-
Notifications
You must be signed in to change notification settings - Fork 0
/
train_arch9.py
422 lines (343 loc) · 19.8 KB
/
train_arch9.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
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import argparse
import datetime
from functools import partial
import json
import traceback
import imlib as im
import numpy as np
import pylib
import tensorflow as tf
import tflib as tl
import data_noise_in as data
import models_712 as models
import os
def matching_loss(x1, x2, y1, y2, margin):
return tf.reduce_mean(tf.nn.relu(models.inner_product(x1, x2) - models.inner_product(y1, y2) + margin))
parser = argparse.ArgumentParser()
# settings
dataroot_default = './data/CelebA'
parser.add_argument('--dataroot', type=str, default=dataroot_default)
parser.add_argument('--dataset', type=str, default='celeba')
parser.add_argument('--gpu', type=str, default='0,1',
help='Specify which gpu to use by `CUDA_VISIBLE_DEVICES=num python train.py **kwargs`\
or `python train.py --gpu num` if you\'re running on a multi-gpu enviroment.\
You need to do nothing if your\'re running on a single-gpu environment or\
the gpu is assigned by a resource manager program.')
parser.add_argument('--threads', type=int, default=-1,
help='Control parallel computation threads,\
please leave it as is if no heavy cpu burden is observed.')
# model
# att_default = ['Bald', 'Bangs', 'Black_Hair', 'Blond_Hair', 'Brown_Hair', 'Bushy_Eyebrows', 'Eyeglasses',
# 'Male', 'Mouth_Slightly_Open', 'Mustache', 'No_Beard', 'Pale_Skin', 'Young']
att_default = ['Young', 'Mouth_Slightly_Open', 'Smiling', 'Black_Hair', 'Blond_Hair', 'Brown_Hair', 'Gray_Hair',
'Receding_Hairline', 'Bangs', 'Male', 'No_Beard', 'Mustache', 'Goatee','Sideburns']
parser.add_argument('--atts', default=att_default, choices=data.Celeba.att_dict.keys(), nargs='+',
help='Attributes to modify by the model')
parser.add_argument('--img_size', type=int, default=128, help='input image size')
# generator
parser.add_argument('--n_downblks_gen', type=int, default=3)
parser.add_argument('--n_intermedblks_gen', type=int, default=2)
parser.add_argument('--n_upblks_gen', type=int, default=3)
parser.add_argument('--ch_gen', type=int, default=32)
# mappingnet
parser.add_argument('--n_layers_map', type=int, default=4)
parser.add_argument('--n_mlp_map', type=int, default=3)
parser.add_argument('--fc_dim_map', type=int, default=64)
parser.add_argument('--dim_noise', type=int, default=16)
# encoder
parser.add_argument('--n_downblks_enc', type=int, default=1)
parser.add_argument('--n_intermedblks_enc', type=int, default=3)
parser.add_argument('--n_mlp_enc', type=int, default=3)
parser.add_argument('--ch_enc', type=int, default=16)
# discriminator
parser.add_argument('--n_resblks_dis', type=int, default=4)
parser.add_argument('--ch_dis', type=int, default=16)
parser.add_argument('--rec_loss_weight', type=float, default=100.0)
# training
parser.add_argument('--mode', default='wgan', choices=['wgan', 'lsgan', 'dcgan'])
parser.add_argument('--epoch', type=int, default=100, help='# of epochs') #200
parser.add_argument('--init_epoch', type=int, default=100, help='# of epochs with init lr.') # 100
parser.add_argument('--batch_size', type=int, default=60)
parser.add_argument('--lr', type=float, default=0.0002, help='learning rate')
parser.add_argument('--n_d', type=int, default=5, help='# of d updates per g update')
parser.add_argument('--n_sample', type=int, default=8, help='# of sample images')
parser.add_argument('--save_freq', type=int, default=0,
help='save model evary save_freq iters, 0 means to save evary epoch.')
parser.add_argument('--sample_freq', type=int, default=0,
help='eval on validation set every sample_freq iters, 0 means to save every epoch.')
