-
Notifications
You must be signed in to change notification settings - Fork 7
/
Copy pathccyclegan_t27__hyper_params.py
486 lines (421 loc) · 21.5 KB
/
ccyclegan_t27__hyper_params.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
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
from __future__ import print_function, division
import scipy
from keras.datasets import mnist
from keras_contrib.layers.normalization.instancenormalization import InstanceNormalization
from keras.layers import Input, Dense, Reshape, Flatten, Dropout, Concatenate
from keras.layers import BatchNormalization, Activation, ZeroPadding2D
from keras.layers.advanced_activations import LeakyReLU
from keras.layers.convolutional import UpSampling2D, Conv2D
from keras.layers import BatchNormalization, Activation, Embedding, ZeroPadding2D
from keras.models import Sequential, Model
from keras.optimizers import Adam
from keras.layers import Input, Dense, Reshape, Flatten, Dropout, multiply
from keras.layers import Reshape
import datetime
import matplotlib.pyplot as plt
import sys
from data_loader import DataLoader
from keras.layers import Concatenate, Dense, LSTM, Input, concatenate
from keras import backend as K
import numpy as np
import pandas as pd
import os
import random
import cv2
import tensorflow as tf
from keras.utils import to_categorical
from sklearn.metrics import accuracy_score
from models import *
from fid import get_fid
from itertools import product
class CCycleGAN():
def __init__(self,img_rows = 48,img_cols = 48,channels = 1, num_classes=7,
d_gan_loss_w=1,d_cl_loss_w=1,
g_gan_loss_w=1,g_cl_loss_w=1,
rec_loss_w=1,
adam_lr=0.0002,adam_beta_1=0.5,adam_beta_2=0.999):
# Input shape
self.img_rows = img_rows
self.img_cols = img_cols
self.channels = channels
self.img_shape = (self.img_rows, self.img_cols, self.channels)
self.num_classes = num_classes
# Loss weights
self.d_gan_loss_w = d_gan_loss_w
self.d_cl_loss_w = d_cl_loss_w
self.g_gan_loss_w = g_gan_loss_w
self.g_cl_loss_w = g_cl_loss_w
self.rec_loss_w = rec_loss_w
# optmizer params
self.adam_lr = adam_lr
self.adam_beta_1 = adam_beta_1
self.adam_beta_2 = adam_beta_2
# Configure data loader
self.dataset_name = 'fer2013'
self.data_loader = DataLoader(dataset_name=self.dataset_name,img_res=self.img_shape,use_test_in_batch=True)
# label dict
self.lab_dict = {0: "Angry", 1: "Disgust" , 2: "Fear" , 3: "Happy" , 4: "Sad" , 5: "Surprise" , 6: "Neutral"}
# Number of filters in the first layer of G and D
self.gf = 32
self.df = 64
optimizer = Adam(self.adam_lr, self.adam_beta_1, self.adam_beta_2)
# Build and compile the discriminators
self.d = build_discriminator(img_shape=self.img_shape,df=64,num_classes=self.num_classes,act_multi_label='sigmoid')
print("******** Discriminator/Classifier ********")
self.d.summary()
self.d.compile(loss=[
'binary_crossentropy', # gan
'binary_crossentropy' # class
],
optimizer=optimizer,
metrics=['accuracy'],
loss_weights=[
self.d_gan_loss_w , # gan
self.d_cl_loss_w # class
])
#-------------------------
# Construct Computational
# Graph of Generators
#-------------------------
# Build the generators
self.g_enc , self.g_dec = build_generator_enc_dec(img_shape=(48,48,1),gf=64,num_classes=7,channels=1,tranform_layer=True)
print("******** Generator_ENC ********")
self.g_enc.summary()
print("******** Generator_DEC ********")
self.g_dec.summary()
