forked from ReidWilliams/GANs
-
Notifications
You must be signed in to change notification settings - Fork 0
/
model.py
259 lines (203 loc) · 10.6 KB
/
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
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
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
# Model brings together the network, the loss function, the feed of
# training images, and a training loop
import tensorflow as tf
from PIL import Image
import numpy as np
import os
from feed import Feed
from architecture import GAN
from utils import pixels01, pixels11, tile
# print and flush
def printnow(x, end='\n'): print(x, flush=True, end=end)
# safe create directories
def makedirs(d):
if not os.path.exists(d): os.makedirs(d)
# This model uses the same loss function as DCGAN
class Model:
def __init__(self, feed, batch_size=64, img_shape=(64, 64),
G_lr=0.0004, D_lr=0.0004, G_beta1=0.5, D_beta1=0.5,
zsize=128, save_freq=10, output_cols=4, output_rows=4,
sess=None, checkpoints_path=None):
self.batch_size = batch_size
if ((img_shape[0] % 32 != 0) or (img_shape[1] % 32 != 0)):
raise ValueException("Image dimensions need to be divisible by 32. \
Dimensions received was %s." % img_shape)
self.img_shape = img_shape + (3,) # add (r,g,b) channels dimension
# learning rates for Adam optimizer
self.G_lr = G_lr
self.D_lr = D_lr
self.G_beta1 = G_beta1
self.D_beta1 = D_beta1
# size of latent vector
self.zsize = zsize
# save session and examples after this many batches
self.save_freq = int(save_freq)
# cols and rows of output image tile
self.output_cols = output_cols
self.output_rows = output_rows
pwd = os.getcwd()
self.dirs = {
'output': os.path.join(pwd, 'output'),
'logs': os.path.join(pwd, 'logs'),
'checkpoints': os.path.join(pwd, 'checkpoints')
}
# set or create tensorflow session
self.sess = sess
if not self.sess:
self.sess = tf.InteractiveSession()
# create directories if they don't exist
makedirs(self.dirs['logs'])
makedirs(self.dirs['output'])
makedirs(self.dirs['checkpoints'])
self.checkpoints_path = checkpoints_path or os.path.join(self.dirs['checkpoints'], 'checkpoint.ckpt')
# get number of files in output so we can continue where a previous process
# left off without overwriting
self.output_img_idx = len([f for f in os.listdir(self.dirs['output']) \
if os.path.isfile(os.path.join(self.dirs['output'], f))])
# data feed for training
self.feed = feed
# bool used by batch normalization. BN behavior is different when training
# vs predicting
self.is_training = tf.placeholder(tf.bool)
self.arch = GAN(self.is_training, img_shape=self.img_shape, zsize=128)
# how many times to train discriminator per minibatch
# This is a hyperparameter that can be tuned, it's >1 in wgans
self.D_train_iters = 2
# Build the network
def build_model(self):
# real image inputs from training data feed for training the
# discriminator
self.X = tf.placeholder(tf.float32, (None,) + self.img_shape)
# for feeding random draws of z (latent variable) to the generator
self.Z = tf.placeholder(tf.float32, (None, self.zsize))
# Instantiate a generator network. It takes an input
# of a latent vector
self.Gz = self.arch.generator(self.Z)
# discriminator connected to real image input (X)
self.Dreal, self.Dreal_logits, self.Dreal_similarity = \
self.arch.discriminator(self.X)
# create a second instance of the discriminator and connect to the
# output of the generator. reuse=True means this second instance will
# share the same network weights and biases as the first instance. We want
# this because training the discriminator weights happens using a loss
# function that is a function of both the discriminator applied to a generated
# image and the discriminator applied to a real image.
self.Dz, self.Dz_logits, _ = \
self.arch.discriminator(self.Gz, reuse=True)
# Build the loss function.
def build_losses(self):
self.Dreal_loss = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(
labels=(tf.ones_like(self.Dreal_logits) - 0.25),
logits=self.Dreal_logits))
self.Dz_loss = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(
labels=tf.zeros_like(self.Dz_logits),
logits=self.Dz_logits))
# discriminator loss function from DCGAN
# discriminator wants to label real images with 1, generated with 0
self.D_loss = self.Dreal_loss + self.Dz_loss
# generator loss function. Make the generator think generated images
# are real
self.G_loss = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(
labels=tf.ones_like(self.Dz_logits),
logits=self.Dz_logits))
def build_optimizers(self):
# explicitly grab lists of variables for each type of network. This is used
# below to set up TF operations that train only one part of the network at
# a time (either generator or discriminator)
G_vars = [i for i in tf.trainable_variables() if 'generator' in i.name]
D_vars = [i for i in tf.trainable_variables() if 'discriminator' in i.name]