# others
parser.add_argument('--use_cropped_img', action='store_true')
parser.add_argument('--experiment_name', default=datetime.datetime.now().strftime("%Y.%m.%d-%H%M%S"))
parser.add_argument('--num_ckpt', type=int, default=10)
parser.add_argument('--clear', default=False, action='store_true')
args = parser.parse_args()
num_gpu = 1
n_att = len(args.atts)
pylib.mkdir('./output/%s' % args.experiment_name)
with open('./output/%s/setting.txt' % args.experiment_name, 'w') as f:
f.write(json.dumps(vars(args), indent=4, separators=(',', ':')))
# ==============================================================================
# = graphs =
# ==============================================================================
# data
if args.threads >= 0:
cpu_config = tf.ConfigProto(intra_op_parallelism_threads=args.threads // 2,
inter_op_parallelism_threads=args.threads // 2,
device_count={'CPU': args.threads})
sess = tf.Session(config=cpu_config)
else:
sess = tl.session()
crop_ = not args.use_cropped_img
if args.dataset == 'celeba':
tr_data = data.Celeba(args.dataroot, att_default, args.img_size, args.batch_size, part='train', sess=sess, crop=crop_, is_tfrecord = True)
val_data = data.Celeba(args.dataroot, args.atts, args.img_size, args.n_sample, part='val', shuffle=False, sess=sess, crop=crop_)
else:
tr_data = data.x2y(args.dataset, args.dataroot, args.img_size, args.batch_size, part='train', sess=sess, is_tfrecord = True)
val_data = data.x2y(args.dataset, args.dataroot, args.img_size, args.n_sample, part='val', shuffle=False, sess=sess, is_tfrecord = False)
# models
Generator = partial(models.Generator2_wRelu, n_downblks = args.n_downblks_gen, n_intermedblks = args.n_intermedblks_gen, n_upblks = args.n_upblks_gen, ch = args.ch_gen)
MappingNet = partial(models.MappingNet_multiStream_deconv_wIN_wRelu_concat, n_mlp = args.n_mlp_map, n_layers = args.n_layers_map, fc_dim = args.fc_dim_map, ch = args.ch_enc*2)
Encoder = partial(models.Encoder4_wIN_wRelu_concat,n_downblks = args.n_downblks_enc, n_intermedblks = args.n_intermedblks_enc, n_mlp = args.n_mlp_enc, ch = args.ch_enc)
# Discriminator = partial(models.PatchDiscriminator, n_att=n_att, n_resblks = args.n_resblks_dis, ch = args.ch_dis)
# Discriminator = partial(models.Discriminator_multiTask, n_att=n_att, n_resblks = args.n_resblks_dis, ch = args.ch_dis)
Discriminator = partial(models.Discriminator, n_att=n_att, n_resblks = args.n_resblks_dis, ch = args.ch_dis)
# inputs
r_M = tf.get_variable(name='G_R', shape=[182000, args.dim_noise], dtype=tf.float32, initializer=tf.random_normal_initializer(mean=0.0, stddev=1.0))
lr = tf.placeholder(dtype=tf.float32, shape=[])
xs_s = tr_data.batch_op[0]
ls_s = tf.to_float(tr_data.batch_op[1])
rs_s = tr_data.batch_op[2]
if args.dataset == 'celeba':
# b_s = tf.random_shuffle(a_s)
permuted_index = tf.random_shuffle(tf.range(args.batch_size))
xt_s = tf.gather(xs_s, permuted_index)
lt_s = tf.gather(ls_s, permuted_index)
rt_i = tf.gather(rs_s, permuted_index)
rt_s = tf.reshape(tf.gather(r_M, rt_i), [rt_i.shape[0], args.dim_noise])
else:
pass
d_opt = tf.train.AdamOptimizer(lr, beta1=0.5)
g_opt = tf.train.AdamOptimizer(lr, beta1=0.5)
tower_d_grads = []
tower_g_grads = []
tower_d_loss_gan = []
tower_gp = []
tower_d_loss_cls = []
tower_g_loss_sim = []
tower_g_loss_r = []
tower_g_loss_r0 = []
tower_g_loss_gan = []
tower_g_loss_cls = []
tower_g_loss_cyc = []
tower_g_loss_reg = []
tower_g_loss_rec = []
tower_g_loss_interp = []