# Input images from both domains
img = Input(shape=self.img_shape)
label0 = Input(shape=(self.num_classes,))
label1 = Input(shape=(self.num_classes,))
# Translate images to the other domain
z1,z2,z3,z4 = self.g_enc(img)
fake = self.g_dec([z1,z2,z3,z4,label1])
# Translate images back to original domain
reconstr = self.g_dec([z1,z2,z3,z4,label0])
# For the combined model we will only train the generators
self.d.trainable = False
# Discriminators determines validity of translated images gan_prob,class_prob [label,img], [gan_prob,class_prob]
gan_valid , class_valid = self.d(fake)
# Combined model trains generators to fool discriminators
self.combined = Model(inputs=[img,label0,label1],
outputs=[ gan_valid, class_valid,
reconstr])
self.combined.compile(loss=['binary_crossentropy','categorical_crossentropy',
'mae'],
loss_weights=[
self.g_gan_loss_w , # g_loss gan
self.g_cl_loss_w , # g_loss class
self.rec_loss_w # reconstruction loss
],
optimizer=optimizer)
def generate_new_labels(self,labels0):
labels1 = []
for i in range(len(labels0)):
allowed_values = list(range(0, self.num_classes))
allowed_values.remove(labels0[i])
labels1.append(random.choice(allowed_values))
return np.array(labels1,'int32')
def generate_new_labels_all(self,labels0):
labels_all = []
for i in range(len(labels0)):
allowed_values = list(range(0, self.num_classes))
allowed_values.remove(labels0[i])
labels_all.append(np.array(allowed_values,'int32'))
return np.array(labels_all,'int32')
def train(self, epochs, batch_size=1, sample_interval=50 , d_g_ratio=5, tun_grid=None):
start_time = datetime.datetime.now()
# logs
epoch_history, batch_i_history, = [] , []
d_gan_loss_history, d_gan_accuracy_history, d_cl_loss_history, d_cl_accuracy_history = [], [], [], []
g_gan_loss_history, g_cl_loss_history = [] , []
reconstr_history = []
# Adversarial loss ground truths
valid = np.ones((batch_size,1) )
fake = np.zeros((batch_size,1) )
null_labels = np.zeros((batch_size,7) )
for epoch in range(epochs):
for batch_i, (labels0 , imgs) in enumerate(self.data_loader.load_batch(batch_size=batch_size)):
labels1_all = self.generate_new_labels_all(labels0)
labels0_cat = to_categorical(labels0, num_classes=self.num_classes)
#
labels1_all_1 = to_categorical(labels1_all[:,0], num_classes=self.num_classes)
labels1_all_2 = to_categorical(labels1_all[:,1], num_classes=self.num_classes)
labels1_all_3 = to_categorical(labels1_all[:,2], num_classes=self.num_classes)
labels1_all_4 = to_categorical(labels1_all[:,3], num_classes=self.num_classes)
labels1_all_5 = to_categorical(labels1_all[:,4], num_classes=self.num_classes)
labels1_all_6 = to_categorical(labels1_all[:,5], num_classes=self.num_classes)
# ----------------------
# Train Discriminators
# ----------------------
# Translate images to opposite domain
zs1,zs2,zs3,zs4 = self.g_enc.predict(imgs)
fakes_1 = self.g_dec.predict([zs1,zs2,zs3,zs4,labels1_all_1])
fakes_2 = self.g_dec.predict([zs1,zs2,zs3,zs4,labels1_all_2])
fakes_3 = self.g_dec.predict([zs1,zs2,zs3,zs4,labels1_all_3])
fakes_4 = self.g_dec.predict([zs1,zs2,zs3,zs4,labels1_all_4])
fakes_5 = self.g_dec.predict([zs1,zs2,zs3,zs4,labels1_all_5])
fakes_6 = self.g_dec.predict([zs1,zs2,zs3,zs4,labels1_all_6])
# Train the discriminators (original images = real / translated = Fake)
idx = np.random.permutation(self.num_classes*labels0.shape[0])
_labels_cat = np.concatenate([labels0_cat,
null_labels,
null_labels,
null_labels,
