-
Notifications
You must be signed in to change notification settings - Fork 80
/
train_ali.py
228 lines (200 loc) · 10.1 KB
/
train_ali.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
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
import argparse
import logging
from blocks.algorithms import Adam
from blocks.bricks import LeakyRectifier, Logistic
from blocks.bricks.conv import ConvolutionalSequence
from blocks.extensions import FinishAfter, Timing, Printing, ProgressBar
from blocks.extensions.monitoring import DataStreamMonitoring
from blocks.extensions.saveload import Checkpoint
from blocks.filter import VariableFilter
from blocks.graph import ComputationGraph, apply_dropout
from blocks.graph.bn import (batch_normalization,
get_batch_normalization_updates)
from blocks.initialization import IsotropicGaussian, Constant
from blocks.main_loop import MainLoop
from blocks.model import Model
from blocks.roles import INPUT
from theano import tensor
from ali.algorithms import ali_algorithm
from ali.bricks import (ALI, GaussianConditional, DeterministicConditional,
XZJointDiscriminator)
from ali.streams import create_celeba_data_streams
from ali.utils import get_log_odds, conv_brick, conv_transpose_brick, bn_brick
from wrapper.interface import AliModel
from utils.samplecheckpoint import SampleCheckpoint
from utils.fuel_helper import create_custom_streams
NUM_EPOCHS = 123
IMAGE_SIZE = (64, 64)
NUM_CHANNELS = 3
NLAT = 256
GAUSSIAN_INIT = IsotropicGaussian(std=0.01)
ZERO_INIT = Constant(0)
LEARNING_RATE = 1e-4
BETA1 = 0.5
LEAK = 0.02
def create_model_brick(model_stream):
layers = [
conv_brick(2, 1, 64), bn_brick(), LeakyRectifier(leak=LEAK),
conv_brick(7, 2, 128), bn_brick(), LeakyRectifier(leak=LEAK),
conv_brick(5, 2, 256), bn_brick(), LeakyRectifier(leak=LEAK),
conv_brick(7, 2, 256), bn_brick(), LeakyRectifier(leak=LEAK),
conv_brick(4, 1, 512), bn_brick(), LeakyRectifier(leak=LEAK),
conv_brick(1, 1, 2 * NLAT)]
encoder_mapping = ConvolutionalSequence(
layers=layers, num_channels=NUM_CHANNELS, image_size=IMAGE_SIZE,
use_bias=False, name='encoder_mapping')
encoder = GaussianConditional(encoder_mapping, name='encoder')
layers = [
conv_transpose_brick(4, 1, 512), bn_brick(), LeakyRectifier(leak=LEAK),
conv_transpose_brick(7, 2, 256), bn_brick(), LeakyRectifier(leak=LEAK),
conv_transpose_brick(5, 2, 256), bn_brick(), LeakyRectifier(leak=LEAK),
conv_transpose_brick(7, 2, 128), bn_brick(), LeakyRectifier(leak=LEAK),
conv_transpose_brick(2, 1, 64), bn_brick(), LeakyRectifier(leak=LEAK),
conv_brick(1, 1, NUM_CHANNELS), Logistic()]
decoder_mapping = ConvolutionalSequence(
layers=layers, num_channels=NLAT, image_size=(1, 1), use_bias=False,
name='decoder_mapping')
decoder = DeterministicConditional(decoder_mapping, name='decoder')
layers = [
conv_brick(2, 1, 64), LeakyRectifier(leak=LEAK),
conv_brick(7, 2, 128), bn_brick(), LeakyRectifier(leak=LEAK),
conv_brick(5, 2, 256), bn_brick(), LeakyRectifier(leak=LEAK),
conv_brick(7, 2, 256), bn_brick(), LeakyRectifier(leak=LEAK),
conv_brick(4, 1, 512), bn_brick(), LeakyRectifier(leak=LEAK)]
x_discriminator = ConvolutionalSequence(
layers=layers, num_channels=NUM_CHANNELS, image_size=IMAGE_SIZE,
use_bias=False, name='x_discriminator')
x_discriminator.push_allocation_config()
layers = [
conv_brick(1, 1, 1024), LeakyRectifier(leak=LEAK),
conv_brick(1, 1, 1024), LeakyRectifier(leak=LEAK)]
z_discriminator = ConvolutionalSequence(
layers=layers, num_channels=NLAT, image_size=(1, 1), use_bias=False,
name='z_discriminator')
z_discriminator.push_allocation_config()
layers = [
conv_brick(1, 1, 2048), LeakyRectifier(leak=LEAK),
conv_brick(1, 1, 2048), LeakyRectifier(leak=LEAK),
conv_brick(1, 1, 1)]
joint_discriminator = ConvolutionalSequence(
layers=layers,
num_channels=(x_discriminator.get_dim('output')[0] +
z_discriminator.get_dim('output')[0]),
image_size=(1, 1),
name='joint_discriminator')
discriminator = XZJointDiscriminator(
x_discriminator, z_discriminator, joint_discriminator,
name='discriminator')
ali = ALI(encoder, decoder, discriminator,
weights_init=GAUSSIAN_INIT, biases_init=ZERO_INIT,
name='ali')
ali.push_allocation_config()
encoder_mapping.layers[-1].use_bias = True
encoder_mapping.layers[-1].tied_biases = False
decoder_mapping.layers[-2].use_bias = True
decoder_mapping.layers[-2].tied_biases = False
x_discriminator.layers[0].use_bias = True
x_discriminator.layers[0].tied_biases = True
ali.initialize()
