-
Notifications
You must be signed in to change notification settings - Fork 0
/
solvers.py
452 lines (378 loc) · 17.8 KB
/
solvers.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
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
from torch.nn import init
import matplotlib.pyplot as plt
import copy
from params import *
from config import *
import utils
import layers
import nets
def check_accuracy(loader, model):
if loader.dataset.train:
print('Checking accuracy on validation set')
else:
print('Checking accuracy on test set')
num_correct = 0
num_samples = 0
model.eval() # set model to evaluation mode
with torch.no_grad():
for x, y in loader:
x = x.to(device=device, dtype=dtype) # move to device, e.g. GPU
y = y.to(device=device, dtype=torch.long)
scores = model(x)
_, preds = scores.max(1)
num_correct += (preds == y).sum()
num_samples += preds.size(0)
acc = float(num_correct) / num_samples
print('Got %d / %d correct (%.2f)' % (num_correct, num_samples, 100 * acc))
class Solver():
def __init__(self,dloader,params):
self.dloader = dloader
self.params = params
def get_optimizer(self, model, which = 'adam'):
#optimizer = optim.Adam( filter(lambda p: p.requires_grad, model.parameters() ), lr=1e-4, betas=(0.5,0.999))
if which=='adam':
optimizer = optim.Adam( model.parameters(), lr=self.params.learning_rate, betas=self.params.adam_beta)
elif which=='SGD':
optimizer = optim.SGD( model.parameters(), lr = self.params.learning_rate, momentum=0.9)
return optimizer
class ReconSolver(Solver):
def __init__(self,model, dloader,params):
super().__init__(dloader, params)
self.model = model
def test(self,classes_num=None):
if classes_num is None:
classes_num = self.params.show_classes_num
img_size = self.params.img_size
img_channel = self.params.img_channel
img_list = []
for i in range(classes_num):
show_class = self.params.show_classes[i]
img = self.dloader.img_grouped[show_class[0]][show_class[1]]
img_list.append(img)
img = torch.stack(img_list)
img = img.to(device=device_CPU,dtype=dtype)
imgr=img.view(-1,img_channel,img_size,img_size)
imgo = self.model(imgr.to(device=device,dtype=dtype))
imgo = imgo.view(-1,img_channel*img_size*img_size).detach().to(device=device_CPU,dtype=dtype)
#print(img.shape,imgo.shape,torch.cat((img,imgo)).shape)
utils.show_images(torch.cat((img,imgo)),self.params)
class ClassifierSolver(Solver):
def __init__(self, s_enc, s_classifier, dloader, params):
super().__init__(dloader, params)
self.s_enc = s_enc
self.s_classifier = s_classifier
self.classifier = nn.Sequential(
s_enc,
s_classifier
)
self.dloader = dloader
self.it_count = 1
def train(self,epochs=1,freeze_enc = False, silence=False):
print_every = 100
model = self.classifier
model = model.to(device=device) # move the model parameters to device
if freeze_enc==True:
optimizer = self.get_optimizer(self.s_classifier)
else:
optimizer = self.get_optimizer(model)
for epoch in range(epochs):
for it, (x,y) in enumerate(self.dloader.loader_train):
model.train()
x = x.to(device = device, dtype=dtype)
y = y.to(device = device, dtype=torch.long) # QUESTION!!
scores = model(x)
loss = F.cross_entropy(scores, y)
optimizer.zero_grad()
loss.backward()
optimizer.step()
if silence==False and self.it_count % print_every == 0:
acc = self.test(mode = 'validation', silence=True)
print('iteration %d, loss = %.4f, acc=%.2f' % (self.it_count,loss.item(),acc),end='\r')
self.it_count += 1
def predict(self, x):
x = x.to(device=device, dtype=dtype) # move to device, e.g. GPU
scores = self.classifier(x)
_, preds = scores.max(1)
return preds
def test(self,mode='validation', silence=False):
if mode=='validation':
loader = self.dloader.loader_val
if silence==False:
print('Checking accuracy on validation set')
elif mode=='test':
loader = self.dloader.loader_test
if silence==False:
print('Checking accuracy on test set')
else:
print("ERROR: wrong mode!")