null_labels,
null_labels,
null_labels])
_imgs = np.concatenate([imgs,
fakes_1,
fakes_2,
fakes_3,
fakes_4,
fakes_5,
fakes_6])
_vf = np.concatenate([valid,fake,fake,fake,fake,fake,fake])
_labels_cat = _labels_cat[idx]
_imgs = _imgs[idx]
_vf = _vf[idx]
d_loss = self.d.train_on_batch(_imgs, [_vf,_labels_cat])
if batch_i % d_g_ratio == 0:
# ------------------
# Train Generators
# ------------------
_imgs = np.concatenate([
imgs,
imgs,
imgs,
imgs,
imgs,
imgs])
_labels0_cat = np.concatenate([
labels0_cat,
labels0_cat,
labels0_cat,
labels0_cat,
labels0_cat,
labels0_cat])
_labels1_all_other = np.concatenate([
labels1_all_1,
labels1_all_2,
labels1_all_3,
labels1_all_4,
labels1_all_5,
labels1_all_6])
# I know this should be outside the loop; left here to make code more understandable
_valid = np.concatenate([
valid,
valid,
valid,
valid,
valid,
valid])
idx = np.random.permutation((self.num_classes-1)*labels0.shape[0])
_imgs = _imgs[idx]
_labels0_cat = _labels0_cat[idx]
_labels1_all_other = _labels1_all_other[idx]
_valid = _valid[idx]
# Train the generators
g_loss = self.combined.train_on_batch([_imgs, _labels0_cat, _labels1_all_other],
[_valid, _labels1_all_other, _imgs])
elapsed_time = datetime.datetime.now() - start_time
print ("[Epoch %d/%d] [Batch %d/%d] [D_gan loss: %f, acc_gan: %3d%%] [D_cl loss: %f, acc_cl: %3d%%] [G_gan loss: %05f, G_cl: %05f, recon: %05f] time: %s " \
% ( epoch, epochs,
batch_i, self.data_loader.n_batches,
d_loss[1],100*d_loss[3],d_loss[2],100*d_loss[4],
g_loss[1],g_loss[2],g_loss[3],
elapsed_time))
# log
epoch_history.append(epoch)
batch_i_history.append(batch_i)
d_gan_loss_history.append(d_loss[1])
d_gan_accuracy_history.append(100*d_loss[3])
d_cl_loss_history.append(d_loss[2])
d_cl_accuracy_history.append(100*d_loss[4])
g_gan_loss_history.append(g_loss[1])
g_cl_loss_history.append(g_loss[2])
reconstr_history.append(g_loss[3])
# If at save interval => save generated image samples
if batch_i % sample_interval == 0:
pass
#self.sample_images(epoch, batch_i)
#self.sample_images(epoch, batch_i,use_leo=True)
#train_history = pd.DataFrame({
# 'epoch': epoch_history,
# 'batch': batch_i_history,
# 'd_gan_loss': d_gan_loss_history,
# 'd_gan_accuracy' : d_gan_accuracy_history,
# 'd_cl_loss': d_cl_loss_history,
# 'd_cl_accuracy': d_cl_accuracy_history,
# 'g_gan_loss': g_gan_loss_history,
# 'g_cl_loss': g_cl_loss_history,
# 'reconstr_loss': reconstr_history
#})
#train_history.to_csv(str(sys.argv[0]).split('.')[0]+'_train_log.csv',index=False)
## ---> here measure and log FIS
## FIS
print(">>> measuring FIS ... ")
self.measure_fis(epoch,tun_grid)
def measure_fis(self, epoch,tun_grid,sample_size=1000):
for i in range(self.num_classes):
_ , batch_images = self.data_loader.load_data(domain=None, batch_size=sample_size, is_testing=True,convertRGB=False)
_ , batch_images_i = self.data_loader.load_data(domain=i, batch_size=sample_size, is_testing=True,convertRGB=True)
for j in range(sample_size):
batch_images_i[j] = (batch_images_i[j]+1.)*127.5 #un-normalize
#batch_images = np.transpose(batch_images, (0,3,1,2) )
batch_images_i = np.transpose(batch_images_i, (0,3,1,2) )
#
labels1_all_i = to_categorical([i]*sample_size, num_classes=self.num_classes)
zs1_,zs2_,zs3_,zs4_ = self.g_enc.predict(batch_images)
_fake_i = self.g_dec.predict([zs1_,zs2_,zs3_,zs4_,labels1_all_i])
fake_i = np.zeros((sample_size,self.img_rows,self.img_cols,3))