# Create optimizers.
G_opt = tf.train.AdamOptimizer(learning_rate=self.G_lr, beta1=self.G_beta1)
D_opt = tf.train.AdamOptimizer(learning_rate=self.D_lr, beta1=self.D_beta1)
# In tensor flow, you set up training by handing an optimizer object a tensor
# this is the output of a loss function, and (in this case) a set of variables
# that can be changed. You get back a training operation that you then run
# (see below) to take a step in training.
# pass var_list explicitly so that during training of (e.g.) generator, discriminator
# weights and biases aren't updated.
self.G_train = G_opt.minimize(self.G_loss, var_list=G_vars)
self.D_train = D_opt.minimize(self.D_loss, var_list=D_vars)
def setup_session(self):
# store epoch as tf variable so we can save it in the session
# this is nice for logging so that restarting the process doesn't reset the
# epoch count.
self.epoch = tf.get_variable('epoch', dtype='int32', initializer=tf.constant(0))
# random numbers to generate outputs. Store in tf variable so it gets
# stored in session. This is useful so that generated images that are saved
# during training come from the same latent variable inputs. This lets you
# see the gradual change / improvement of outputs even if the process dies
# and gets restarted
self.example_noise = tf.get_variable('noise', dtype='float32',
initializer=tf.constant(np.random.normal(size=(self.batch_size, self.zsize)).astype('float32')))
self.saver = tf.train.Saver()
try:
print('trying to restore session from %s' % self.checkpoints_path)
self.saver.restore(self.sess, self.checkpoints_path)
print('restored session')
except Exception as e:
print('failed to restore session, creating a new one')
print(e)
tf.global_variables_initializer().run()
# log some basic data for tensorboard
def setup_logging(self):
self.writer = tf.summary.FileWriter(self.dirs['logs'], self.sess.graph)
self.G_stats = tf.summary.merge([
tf.summary.scalar('G_loss', self.G_loss)
])
Dreal_mean = tf.reduce_mean(tf.sigmoid(self.Dreal_logits))
Dz_mean = tf.reduce_mean(tf.sigmoid(self.Dz_logits))
self.D_stats = tf.summary.merge([
tf.summary.scalar('Dreal_out', Dreal_mean),
tf.summary.scalar('Dz_out', Dz_mean),
tf.summary.scalar('D_loss', self.D_loss)
])
def train(self):
batches = self.feed.nbatches()
printnow('training with %s batches per epoch' % batches)
printnow('saving session and examples every %s batches' % self.save_freq)
# order the logged data for tensorboard
logcounter = 0
epoch = self.epoch.eval() # have to do this b/c self.epoch is a tensorflow var
while True:
for batch in range(batches):
# training image pixel values are [0,1] but DCGAN and it seems most
# GAN architectures benefit from / use [-1,1]
xfeed = pixels11(self.feed.feed(batch)) # convert to [-1, 1]
zfeed = np.random.normal(size=(self.batch_size, self.zsize)).astype('float32')
# train discriminator (possibly more than once) by running
# the training operation inside the session
for i in range(self.D_train_iters):
_, summary = self.sess.run(
[ self.D_train, self.D_stats ],
feed_dict={ self.X: xfeed, self.Z: zfeed, self.is_training: True })
self.writer.add_summary(summary, logcounter)
# train generator
_, summary = self.sess.run(
[ self.G_train, self.G_stats],
feed_dict={ self.X: xfeed, self.Z: zfeed, self.is_training: True })
self.writer.add_summary(summary, logcounter)
logcounter += 1
if (batch % self.save_freq == 0):
printnow('Epoch %s, batch %s/%s, saving session and examples' % (epoch, batch, batches))
# update TF epoch variable so restart of process picks up at same
# epoch where it died
self.sess.run(self.epoch.assign(epoch))
self.save_session()
self.output_examples()
epoch += 1
def save_session(self):
self.saver.save(self.sess, self.checkpoints_path)
def output_examples(self):
cols = self.output_cols
rows = self.output_rows
nimgs = cols*rows
zfeed = self.example_noise.eval() # need to eval to get value since it's a tf variable
imgs = self.sess.run(self.Gz, feed_dict={ self.Z: zfeed, self.is_training: False })
imgs = imgs[:nimgs]
# conver [-1,1] back to [0,1] before saving
imgs = pixels01(imgs)
path = os.path.join(self.dirs['output'], '%06d.jpg' % self.output_img_idx)
tiled = tile(imgs, (rows, cols))
as_ints = (tiled * 255.0).astype('uint8')
Image.fromarray(as_ints).save(path)
self.output_img_idx += 1