-
Notifications
You must be signed in to change notification settings - Fork 39
/
Copy pathtrainer.py
499 lines (448 loc) · 21.1 KB
/
trainer.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
487
488
489
490
491
492
493
494
495
496
497
498
499
from __future__ import print_function
from six.moves import range
import torch
import torch.optim as optim
from torch.autograd import Variable
import torch.backends.cudnn as cudnn
from PIL import Image
from cfg.config import cfg
from miscc.utils import mkdir_p
from miscc.utils import build_super_images, build_super_images2
from miscc.utils import weights_init, load_params, copy_G_params
from model import G_DCGAN, G_NET
from datasets import prepare_data
from model import RNN_ENCODER, CNN_ENCODER, CAPTION_CNN, CAPTION_RNN
from miscc.losses import words_loss
from miscc.losses import discriminator_loss, generator_loss, KL_loss
import os
import time
import numpy as np
# MirrorGAN
class Trainer(object):
def __init__(self, output_dir, data_loader, n_words, ixtoword):
if cfg.TRAIN.FLAG:
self.model_dir = os.path.join(output_dir, 'Model')
self.image_dir = os.path.join(output_dir, 'Image')
mkdir_p(self.model_dir)
mkdir_p(self.image_dir)
torch.cuda.set_device(cfg.GPU_ID)
cudnn.benchmark = True
self.batch_size = cfg.TRAIN.BATCH_SIZE
self.max_epoch = cfg.TRAIN.MAX_EPOCH
self.snapshot_interval = cfg.TRAIN.SNAPSHOT_INTERVAL
self.n_words = n_words
self.ixtoword = ixtoword
self.data_loader = data_loader
self.num_batches = len(self.data_loader)
def build_models(self):
# text encoders
if cfg.TRAIN.NET_E == '':
print('Error: no pretrained text-image encoders')
return
image_encoder = CNN_ENCODER(cfg.TEXT.EMBEDDING_DIM)
img_encoder_path = cfg.TRAIN.NET_E.replace('text_encoder', 'image_encoder')
state_dict = \
torch.load(img_encoder_path, map_location=lambda storage, loc: storage)
image_encoder.load_state_dict(state_dict)
for p in image_encoder.parameters():
p.requires_grad = False
print('Load image encoder from:', img_encoder_path)
image_encoder.eval()
text_encoder = \
RNN_ENCODER(self.n_words, nhidden=cfg.TEXT.EMBEDDING_DIM)
state_dict = \
torch.load(cfg.TRAIN.NET_E,
map_location=lambda storage, loc: storage)
text_encoder.load_state_dict(state_dict)
for p in text_encoder.parameters():
p.requires_grad = False
print('Load text encoder from:', cfg.TRAIN.NET_E)
text_encoder.eval()
# Caption models - cnn_encoder and rnn_decoder
caption_cnn = CAPTION_CNN(cfg.CAP.embed_size)
caption_cnn.load_state_dict(torch.load(cfg.CAP.caption_cnn_path, map_location=lambda storage, loc: storage))
for p in caption_cnn.parameters():
p.requires_grad = False
print('Load caption model from:', cfg.CAP.caption_cnn_path)
caption_cnn.eval()
caption_rnn = CAPTION_RNN(cfg.CAP.embed_size, cfg.CAP.hidden_size * 2, self.n_words, cfg.CAP.num_layers)
caption_rnn.load_state_dict(torch.load(cfg.CAP.caption_rnn_path, map_location=lambda storage, loc: storage))
for p in caption_rnn.parameters():
p.requires_grad = False
print('Load caption model from:', cfg.CAP.caption_rnn_path)
# Generator and Discriminator:
netsD = []
if cfg.GAN.B_DCGAN:
if cfg.TREE.BRANCH_NUM == 1:
from model import D_NET64 as D_NET
elif cfg.TREE.BRANCH_NUM == 2:
from model import D_NET128 as D_NET
else: # cfg.TREE.BRANCH_NUM == 3:
from model import D_NET256 as D_NET
netG = G_DCGAN()
netsD = [D_NET(b_jcu=False)]
else:
from model import D_NET64, D_NET128, D_NET256
netG = G_NET()
if cfg.TREE.BRANCH_NUM > 0:
netsD.append(D_NET64())
if cfg.TREE.BRANCH_NUM > 1:
netsD.append(D_NET128())
if cfg.TREE.BRANCH_NUM > 2:
netsD.append(D_NET256())
netG.apply(weights_init)
# print(netG)
for i in range(len(netsD)):
netsD[i].apply(weights_init)
# print(netsD[i])
print('# of netsD', len(netsD))
epoch = 0
if cfg.TRAIN.NET_G != '':
state_dict = \
torch.load(cfg.TRAIN.NET_G, map_location=lambda storage, loc: storage)
netG.load_state_dict(state_dict)
print('Load G from: ', cfg.TRAIN.NET_G)
istart = cfg.TRAIN.NET_G.rfind('_') + 1
iend = cfg.TRAIN.NET_G.rfind('.')
