-
Notifications
You must be signed in to change notification settings - Fork 0
/
train.py
144 lines (114 loc) · 5.56 KB
/
train.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
import argparse
import os
import numpy as np
import cntk as C
from cntk import Trainer
from cntk.learners import (adam, UnitType, learning_rate_schedule,
momentum_as_time_constant_schedule, momentum_schedule)
from cntk.logging import *
import matplotlib.pyplot as plt
from utils import *
from model import *
parser = argparse.ArgumentParser()
parser.add_argument('--lr', type=float, default=0.001, help='learning rate, default=0.001')
parser.add_argument('--mm', type=float, default=0.5, help='beta1 for adam. default=0.5')
parser.add_argument('--mm_var', type=float, default=0.999, help='beta2 for adam. default=0.999')
parser.add_argument('--z_dim', type=int, default=64, help='dimension of seed vector. default=64')
parser.add_argument('--kt', type=float, default=0.0, help='parameter kt. default=0.0')
parser.add_argument('--gamma', type=float, default=1.0, help='parameter gamma. default=1.0')
parser.add_argument('--lamda', type=float, default=0.001, help='parameter lambda. default=0.001')
parser.add_argument('--batchSize', type=int, default=64, help='mini-batch size. default=64')
parser.add_argument('--imageSize', type=int, default=28, help='height / width of the input image to network. default=28')
parser.add_argument('--dataPath', default='Train-28x28_cntk_text.txt', help='Data path. default=Train-28x28_cntk_text.txt')
opt = parser.parse_args()
num_minibatches = 300000 # iterations
lr_update_step = 60000
print_frequency_mbsize = 1000
input_dim = opt.imageSize ** 2 # assumption: height==width
noise_shape = opt.z_dim
kt = opt.kt
if __name__=='__main__':
check_path(opt.dataPath)
reader_train = create_reader(opt.dataPath, True, input_dim)
##
input_dynamic_axes = [C.Axis.default_batch_axis()]
Z = C.input_variable(noise_shape, dynamic_axes=input_dynamic_axes)
X_real = C.input_variable(input_dim, dynamic_axes=input_dynamic_axes)
X_real_scaled = X_real / 127.5 -1
kt_in = C.constant(kt)
# Create the model function for the generator and discriminator models
X_fake = generator(Z, opt)
D_real = discriminator(X_real_scaled, opt)
D_fake = D_real.clone(
method = 'share',
substitutions = {X_real_scaled.output: X_fake.output})
# Create loss functions and configure optimazation algorithms
D_real_loss = l1_loss(X_real_scaled, D_real)
D_fake_loss = l1_loss(X_fake, D_fake)
G_loss = D_fake_loss
D_loss = D_real_loss - D_fake_loss * kt_in
lr_schedule = list( opt.lr * np.asarray([0.95**t for t in range(0, num_minibatches//lr_update_step+1)]) )
G_learner = adam(
parameters = X_fake.parameters,
lr = learning_rate_schedule(lr_schedule, UnitType.sample, lr_update_step*opt.batchSize),
momentum = momentum_schedule(opt.mm),
variance_momentum = momentum_schedule(opt.mm_var)
)
D_learner = adam(
parameters = D_real.parameters,
lr = learning_rate_schedule(lr_schedule, UnitType.sample, lr_update_step*opt.batchSize),
momentum = momentum_schedule(opt.mm),
variance_momentum = momentum_schedule(opt.mm_var)
)
pp_G = ProgressPrinter(print_frequency_mbsize, metric_is_pct=False)
pp_D = ProgressPrinter(print_frequency_mbsize, metric_is_pct=False)
tensorboard_logdir = 'log/'
tb = TensorBoardProgressWriter(freq = print_frequency_mbsize, log_dir=tensorboard_logdir)
# Instantiate the trainers
G_trainer = Trainer(
X_fake,
(G_loss, None),
G_learner
)
D_trainer = Trainer(
D_real,
(D_loss, None),
D_learner
)
##
input_map = {X_real: reader_train.streams.features}
m_global_pre = 10
X_fake_node = C.combine([X_fake.owner])
sample_seed = noise_sample(25, opt.z_dim)
exists_or_mkdir('./samples')
for train_step in range(num_minibatches):
Z_data = noise_sample(opt.batchSize, opt.z_dim)
batch_inputs = {Z: Z_data}
G_trainer.train_minibatch(batch_inputs)
Z_data = noise_sample(opt.batchSize, opt.z_dim)
X_data = reader_train.next_minibatch(opt.batchSize, input_map)
batch_inputs = {X_real: X_data[X_real].data, Z: Z_data}
D_trainer.train_minibatch(batch_inputs)
pp_G.update_with_trainer(G_trainer)
pp_D.update_with_trainer(D_trainer)
temp = C.combine([C.reduce_mean(D_real_loss, axis=C.Axis.all_axes()),
C.reduce_mean(D_fake_loss, axis=C.Axis.all_axes())]
).eval({X_real: X_data[X_real].data, Z: Z_data})
val = list(temp.values())
kt = np.clip(kt + opt.lamda*(opt.gamma * val[0] - val[1]), 0, 1).astype(np.float32)
C.assign(kt_in, kt).eval()
m_global = val[0] + np.abs(opt.gamma * val[0] - val[1])
tb.write_value("m_global", m_global, train_step)
tb.write_value("kt", kt, train_step)
if train_step % 1000 == 0:
output = np.rint((X_fake_node.eval({Z: sample_seed}) + 1) * 127.5)
output = np.clip(output, 0, 255)
output = output.astype(np.uint8)
output = np.reshape(output, [-1, opt.imageSize, opt.imageSize])
save_images(output, [5, 5], './samples/{:05d}.png'.format(train_step))
if train_step % 100 == 0 and m_global < m_global_pre:
G_trainer.save_checkpoint('models/BEGAN_G_{}.dnn'.format(train_step))
D_trainer.save_checkpoint('models/BEGAN_D_{}.dnn'.format(train_step))
m_global_pre = m_global
X_fake_output = X_fake_node.eval({Z: noise_sample(36, opt.z_dim)})
plot_images(X_fake_output, subplot_shape=[6, 6])