-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathtrain-gqn.py
130 lines (92 loc) · 3.92 KB
/
train-gqn.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
import tensorflow as tf
import tensorflow.contrib.summary as summary
from gqn import GenerativeQueryNetwork
from gqn_datasets.data_reader import DataReader
from utils import *
import os
# Training hyper-parameters
mu_i = 5e-4
mu_f = 5e-5
mu_n = 1.6e6
sigma_i = 2.0
sigma_f = 0.7
sigma_n = 4e4 # 2e5
batch_size = 36
test_batch_size = 10
context_size = 5
S_max = int(2e6)
# Model hyper-parameters
im_size = 256
x_dim = 3
r_dim = 64 # 256
h_dim = 64 # 256
z_dim = 32 # 256
L = 12
# Overhead
root = "/localdata/auguste/gqn-dataset"
logs_path = "/localdata/auguste/logs_test"
if __name__ == '__main__':
session_name = get_session_name()
session_logs_path = os.path.join(logs_path, session_name)
global_step = tf.train.get_or_create_global_step()
train_data_reader = DataReader(
'shepard_metzler_5_parts', batch_size, context_size, root)
test_data_reader = DataReader(
'shepard_metzler_5_parts', test_batch_size,
context_size, root, mode='test'
)
model = GenerativeQueryNetwork(x_dim, r_dim, h_dim, z_dim)
lr = tf.train.polynomial_decay(mu_i, global_step, mu_n, mu_f)
sigma_s = tf.train.polynomial_decay(sigma_i, global_step, sigma_n, sigma_f)
optimizer = tf.train.AdamOptimizer(learning_rate=lr)
writer = summary.create_file_writer(session_logs_path, max_queue=1)
writer.set_as_default()
with summary.record_summaries_every_n_global_steps(500):
batch = train_data_reader.read()
x_mu, x_q, r, kl = model(batch)
output_dist = tf.distributions.Normal(loc=x_mu, scale=sigma_s)
log_likelihood = tf.reduce_logsumexp(output_dist.log_prob(x_q))
loss = kl - log_likelihood
optimize = optimizer.minimize(loss, global_step=global_step)
output_dist_const_var = tf.distributions.Normal(
loc=x_mu, scale=sigma_f)
log_likelihood_const_var = tf.reduce_logsumexp(
output_dist_const_var.log_prob(x_q))
summary.scalar(
"log-likelihood", log_likelihood, family="train")
summary.scalar(
"log-likelihood constant variance",
log_likelihood_const_var, family="train"
)
summary.image("inference output", cast_im(x_mu[0:3]), max_images=3)
summary.image("inference target", cast_im(x_q[0:3]), max_images=3)
summary.scalar("learning_rate", lr, family="hyper-parameters")
summary.scalar("sigma_s", sigma_s, family="hyper-parameters")
# Test set
batch = test_data_reader.read()
x_mu, x_q, r = model.sample(batch)
output_dist = tf.distributions.Normal(loc=x_mu, scale=sigma_s)
log_likelihood = tf.reduce_logsumexp(output_dist.log_prob(x_q))
output_dist_const_var = tf.distributions.Normal(
loc=x_mu, scale=sigma_f)
log_likelihood_const_var = tf.reduce_logsumexp(
output_dist_const_var.log_prob(x_q))
summary.scalar(
"log-likelihood", log_likelihood, family="test")
summary.scalar(
"log-likelihood constant variance",
log_likelihood_const_var, family="test"
)
summary.image("generation output", cast_im(x_mu[0:3]), max_images=3)
summary.image("generation target", cast_im(x_q[0:3]), max_images=3)
context = batch.query.context.frames[0:3]
batch_size, context_size, *x_shape = context.shape
context = tf.reshape(context, [-1, *x_shape])
summary.image("generation context", cast_im(context), max_images=5)
with tf.Session() as sess:
tf.global_variables_initializer().run()
summary.initialize(graph=tf.get_default_graph())
for s in range(S_max):
l, *_ = sess.run([loss, optimize, summary.all_summary_ops()])
if s % 100 == 0:
print("Iteration: {} Loss: {}".format(s, l))