raw_marginals, = next(model_stream.get_epoch_iterator())
b_value = get_log_odds(raw_marginals)
decoder_mapping.layers[-2].b.set_value(b_value)
return ali
def create_models(model_stream):
ali = create_model_brick(model_stream)
x = tensor.tensor4('features')
z = ali.theano_rng.normal(size=(x.shape[0], NLAT, 1, 1))
def _create_model(with_dropout):
cg = ComputationGraph(ali.compute_losses(x, z))
if with_dropout:
inputs = VariableFilter(
bricks=([ali.discriminator.x_discriminator.layers[0]] +
ali.discriminator.x_discriminator.layers[2::3] +
ali.discriminator.z_discriminator.layers[::2] +
ali.discriminator.joint_discriminator.layers[::2]),
roles=[INPUT])(cg.variables)
cg = apply_dropout(cg, inputs, 0.2)
return Model(cg.outputs)
model = _create_model(with_dropout=False)
with batch_normalization(ali):
bn_model = _create_model(with_dropout=True)
pop_updates = list(
set(get_batch_normalization_updates(bn_model, allow_duplicates=True)))
bn_updates = [(p, m * 0.05 + p * 0.95) for p, m in pop_updates]
return model, bn_model, bn_updates
def create_main_loop(save_path, subdir, dataset, color_convert,
batch_size, monitor_every, checkpoint_every, image_size):
if dataset is None:
streams = create_celeba_data_streams(batch_size, batch_size)
model_stream = create_celeba_data_streams(500, 500)[0]
else:
streams = create_custom_streams(filename=dataset,
training_batch_size=batch_size,
monitoring_batch_size=batch_size,
include_targets=False,
color_convert=color_convert)
model_stream = create_custom_streams(filename=dataset,
training_batch_size=500,
monitoring_batch_size=500,
include_targets=False,
color_convert=color_convert)[0]
main_loop_stream, train_monitor_stream, valid_monitor_stream = streams[:3]
model, bn_model, bn_updates = create_models(model_stream)
ali, = bn_model.top_bricks
discriminator_loss, generator_loss = bn_model.outputs
step_rule = Adam(learning_rate=LEARNING_RATE, beta1=BETA1)
algorithm = ali_algorithm(discriminator_loss, ali.discriminator_parameters,
step_rule, generator_loss,
ali.generator_parameters, step_rule)
algorithm.add_updates(bn_updates)
bn_monitored_variables = (
[v for v in bn_model.auxiliary_variables if 'norm' not in v.name] +
bn_model.outputs)
monitored_variables = (
[v for v in model.auxiliary_variables if 'norm' not in v.name] +
model.outputs)
extensions = [
Timing(),
FinishAfter(after_n_epochs=NUM_EPOCHS),
DataStreamMonitoring(
bn_monitored_variables, train_monitor_stream, prefix="train",
updates=bn_updates, before_first_epoch=True,
every_n_epochs=monitor_every),
DataStreamMonitoring(
monitored_variables, valid_monitor_stream, prefix="valid",
before_first_epoch=False, every_n_epochs=monitor_every),
Checkpoint(save_path, every_n_epochs=checkpoint_every,
before_training=True, after_epoch=True, after_training=True,
use_cpickle=True),
SampleCheckpoint(interface=AliModel, z_dim=NLAT, image_size=IMAGE_SIZE, channels=NUM_CHANNELS, dataset="celeba_64", split="valid", save_subdir=subdir, before_training=True, after_epoch=True),
ProgressBar(),
Printing(),
]
main_loop = MainLoop(model=bn_model, data_stream=main_loop_stream,
algorithm=algorithm, extensions=extensions)
return main_loop
if __name__ == "__main__":
logging.basicConfig(level=logging.INFO)
parser = argparse.ArgumentParser(description="Train ALI on CelebA")
parser.add_argument('--model', dest='model', type=str,
default="ali_celeba.zip", help="Model to save")
parser.add_argument("--subdir", dest='subdir', type=str, default="output",
help="Subdirectory for output files (images)")
parser.add_argument('--dataset', dest='dataset', default=None,
help="Dataset for training.")
parser.add_argument("--image-size", dest='image_size', type=int, default=64,
help="size of (offset) images")
parser.add_argument('--color-convert', dest='color_convert',
default=False, action='store_true',
help="Convert source dataset to color from grayscale.")
parser.add_argument("--batch-size", type=int, dest="batch_size",
default=100, help="Size of each mini-batch")
parser.add_argument("--monitor-every", type=int, dest="monitor_every",
default=4, help="Frequency in epochs for monitoring")
parser.add_argument("--checkpoint-every", type=int,
dest="checkpoint_every", default=1,
help="Frequency in epochs for checkpointing")
args = parser.parse_args()
create_main_loop(args.model, args.subdir, args.dataset,
args.color_convert, args.batch_size, args.monitor_every,
args.checkpoint_every, args.image_size).run()