for j in range(sample_size):
_fake_i[j] = (_fake_i[j]+1.)*127.5
fake_i[j] = cv2.cvtColor(_fake_i[j], cv2.COLOR_GRAY2RGB)
fake_i = np.transpose(fake_i, (0,3,1,2) )
fid_i = get_fid(fake_i, batch_images_i)
print(i,fid_i)
# update grid
assert tun_grid[(tun_grid['d_gan_loss_w']==self.d_gan_loss_w)
& (tun_grid['d_cl_loss_w']==self.d_cl_loss_w)
& (tun_grid['g_gan_loss_w']==self.g_gan_loss_w)
& (tun_grid['g_cl_loss_w']==self.g_cl_loss_w)
& (tun_grid['rec_loss_w']==self.rec_loss_w)
& (tun_grid['adam_lr']==self.adam_lr)
& (tun_grid['epoch']==epoch)].shape[0] == 1
tun_grid.loc[(tun_grid['d_gan_loss_w']==self.d_gan_loss_w)
& (tun_grid['d_cl_loss_w']==self.d_cl_loss_w)
& (tun_grid['g_gan_loss_w']==self.g_gan_loss_w)
& (tun_grid['g_cl_loss_w']==self.g_cl_loss_w)
& (tun_grid['rec_loss_w']==self.rec_loss_w)
& (tun_grid['adam_lr']==self.adam_lr)
& (tun_grid['epoch']==epoch) , 'lab'+str(i)+'_FIS' ] = fid_i
# store
tun_grid.to_csv(str(sys.argv[0]).split('.')[0]+'_fis.csv',index=False)
def sample_images(self, epoch, batch_i, use_leo=False):
## disc
labels0_d , imgs_d = self.data_loader.load_data(batch_size=64, is_testing=True)
gan_pred_prob,class_pred_prob = self.d.predict(imgs_d)
gan_pred = (gan_pred_prob > 0.5)*1.0
gan_pred = gan_pred.reshape((64,))
class_pred = np.argmax(class_pred_prob,axis=1)
gan_test_accuracy = accuracy_score(y_true=np.ones(64), y_pred=gan_pred)
class_test_accuracy = accuracy_score(y_true=labels0_d, y_pred=class_pred)
print("*** TEST *** [D_gan accuracy :",gan_test_accuracy,"] [D_cl accuracy :",class_test_accuracy,"]")
## gen
if use_leo:
labels0_ , imgs_ = self.data_loader.load_leo()
else:
labels0_ , imgs_ = self.data_loader.load_data(batch_size=1, is_testing=True)
labels1_all = self.generate_new_labels_all(labels0_)
labels0_cat = to_categorical(labels0_, num_classes=self.num_classes)
labels1_all_1 = to_categorical(labels1_all[:,0], num_classes=self.num_classes)
labels1_all_2 = to_categorical(labels1_all[:,1], num_classes=self.num_classes)
labels1_all_3 = to_categorical(labels1_all[:,2], num_classes=self.num_classes)
labels1_all_4 = to_categorical(labels1_all[:,3], num_classes=self.num_classes)
labels1_all_5 = to_categorical(labels1_all[:,4], num_classes=self.num_classes)
labels1_all_6 = to_categorical(labels1_all[:,5], num_classes=self.num_classes)
# Translate images
zs1_,zs2_,zs3_,zs4_ = self.g_enc.predict(imgs_)
fake_1 = self.g_dec.predict([zs1_,zs2_,zs3_,zs4_,labels1_all_1])
fake_2 = self.g_dec.predict([zs1_,zs2_,zs3_,zs4_,labels1_all_2])
fake_3 = self.g_dec.predict([zs1_,zs2_,zs3_,zs4_,labels1_all_3])
fake_4 = self.g_dec.predict([zs1_,zs2_,zs3_,zs4_,labels1_all_4])
fake_5 = self.g_dec.predict([zs1_,zs2_,zs3_,zs4_,labels1_all_5])
fake_6 = self.g_dec.predict([zs1_,zs2_,zs3_,zs4_,labels1_all_6])
# Reconstruct image
reconstr_ = self.g_dec.predict([zs1_,zs2_,zs3_,zs4_,labels0_cat])
gen_imgs = np.concatenate([imgs_,
fake_1,
fake_2,
fake_3,
fake_4,
fake_5,
fake_6,
reconstr_])
# Rescale images 0 - 1
gen_imgs = 0.5 * gen_imgs + 0.5
titles = ['Orig:'+str(self.lab_dict[labels0_.item(0)]),
'Trans:'+str(self.lab_dict[labels1_all[:,0].item(0)]),
'Trans:'+str(self.lab_dict[labels1_all[:,1].item(0)]),
'Trans:'+str(self.lab_dict[labels1_all[:,2].item(0)]),
'Trans:'+str(self.lab_dict[labels1_all[:,3].item(0)]),
'Trans:'+str(self.lab_dict[labels1_all[:,4].item(0)]),
'Trans:'+str(self.lab_dict[labels1_all[:,5].item(0)]),
'Reconstr.']
