-
Notifications
You must be signed in to change notification settings - Fork 3
/
train_gan.py
311 lines (237 loc) · 12.2 KB
/
train_gan.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
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
from models import *
from dataloader import *
import conf
import tensorflow as tf
def main():
"""The famous main function that no one knows what it's for"""
run_title = 'DEBUG_RUN' if conf.SIMPLE_RUN else input('Please name this session:')
# Training parameters
epochs = 600_000
lr = 0.00002
lr_decay = 0.5
decay_every = 100_000
save_every = 10_000
beta1 = 0.5
force_gpu = conf.FORCE_GPU
num_gpu = conf.NUM_GPU
# Encoded texts fed from the pipeline
if conf.END_TO_END:
datasource = GanDataLoader_NoEncoder()
else:
datasource = GanDataLoader()
text_right, real_image = datasource.correct_pipe()
text_wrong, real_image2 = datasource.incorrect_pipe()
text_G, real_image_G = datasource.text_only_pipe()
text_encoder(text_G, reuse=False)
# This is to be able to change the learning rate while training
with tf.variable_scope('learning_rate'):
lr_v = tf.Variable(lr, trainable=False)
# Optimizers
optimizer = tf.train.AdamOptimizer(learning_rate=lr_v, beta1=beta1)
def split_tensor_for_gpu(t):
return tf.split(t,num_gpu)
# TODO Double GPU Begin =========
# Split for GPU
c1_txts = split_tensor_for_gpu(text_right)
c1_imgs = split_tensor_for_gpu(real_image)
c2_txts = split_tensor_for_gpu(text_wrong)
c2_imgs = split_tensor_for_gpu(real_image2)
cg_txts = split_tensor_for_gpu(text_G)
G_grads = []
D_grads = []
text_grads = []
G_loss = 0
D_loss = 0
for i in range(num_gpu):
# Runs on GPU
G_grads_vars, D_grads_vars, G_loss_gpu, D_loss_gpu, txt_grads_gpu = loss_tower(i, optimizer, cg_txts[i], c1_imgs[i], c1_txts[i], c2_imgs[i], c2_txts[i])
# normalize and element wise add
if not G_grads:
G_grads = [g_grad / num_gpu for g_grad, g_vars in G_grads_vars]
D_grads = [d_grad / num_gpu for d_grad, d_vars in D_grads_vars]
text_grads = [txt_grad / num_gpu for txt_grad, txt_vars in txt_grads_gpu]
else:
# Element wise add to G_grads collection, G_grads is same size as G_grads_vars' grads
G_grads = [ g_grad / num_gpu + G_grads[j] for j, (g_grad, g_vars) in enumerate(G_grads_vars)]
D_grads = [ d_grad / num_gpu + D_grads[j]for j, (d_grad, d_vars) in enumerate(D_grads_vars)]
text_grads = [txt_grad / num_gpu + text_grads[j] for j, (txt_grad, txt_vars) in enumerate(txt_grads_gpu)]
G_loss = G_loss_gpu / num_gpu + G_loss
D_loss = D_loss_gpu / num_gpu + D_loss
# sum and normalize
## extract vars
#G_vars = [var for grad, var in G_grads_vars] # G0 and G1 share vars, so doesn't matter
#D_vars = [var for grad, var in D_grads_vars] # D0 and D1 share vars, so doesn't matter
G_vars = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES, scope='generator')
D_vars = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES, scope='discriminator')
if conf.END_TO_END:
# Encoder_grads_G = G_grads[22:] # TODO Hard coded split between GAN grads and encoder grads
# G_grads = G_grads[:22]
# Encoder_grads_D = D_grads[24:]
# D_grads = D_grads[:24]
Encode_vars = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES, scope='txt_encode')
# Encoder_grads = [(g + d)/2 for g, d in zip(Encoder_grads_G, Encoder_grads_D)]
#E_opt = optimizer.apply_gradients(zip(Encoder_grads, Encode_vars))
E_opt = optimizer.apply_gradients(zip(text_grads, Encode_vars))
else:
Encode_vars = []
E_opt = tf.constant(5)
#G_vars += Encode_vars
D_vars += Encode_vars
G_opt = optimizer.apply_gradients(zip(G_grads, G_vars))
D_opt = optimizer.apply_gradients(zip(D_grads, D_vars))
