for step in range(num_batches):
idx = np.random.randint(x_train.shape[0], size=batch_size)
batch_x = x_train[idx, :, :]
batch_y = y_train[idx, :, :]
# Run optimization op (backprop)
sess.run(train_op, feed_dict={X: batch_x, Y: batch_y, keep_prob: keep_prob_train})
epoch_mae = sess.run([mae_op], feed_dict={X: x_test, Y: y_test, keep_prob: 1.0})
# epoch_cost /= num_batches
print("Epoch " + str(epoch) + ", Epoch MAE= " + str(epoch_mae))
print("Optimization Finished!")
save_path = saver.save(sess, "./Improved_Model_" + str(experiment) + ".ckpt")
print("Model saved in file: %s" % save_path)