r, c = 2, 4
fig, axs = plt.subplots(r, c)
plt.subplots_adjust(hspace=0)
if not os.path.exists( "images/%s/"% (self.dataset_name)):
os.makedirs( "images/%s/"% (self.dataset_name) )
cnt = 0
for i in range(r):
for j in range(c):
axs[i,j].imshow(gen_imgs[cnt].reshape((self.img_rows,self.img_cols)),cmap='gray')
axs[i,j].set_title(titles[cnt])
axs[i,j].axis('off')
cnt += 1
if use_leo:
fig.savefig("images/%s/%d_%d_leo.png" % (self.dataset_name, epoch, batch_i))
else:
fig.savefig("images/%s/%d_%d.png" % (self.dataset_name, epoch, batch_i))
plt.close()
def generate_tuning_grid(MAX_EPOCH = 90,
tuned_parameters = {'d_gan_loss_w': [1],
'd_cl_loss_w' : [1],
'g_gan_loss_w': [1,2],
'g_cl_loss_w': [1,2],
'rec_loss_w': [1],
'adam_lr': [0.0002,0.0001]}):
tuned_parameters['epoch'] = list(range(MAX_EPOCH))
#print([dict(zip(tuned_parameters.keys(),v)) for v in product(*tuned_parameters.values())])
row_list = [dict(zip(tuned_parameters.keys(),v)) for v in product(*tuned_parameters.values())]
row_dict = {i: row_list[i] for i in range(len(row_list)) }
tun_grid = pd.DataFrame.from_dict(row_dict, orient='index')
#
#tun_grid['epoch'] = -1
#
tun_grid['lab0_FIS'] = -1
tun_grid['lab1_FIS'] = -1
tun_grid['lab2_FIS'] = -1
tun_grid['lab3_FIS'] = -1
tun_grid['lab4_FIS'] = -1
tun_grid['lab5_FIS'] = -1
tun_grid['lab6_FIS'] = -1
#
#tun_grid['FIS'] = -1
return tun_grid
if __name__ == '__main__':
## create hyper-params grid
RECOVER_POINT = 2 # set to 0 to do all
MAX_EPOCH = 200
tuned_parameters = {'d_gan_loss_w': [1],
'd_cl_loss_w' : [1],
'g_gan_loss_w': [1,2],
'g_cl_loss_w': [1,2],
'rec_loss_w': [1],
'adam_lr': [0.0002,0.0001]}
run_opts = [dict(zip(tuned_parameters.keys(),v)) for v in product(*tuned_parameters.values())]
if RECOVER_POINT == 0:
tun_grid = generate_tuning_grid(MAX_EPOCH = MAX_EPOCH, tuned_parameters = tuned_parameters)
else:
print(">> loading tune grid ...")
tun_grid = pd.read_csv('ccyclegan_t27__hyper_params_fis_e200.csv')
print(tun_grid.head())
print("...")
print(tun_grid.tail())
## tune
for i,opt in enumerate(run_opts):
#K.clear_session()
print("*************************************************************")
print(i,"/",len(run_opts))
print(opt)
if i < RECOVER_POINT:
print("> skipping ... ")
else:
print("> computing ...")
gan = CCycleGAN(
d_gan_loss_w=opt['d_gan_loss_w'],d_cl_loss_w=opt['d_cl_loss_w'],
g_gan_loss_w=opt['g_gan_loss_w'],g_cl_loss_w=opt['g_cl_loss_w'],
rec_loss_w=opt['rec_loss_w'],
adam_lr=opt['adam_lr'],adam_beta_1=0.5,adam_beta_2=0.999)
gan.train(epochs=MAX_EPOCH, batch_size=64, tun_grid=tun_grid)