# Single GPU # TODO SINGLE GPU BEGIN ============
# G_grads_vars, D_grads_vars, G_loss, D_loss = loss_tower(0, optimizer, text_G, real_image, text_right, real_image2,
# text_wrong)
#G_opt = optimizer.apply_gradients(G_grads_vars)
#D_opt = optimizer.apply_gradients(D_grads_vars)
# TODO OLD SETUP BEGIN ========
# Outputs from G and D
# fake_image = generator_resnet(text_G)
# S_r = discriminator_resnet(real_image, text_right)
# S_w = discriminator_resnet(real_image2, text_wrong)
# S_f = discriminator_resnet(fake_image, text_G)
#
#
# # Loss functions for G and D
# G_loss = -tf.reduce_mean(tf.log(S_f))
# D_loss = -tf.reduce_mean(tf.log(S_r) + (tf.log(1 - S_w) + tf.log(1 - S_f))/2)
# tf.summary.scalar('generator_loss', G_loss)
# tf.summary.scalar('discriminator_loss', D_loss)
# Parameters we want to train, and their gradients
# G_vars = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES, scope='generator')
# D_vars = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES, scope='discriminator')
# G_grads = tf.gradients(G_loss, G_vars)
# D_grads = tf.gradients(D_loss, D_vars)
# G_opt = optimizer.apply_gradients(zip(G_grads, G_vars))
# D_opt = optimizer.apply_gradients(zip(D_grads, D_vars))
# Metrics:
testset_op = setup_testset(datasource)
# Write to tensorboard
setup_accuracy(text_right, real_image, text_wrong, real_image2, text_G)
tf.summary.scalar('generator_loss', G_loss, family='GAN')
tf.summary.scalar('discriminator_loss', D_loss, family='GAN')
hp_str = 'Force_gpu:{}\ndecay_every:{}\ndecay_rate:{}\blearning_rate:{}\nepochs:{}\nforce_gpu:{}'.format(force_gpu,decay_every,decay_every,lr,epochs,force_gpu)
outer_string = tf.convert_to_tensor(hp_str)
tf.summary.text('configuration', outer_string)
# plot weights
# for var in tf.trainable_variables():
# tf.summary.histogram(var.name, var, family='GAN_internal')
# for grad, var in zip(G_grads, G_vars):
# tf.summary.histogram(var.name + '/gradient', grad, family='internal')
# for grad, var in zip(D_grads, D_vars):
# tf.summary.histogram(var.name + '/gradient', grad, family='internal')
merged = tf.summary.merge_all()
saver = tf.train.Saver()
#fake_img_summary_op = tf.summary.image('generated_image', tf.concat([fake_image * 127.5, real_image_G * 127.5], axis=2))
run_name = run_title + datetime.datetime.now().strftime("May_%d_%I_%M%p_GAN")
writer = tf.summary.FileWriter('./tensorboard_logs/%s' % run_name, tf.get_default_graph())
with tf.Session(config=tf.ConfigProto(allow_soft_placement=(not force_gpu))) as sess: # Allow fall back to CPU
#tf.set_random_seed(100)
sess.run(tf.global_variables_initializer())
# Restore text encoder
text_encoder_vars = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES, scope='txt_encode')
# saver = tf.train.Saver(text_encoder_vars)
# saver.restore(sess, 'assets/char-rnn-cnn-19999')
datasource.preprocess_data_and_initialize(sess)
# Run the initializers for the pipeline
t0 = time()
for step in range(epochs):
# Updating the learning rate every 100 epochs (starting after first 1000 update steps)
if step != 0 and step > 10000 and (step % decay_every == 0):
sess.run(tf.assign(lr_v, lr_v * lr_decay))
log = " ** new learning rate: %f" % (lr * lr_decay)
print(log)
# Updates parameters in G and D, only every third time for D
if step % 10 == 0:
print('Update: ', step)
summary, dloss, gloss, _, _, _= sess.run(
[merged, D_loss, G_loss, D_opt, G_opt, E_opt])
print('Discriminator loss: ', dloss)
print('Generator loss: ', gloss)
print('time:', time() - t0 )
# Tensorboard stuff
writer.add_summary(summary, step)
# if step % 2 == 0:
# _, _ = sess.run(
# [D_opt, G_opt])
else:
_, _,_ = sess.run(
[D_opt, G_opt, E_opt])
if step % save_every == 0:
saver.save(sess, 'saved/%s' % run_name, global_step=step)