xs_sample = tf.placeholder(tf.float32, shape=[None, args.img_size, args.img_size, 3])
lt_sample = tf.placeholder(tf.float32, shape=[None, n_att])
ls_sample = tf.placeholder(tf.float32, shape=[None, n_att])
with tf.variable_scope(tf.get_variable_scope()):
for i in range(1):
with tf.name_scope("tower_%d" % i):
xs = xs_s[i * args.batch_size // num_gpu: (i + 1) * args.batch_size // num_gpu]
xt = xt_s[i * args.batch_size // num_gpu: (i + 1) * args.batch_size // num_gpu]
rt = rt_s[i * args.batch_size // num_gpu: (i + 1) * args.batch_size // num_gpu]
ls = ls_s[i * args.batch_size // num_gpu: (i + 1) * args.batch_size // num_gpu]
lt = lt_s[i * args.batch_size // num_gpu: (i + 1) * args.batch_size // num_gpu]
rs = rs_s[i * args.batch_size // num_gpu: (i + 1) * args.batch_size // num_gpu]
# generate
r_1 = tf.random_normal((args.batch_size // num_gpu, args.dim_noise))
s_rand_t1 = MappingNet(r_1, lt - ls)
r_0 = rt
s_rand_t = MappingNet(r_0, lt - ls)
s_rand_s, _ = Encoder(xs, ls - lt)
s_rand1 = s_rand_s + s_rand_t1
xg_rand1 = Generator(xs, s_targ=s_rand1)
s_targ_t, _ = Encoder(xt, lt - ls)
loss_latent = tf.losses.mean_squared_error(s_targ_t, s_rand_t)
r_grads = tf.gradients(loss_latent, r_0)[0]
grads_l2 = tf.reshape(tf.sqrt(tf.reduce_mean(tf.square(r_grads), 1)), [args.batch_size, 1])
normalized_grads = (0.9 / grads_l2 + 5.0) * r_grads
r = tf.stop_gradient(tf.clip_by_value((r_0 + normalized_grads), -1.0, 1.0)) #
tower_assign = []
for index_i in range(args.batch_size):
tower_assign.append(tf.assign(r_M[tf.reshape(rt_i[index_i], [])], r[index_i]))
s_rand_t = MappingNet(r, lt - ls)
s_rand = s_rand_s + s_rand_t
xg_rand = Generator(xs, s_targ=s_rand)
s_targ = s_targ_t + s_rand_s
xg_targ = Generator(xs, s_targ=s_targ)
alpha = tf.random_uniform((args.batch_size // num_gpu, 1, 1, 1), maxval=1)
s_interp = alpha * s_rand + (1 - alpha) * s_targ
xg_interp = Generator(xs, s_targ=s_interp)
s_targ_t_rand, _ = Encoder(tf.stop_gradient(xg_rand), lt - ls)
s_targ_rand = s_targ_t_rand + s_rand_s
s_targ_t_targ, _ = Encoder(tf.stop_gradient(xg_targ), lt - ls)
s_targ_targ = s_targ_t_targ + s_rand_s
s_rand_rec_t = MappingNet(r, ls - ls)
s_rand_rec_s,_ = Encoder(xs, ls - ls)
s_rand_rec = s_rand_rec_t + s_rand_rec_s
xg_rand_rec = Generator(xs, s_targ=s_rand_rec)
s_targ_t_rec, _ = Encoder(xt, ls - ls)
s_targ_rec = s_targ_t_rec + s_rand_rec_s
xg_targ_rec = Generator(xs, s_targ=s_targ_rec)
l_1 = tf.ones_like(lt)
s_targ_t_targ_oth, _ = Encoder(xg_rand, (l_1-tf.abs(lt - ls))*(l_1-2.0*ls))
s_rand_t_oth, _ = Encoder(xs, (l_1-tf.abs(lt - ls))*(l_1-2.0*ls))
# discriminate
xs_logit_gan, xs_logit_att,_= Discriminator(xs)
xgr_logit_gan, xgr_logit_att,xgr_logit_att_feature = Discriminator(xg_rand)
xgt_logit_gan, xgt_logit_att,xgt_logit_att_feature= Discriminator(xg_targ)
xgi_logit_gan, xgi_logit_att, xgi_logit_att_feature = Discriminator(xg_interp)
# discriminator losses
if args.mode == 'wgan': # wgan-gp
wd = tf.reduce_mean(xs_logit_gan) - (tf.reduce_mean(xgr_logit_gan) + tf.reduce_mean(xgt_logit_gan) + tf.reduce_mean(xgi_logit_gan)) / 3.0
d_loss_gan = -wd
gp = (models.gradient_penalty(Discriminator, xs, xg_rand) + models.gradient_penalty(Discriminator,xs, xg_targ) + \
models.gradient_penalty(Discriminator, xs, xg_interp)) / 3.0
d_loss_cls = tf.losses.sigmoid_cross_entropy(ls, xs_logit_att)
tower_d_loss_gan.append(d_loss_gan)
tower_gp.append(gp)
tower_d_loss_cls.append(d_loss_cls)
d_loss = d_loss_gan + gp * 10.0 + d_loss_cls
# generator losses
if args.mode == 'wgan':