num_correct = 0
num_samples = 0
error_count = np.zeros(672)
error_dict = dict()
with torch.no_grad():
for x, y in loader:
x = x.to(device=device, dtype=dtype) # move to device, e.g. GPU
y = y.to(device=device, dtype=torch.long)
preds = self.predict(x)
num_correct += (preds == y).sum()
num_samples += preds.size(0)
acc = float(num_correct) / num_samples
if silence==False:
print('Got %d / %d correct (%.2f)' % (num_correct, num_samples, 100 * acc))
return acc
class DisAdvSolver(Solver):
def __init__(self, s_enc=None, z_enc=None, sz_dec=None, z_adv=None, dloader=None, params=None, loading=False, loadPath = var_save_path, loadSuffix = '', model_name = '', saving_while_training=False):
super().__init__(dloader, params)
self.dloader = dloader
self.params = params
self.model_name = model_name
self.saving_while_training = saving_while_training
if loading==False:
self.s_enc = s_enc
self.z_enc = z_enc
self.sz_dec = sz_dec
self.z_adv = z_adv
else:
self.z_enc = nets.Z_Encoder(self.params)
self.z_adv = nets.Z_AdvLayer(self.params)
self.sz_dec= nets.SZ_Decoder(self.params)
self.s_enc = nets.S_Encoder(self.params)
if loading==True:
self.load_model(loadPath,loadSuffix)
self.recon_w = self.params.recon_w
self.adv_w = self.params.adv_w
self.adv_net = layers.AdvNet(self.z_enc,self.z_adv)
self.disent_net = layers.DisentNet(self.z_enc,self.sz_dec)
self.set_mode('init')
self.adv_solver = self.get_optimizer(self.adv_net)
self.disent_solver = self.get_optimizer(self.disent_net)
self.z_enc_solver = self.get_optimizer(self.z_enc)
self.sz_dec_solver = self.get_optimizer(self.sz_dec)
self.z_adv_solver = self.get_optimizer(self.z_adv, which = 'SGD')
self.reconSolver = ReconSolver(layers.ReconNet(self.s_enc,self.z_enc,self.sz_dec), self.dloader, self.params)
self.show_every = 2000
self.save_num = 1
self.save_every = dloader.iter_per_epoch
self.log_every = 100
self.adv_disent_ratio = self.params.adv_disent_ratio
self.train_log = { 'loss':[], 'adv_acc':[] }
self.it_count = 1
def set_mode(self, mode):
if mode=='disentangle':
layers.set_trainable(self.s_enc,False)
layers.set_trainable(self.z_enc,True)
layers.set_trainable(self.sz_dec,True)
layers.set_trainable(self.z_adv,True)
elif mode == 'adversarial':
layers.set_trainable(self.s_enc,False)
layers.set_trainable(self.z_enc,False)
layers.set_trainable(self.sz_dec,False)
layers.set_trainable(self.z_adv,True)
elif mode == 'init':
layers.set_trainable(self.s_enc,True)
layers.set_trainable(self.z_enc,True)
layers.set_trainable(self.sz_dec,True)
layers.set_trainable(self.z_adv,True)
else:
print("Unknown mode")
def show_switch_latent(self, show_range = None):
if show_range is None:
show_range = self.params.show_classes_num
s_latent=[]
z_latent=[]
img_lists = []
real_lists = []
ground_truth_list = []
s_enc = self.s_enc
z_enc = self.z_enc
sz_dec = self.sz_dec
img_size = self.params.img_size
img_channel = self.params.img_channel
show_classes = self.params.show_classes
# first image is a black one
img_lists.append( torch.tensor( np.zeros(img_channel*img_size*img_size) ).to(device=device,dtype=dtype) )
for classi in range(show_range):
show_class = show_classes[classi]
img = self.dloader.img_grouped[show_class[0]][show_class[1]].view(1,img_channel,img_size,img_size)
real_lists.append(self.dloader.img_grouped[show_class[0]][show_class[1]])
img = img.to(device=device,dtype=dtype)
ground_truth_list.append( img.view( img_channel*img_size*img_size ) )
s_latent.append( s_enc(img) )
z_latent.append( z_enc(img) )
img_lists.extend( ground_truth_list )
for row in range( show_range ):
img_lists.append( ground_truth_list[row] )
for col in range( show_range ):
latent = torch.cat((s_latent[col],z_latent[row]),dim=1)
recon = sz_dec(latent)