pass
if step % 100 == 0:
testset_op(sess, writer, step)
if step % 1000 == 0:
print('1000 epoch time:', time()-t0)
t0 = time()
# Close writer when done training
writer.close()
def loss_tower(gpu_num, optimizer, text_G, real_image, text_right, real_image2, text_wrong, reuse=True):
# Outputs from G and D
with tf.device('/gpu:%d' % gpu_num):
with tf.name_scope('gpu_%d' % gpu_num):
if conf.END_TO_END:
text_right = text_encoder(text_right, reuse)
text_G = text_encoder(text_G, reuse)
text_wrong = text_encoder(text_wrong, reuse)
else:
Encode_vars = []
fake_image = generator_resnet(text_G)
S_r = discriminator_resnet(real_image, text_right)
S_w = discriminator_resnet(real_image2, text_wrong)
S_f = discriminator_resnet(fake_image, text_G)
# Loss functions for G and D
G_loss = -tf.reduce_mean(tf.log(S_f), name='G_loss_gpu%d' % gpu_num)
D_loss = -tf.reduce_mean(tf.log(S_r) + (tf.log(1 - S_w) + tf.log(1 - S_f))/2, name='G_loss_gpu%d' % gpu_num)
if conf.END_TO_END:
Encode_vars = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES, scope='txt_encode')
Txt_loss = -tf.reduce_mean(tf.log(S_r))
txt_grads = optimizer.compute_gradients(Txt_loss, Encode_vars)
G_vars = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES, scope='generator')
D_vars = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES, scope='discriminator')
# Parameters we want to train, and their gradients
G_grads = optimizer.compute_gradients(G_loss, G_vars)
D_grads = optimizer.compute_gradients(D_loss, D_vars + Encode_vars) # disable text encoder training on D
return G_grads, D_grads, G_loss, D_loss, txt_grads
def setup_accuracy( c1_txt, c1_img, c2_txt, c2_img, cg_txt, reuse=True):
with tf.device('/gpu:0'):
if conf.END_TO_END:
c1_txt = text_encoder(c1_txt, reuse)
c2_txt = text_encoder(c2_txt, reuse)
cg_txt = text_encoder(cg_txt, reuse)
txt_in = tf.concat([c1_txt, c2_txt, cg_txt], axis=0)
g_img = generator_resnet(cg_txt, z_size=conf.GAN_BATCH_SIZE)
img_in = tf.concat([c1_img, c2_img, g_img], axis=0)
dout = discriminator_resnet(img_in, txt_in)
dout = tf.reshape(dout, [-1])
ones = tf.ones_like(dout)
zeros = tf.zeros_like(dout)
dout_stepped = tf.where(tf.greater(dout, 0.5),ones,zeros)
labels = tf.concat([tf.ones([conf.GAN_BATCH_SIZE]), tf.zeros([conf.GAN_BATCH_SIZE]), tf.zeros([conf.GAN_BATCH_SIZE])], axis=0)
diff = dout_stepped - labels
c1, c2, cg = tf.split(diff, 3)
c1_accuracy = (1 - tf.count_nonzero(c1) / conf.GAN_BATCH_SIZE) * 100
c2_accuracy = (1 - tf.count_nonzero(c2) / conf.GAN_BATCH_SIZE) * 100
cg_accuracy = (1 - tf.count_nonzero(cg) / conf.GAN_BATCH_SIZE) * 100
tf.summary.scalar('real_pair_accuracy', c1_accuracy, family='DiscriminatorAccuracy')
tf.summary.scalar('wrong_pair_accuracy', c2_accuracy, family='DiscriminatorAccuracy')
tf.summary.scalar('fake_pair_accuracy', cg_accuracy, family='DiscriminatorAccuracy')
def setup_testset(datasource):
# Test pipe setup
# Non-derterministic
sample_size = 10
test_nondeter_txt, test_nondeter_img = datasource.test_pipe(deterministic=False, sample_size = sample_size)
if conf.END_TO_END:
test_nondeter_txt = text_encoder(test_nondeter_txt, reuse=True)
test_batch_G_img = generator_resnet(test_nondeter_txt, z_size=sample_size)
img = tf.concat([test_batch_G_img * 127.5, test_nondeter_img * 127.5], axis=2)
test_batch_summary_op = tf.summary.image('test_batch', img, family='test_images', max_outputs=10)
# To be run inside a session
def testset_op(sess, writer, step):
test_batch_summary = sess.run(test_batch_summary_op)
writer.add_summary(test_batch_summary, step)
return testset_op
# So that we only run things if it's run directly, not if it's imported
if __name__ == '__main__':
main()