g_loss_gan = -(tf.reduce_mean(xgr_logit_gan) + tf.reduce_mean(xgt_logit_gan) + tf.reduce_mean(xgi_logit_gan)) / 3.0
g_loss_cls = (tf.losses.sigmoid_cross_entropy(lt, xgr_logit_att) +
tf.losses.sigmoid_cross_entropy(lt, xgt_logit_att) +
tf.losses.sigmoid_cross_entropy(lt, xgi_logit_att)) / 3.0
g_loss_rec = (tf.losses.absolute_difference(xs, xg_rand_rec) + tf.losses.absolute_difference(xs, xg_targ_rec)) / 2.0
g_loss_cyc = (tf.losses.absolute_difference(s_rand, s_targ_rand) + tf.losses.absolute_difference(s_targ, s_targ_targ) + tf.losses.absolute_difference(s_targ_t_targ_oth, s_rand_t_oth)) / 3.0
g_loss_sim = tf.losses.absolute_difference(s_rand, s_targ)
g_loss_ModeSeeking = tf.losses.absolute_difference(r_1, r)/(tf.losses.absolute_difference(tf.stop_gradient(xg_rand1), xg_rand) + 0.00001)
tower_g_loss_gan.append(g_loss_gan)
tower_g_loss_cls.append(g_loss_ModeSeeking)
tower_g_loss_sim.append(g_loss_sim)
tower_g_loss_r.append(tf.reduce_mean(r))
tower_g_loss_r0.append(tf.reduce_mean(r_0))
tower_g_loss_rec.append(g_loss_rec)
tower_g_loss_cyc.append(tf.reduce_mean(g_loss_cyc))
g_loss = g_loss_gan + g_loss_cls * 10.0 + g_loss_rec * args.rec_loss_weight + g_loss_cyc * args.rec_loss_weight /2.0 + g_loss_sim * args.rec_loss_weight /2.0 + g_loss_ModeSeeking * 0.005#+ g_loss_interp * 10
# optimize
tf.get_variable_scope().reuse_variables()
d_var = tl.trainable_variables('D')
d_grads = d_opt.compute_gradients(d_loss, var_list=d_var)
tower_d_grads.append(d_grads)
g_var = tl.trainable_variables('G')
g_grads = g_opt.compute_gradients(g_loss, var_list=g_var)
tower_g_grads.append(g_grads)
# optimize
d_grads = models.average_gradients(tower_d_grads)
d_step = d_opt.apply_gradients(d_grads)
g_grads = models.average_gradients(tower_g_grads)
g_step = g_opt.apply_gradients(g_grads)
# summary
d_summary = tl.summary({
tf.reduce_mean(tf.stack(tower_d_loss_gan)): 'd_loss_gan',
tf.reduce_mean(tf.stack(tower_gp)): 'gp',
tf.reduce_mean(tf.stack(tower_d_loss_cls)): 'd_loss_cls',
}, scope='D')
g_summary = tl.summary({
tf.reduce_mean(tf.stack(tower_g_loss_gan)): 'g_loss_gan',
tf.reduce_mean(tf.stack(tower_g_loss_cls)): 'g_loss_cls',
tf.reduce_mean(tf.stack(tower_g_loss_rec)): 'g_loss_rec',
tf.reduce_mean(tf.stack(tower_g_loss_cyc)): 'g_loss_cyc',
tf.reduce_mean(tf.stack(tower_g_loss_sim)): 'g_loss_sim',
tf.reduce_mean(tf.stack(tower_g_loss_r)): 'g_loss_r',
tf.reduce_mean(tf.stack(tower_g_loss_r0)): 'g_loss_r0'
}, scope='G')
lr_summary = tl.summary({lr: 'lr'}, scope='Learning_Rate')
d_summary = tf.summary.merge([d_summary, lr_summary])
# validation
test_label = lt_sample - ls_sample
r_sample = tf.random_normal((args.n_sample, args.dim_noise))
s_rand_sample_t = MappingNet(r_sample, test_label)
s_rand_sample_s, _ = Encoder(xs_sample, -test_label)
s_rand_sample = s_rand_sample_s + s_rand_sample_t
xg_rand_sample = Generator(xs_sample, s_targ = s_rand_sample)
# ==============================================================================
# = train =
# ==============================================================================
# iteration counter
it_cnt, update_cnt = tl.counter()
# saver
saver = tf.train.Saver(max_to_keep=args.num_ckpt)
# summary writer
summary_writer = tf.summary.FileWriter('./output/%s/summaries' % args.experiment_name, sess.graph)
# initialization
ckpt_dir = './output/%s/checkpoints' % args.experiment_name
pylib.mkdir(ckpt_dir)
try:
assert args.clear == False
tl.load_checkpoint(ckpt_dir, sess)
except:
print('NOTE: Initializing all parameters...')