img_lists.append(recon.view( img_channel*img_size*img_size ) )
utils.show_images(torch.stack(img_lists).detach().cpu().numpy(),self.params)
utils.show_images(torch.stack(real_lists).detach().cpu().numpy(),self.params)
def show_interpolated(self, inter_step = 4, tuples=None):
if tuples is None:
tuples = self.params.interpolated_tuples
img_size = self.params.img_size
img_channel = self.params.img_channel
inter_img1 = self.dloader.img_grouped[tuples[0][0]][tuples[0][1]].view(1,img_channel,img_size,img_size)
inter_img2 = self.dloader.img_grouped[tuples[1][0]][tuples[1][1]].view(1,img_channel,img_size,img_size)
inter_img1 = inter_img1.to(device=device,dtype=dtype)
inter_img2 = inter_img2.to(device=device,dtype=dtype)
s_enc = self.s_enc
z_enc = self.z_enc
sz_dec = self.sz_dec
s_lat1 = s_enc(inter_img1)
z_lat1 = z_enc(inter_img1)
s_lat2 = s_enc(inter_img2)
z_lat2 = z_enc(inter_img2)
weights = np.arange(0,1,1/(inter_step-1))
weights = np.append(weights,1.)
weights = torch.tensor(weights)
weights = weights.to(device=device,dtype=dtype)
#print(z_lat1,z_lat2)
img_lists = []
for row_w in weights:
for col_w in weights:
s_latent = (1-row_w) * s_lat1 + row_w * s_lat2
z_latent = (1-col_w) * z_lat1 + col_w * z_lat2
latent = torch.cat((s_latent,z_latent),dim=1)
recon = sz_dec(latent)
img_lists.append(recon.view( img_channel*img_size*img_size ) )
utils.show_images(torch.stack(img_lists).detach().cpu().numpy(),self.params)
def train(self, epochs=1):
disent_recon_loss = 1000000.
disent_adv_loss = 1000000.
disent_loss = 1000000.
adv_loss = 1000000.
dloader = self.dloader
self.max_it_count = dloader.iter_per_epoch * epochs
self.cur_it_count = 0
self.save_count = 0
self.save_num = 6
self.show_num = 1
#self.save_every = dloader.iter_per_epoch * epochs // self.save_num
self.save_every = 10000
#self.show_every = dloader.iter_per_epoch * epochs // self.show_num
self.show_every = 10000
for epoch in range(epochs):
for it, (x,y) in enumerate(dloader.loader_train):
x = x.to(device=device, dtype=dtype)
y = y.to(device=device, dtype=torch.long)
if self.it_count==1:
first_x = x
first_y = y
############ train disentangle net
if self.it_count%(self.adv_disent_ratio+1) == 0:
self.set_mode('disentangle')
self.z_enc_solver.zero_grad()
self.sz_dec_solver.zero_grad()
s_latent = self.s_enc(x)
z_latent = self.z_enc(x)
latent = torch.cat((s_latent,z_latent),dim=1)
reconstructed = self.sz_dec(latent)
disent_recon_loss = F.mse_loss(reconstructed, x, size_average=True)
scores = self.z_adv(z_latent)
disent_adv_loss = F.cross_entropy(scores, y)
disent_loss = self.recon_w * disent_recon_loss + self.adv_w * disent_adv_loss
disent_loss.backward()
self.z_enc_solver.step()
self.sz_dec_solver.step()
else:
############# train adv net
self.set_mode('adversarial')
self.z_adv_solver.zero_grad()
z_latent = self.z_enc(x)
scores = self.z_adv(z_latent)
adv_loss = F.cross_entropy(scores, y)
adv_loss.backward()
self.z_adv_solver.step()
if (self.it_count % self.log_every ==0):
self.train_log['loss'].append((disent_recon_loss, adv_loss, disent_loss))
print('progress: %.2f, total iter: %d, losses: recon:%.4f, disent:%.4f, adv:%.4f ' % (self.cur_it_count/self.max_it_count, self.it_count, disent_recon_loss, disent_loss, adv_loss), end= '\r')
if (self.it_count % self.show_every==0):
print()
self.show_switch_latent(10)
self.show_interpolated(10)
self.reconSolver.test()
plt.show()
print()
if ((self.saving_while_training==True) and (self.it_count%self.save_every==0)):
self.save_count += 1
self.save_model(suffix = self.model_name+'_num'+str(self.it_count))
print()
# it_count is total accumulated count, while cur_it_count is count in this training session
self.it_count += 1
self.cur_it_count += 1
plt.show()
def test_disentangle(self):