sess.run(tf.global_variables_initializer())
# train
try:
xa_sample_ipt, a_sample_ipt = val_data.get_next()
b_sample_ipt_list = [a_sample_ipt] # the first is for reconstruction
for i in range(len(args.atts)):
tmp = np.array(a_sample_ipt, copy=True)
tmp[:, i] = 1 - tmp[:, i] # inverse attribute
if args.dataset == 'celeba':
tmp = data.Celeba.check_attribute_conflict(tmp, args.atts[i], args.atts)
b_sample_ipt_list.append(tmp)
it_per_epoch = len(tr_data) // (args.batch_size * (args.n_d + 1))
max_it = args.epoch * it_per_epoch
for it in range(sess.run(it_cnt), max_it):
with pylib.Timer(is_output=False) as t:
sess.run(update_cnt)
# which epoch
epoch = it // it_per_epoch
it_in_epoch = it % it_per_epoch + 1
# learning rate
lr_ipt = args.lr / (10 ** (epoch // args.init_epoch))
# train D
for i in range(args.n_d):
d_summary_opt, _ = sess.run([d_summary, d_step], feed_dict={lr: lr_ipt})
summary_writer.add_summary(d_summary_opt, it)
# train G
g_summary_opt, _, _ = sess.run([g_summary, g_step, tower_assign], feed_dict={lr: lr_ipt})
summary_writer.add_summary(g_summary_opt, it)
# display
if (it + 1) % 1 == 0:
print("Epoch: (%3d) (%5d/%5d) Time: %s!" % (epoch, it_in_epoch, it_per_epoch, t))
# save
if (it + 1) % (args.save_freq if args.save_freq else it_per_epoch*5) == 0:
save_path = saver.save(sess, '%s/Epoch_(%d)_(%dof%d).ckpt' % (ckpt_dir, epoch, it_in_epoch, it_per_epoch))
print('Model is saved at %s!' % save_path)
# sample
if (it + 1) % (args.save_freq if args.save_freq else it_per_epoch) == 0:
x_sample_opt_list = [xa_sample_ipt, np.full((args.n_sample, args.img_size, args.img_size // 10, 3), -1.0)]
_a_sample_ipt = (a_sample_ipt * 2 - 1) * 0.5
for i, b_sample_ipt in enumerate(b_sample_ipt_list):
_b_sample_ipt = (b_sample_ipt * 2 - 1) * 0.5
x_sample_opt_list.append(sess.run(xg_rand_sample, feed_dict={xs_sample: xa_sample_ipt,
lt_sample: _b_sample_ipt,
ls_sample: _a_sample_ipt}))
last_images = x_sample_opt_list[-1]
if i > 0: # add a mark (+/-) in the upper-left corner to identify add/remove an attribute
for nnn in range(last_images.shape[0]):
last_images[nnn, 2:5, 0:7, :] = 1.
if _b_sample_ipt[nnn, i - 1] > 0:
last_images[nnn, 0:7, 2:5, :] = 1.
last_images[nnn, 1:6, 3:4, :] = -1.
last_images[nnn, 3:4, 1:6, :] = -1.
sample = np.concatenate(x_sample_opt_list, 2)
save_dir = './output/%s/sample_training' % args.experiment_name
pylib.mkdir(save_dir)
im.imwrite(im.immerge(sample, args.n_sample, 1), '%s/Epoch_(%d)_(%dof%d).jpg' % \
(save_dir, epoch, it_in_epoch, it_per_epoch))
# xa_ipt, xc_ipt, xac_ipt, xca_ipt = sess.run([xa, xc, xac, xca])
# xa_ipt, xc_ipt, xac_ipt, xca_ipt = xa_ipt[:args.n_sample], xc_ipt[:args.n_sample], xac_ipt[:args.n_sample], xca_ipt[:args.n_sample]
# im.imwrite(im.immerge(np.concatenate([xa_ipt, xc_ipt, xac_ipt, xca_ipt], 2), args.n_sample, 1),
# '%s/Epoch_(%d)_(%dof%d)_2img.jpg' % (save_dir, epoch, it_in_epoch, it_per_epoch))
except:
traceback.print_exc()
finally:
save_path = saver.save(sess, '%s/Epoch_(%d)_(%dof%d).ckpt' % (ckpt_dir, epoch, it_in_epoch, it_per_epoch))
print('Model is saved at %s!' % save_path)
sess.close()