# S_Classifier and Z_AdvLayer are same, except input dimension...
print('training a classifier on top of z encoder...')
self.z_adv = nets.Z_AdvLayer(self.params)
z_enc_clone = copy.deepcopy(self.z_enc)
z_adv_clone = copy.deepcopy(self.z_adv)
z_enc_tester = ClassifierSolver(z_enc_clone, z_adv_clone ,self.dloader, self.params)
z_enc_tester.train(1,freeze_enc=True,silence=True)
print('testing this classifier...')
z_enc_tester.test(mode='test')
def plot_history(self):
print('reconstruction loss curve')
loss_hist = self.train_log['loss']
recon_loss_hist = [tup[0] for tup in loss_hist]
plt.plot(recon_loss_hist)
plt.show()
def save_model(self,path = var_save_path, suffix = None):
if suffix is None:
suffix = self.model_name
save_list = [self.s_enc, self.z_enc, self.sz_dec, self.z_adv]
save_list[0].m_name = 's_enc'; save_list[1].m_name = 'z_enc';
save_list[2].m_name = 'sz_dec'; save_list[3].m_name = 'z_adv';
utils.save_models(save_list,path, mode='param', mode_param = suffix)
def load_model(self, path = var_save_path, suffix = ''):
load_list = [self.s_enc, self.z_enc, self.sz_dec, self.z_adv]
load_list[0].m_name = 's_enc'; load_list[1].m_name = 'z_enc';
load_list[2].m_name = 'sz_dec'; load_list[3].m_name = 'z_adv';
utils.load_models(load_list,path = path, suffix = suffix)
class HPTuner():
def __init__(self,s_enc, dloader, params):
self.combinations = params.hyperparam_combinations()
self.epoch_num = 100
self.index = 0
self.max_index = len(self.combinations)
self.dloader = dloader
self.params = params
self.s_enc = s_enc
def tune(self, epoch_num=100):
self.epoch_num = epoch_num
is_saving = False
save_list = []
try:
print('learning_rate, hlayer_size, training_epochs_num, reg_strengths')
while True:
hyperparameters = self.combinations[self.index]
print('training in this scheme:','{}/{}'.format(self.index+1,self.max_index))
print(hyperparameters)
print()
z_enc = nets.Z_Encoder(self.params)
z_adv = nets.Z_AdvLayer(self.params)
sz_dec= nets.SZ_Decoder(self.params)
s_enc = self.s_enc
solver = DisAdvSolver(s_enc, z_enc, sz_dec, z_adv, self.dloader, self.params)
solver.show_every = self.dloader.iter_per_epoch * self.epoch_num
solver.train(self.epoch_num)
solver.plot_history()
print('saving model...')
is_saving = True
solver.save_model(suffix = str(hyperparameters))
is_saving = False
self.index += 1
if self.index == self.max_index:
break
except KeyboardInterrupt:
print('KeyboardInterrupt! you can keep going by rerun the cell, it will continue from where it was interrupted')
if is_saving == True:
print('resume saving...')
solver.save_model(suffix = str(hyperparameters))