epoch = cfg.TRAIN.NET_G[istart:iend]
epoch = int(epoch) + 1
if cfg.TRAIN.B_NET_D:
Gname = cfg.TRAIN.NET_G
for i in range(len(netsD)):
s_tmp = Gname[:Gname.rfind('/')]
Dname = '%s/netD%d.pth' % (s_tmp, i)
print('Load D from: ', Dname)
state_dict = \
torch.load(Dname, map_location=lambda storage, loc: storage)
netsD[i].load_state_dict(state_dict)
if cfg.CUDA:
text_encoder = text_encoder.cuda()
image_encoder = image_encoder.cuda()
caption_cnn = caption_cnn.cuda()
caption_rnn = caption_rnn.cuda()
netG.cuda()
for i in range(len(netsD)):
netsD[i].cuda()
return [text_encoder, image_encoder, caption_cnn, caption_rnn, netG, netsD, epoch]
def define_optimizers(self, netG, netsD):
optimizersD = []
num_Ds = len(netsD)
for i in range(num_Ds):
opt = optim.Adam(netsD[i].parameters(),
lr=cfg.TRAIN.DISCRIMINATOR_LR,
betas=(0.5, 0.999))
optimizersD.append(opt)
optimizerG = optim.Adam(netG.parameters(),
lr=cfg.TRAIN.GENERATOR_LR,
betas=(0.5, 0.999))
return optimizerG, optimizersD
def prepare_labels(self):
batch_size = self.batch_size
real_labels = Variable(torch.FloatTensor(batch_size).fill_(1))
fake_labels = Variable(torch.FloatTensor(batch_size).fill_(0))
match_labels = Variable(torch.LongTensor(range(batch_size)))
if cfg.CUDA:
real_labels = real_labels.cuda()
fake_labels = fake_labels.cuda()
match_labels = match_labels.cuda()
return real_labels, fake_labels, match_labels
def save_model(self, netG, avg_param_G, netsD, epoch):
backup_para = copy_G_params(netG)
load_params(netG, avg_param_G)
torch.save(netG.state_dict(),
'%s/netG_epoch_%d.pth' % (self.model_dir, epoch))
load_params(netG, backup_para)
#
for i in range(len(netsD)):
netD = netsD[i]
torch.save(netD.state_dict(),
'%s/netD%d.pth' % (self.model_dir, i))
print('Save G/Ds models.')
def set_requires_grad_value(self, models_list, brequires):
for i in range(len(models_list)):
for p in models_list[i].parameters():
p.requires_grad = brequires
def save_img_results(self, netG, noise, sent_emb, words_embs, mask,
image_encoder, captions, cap_lens,
gen_iterations, name='current'):
# Save images
fake_imgs, attention_maps, _, _ = netG(noise, sent_emb, words_embs, mask)
for i in range(len(attention_maps)):
if len(fake_imgs) > 1:
img = fake_imgs[i + 1].detach().cpu()
lr_img = fake_imgs[i].detach().cpu()
else:
img = fake_imgs[0].detach().cpu()
lr_img = None
attn_maps = attention_maps[i]
att_sze = attn_maps.size(2)
img_set, _ = \
build_super_images(img, captions, self.ixtoword,
attn_maps, att_sze, lr_imgs=lr_img)
if img_set is not None:
im = Image.fromarray(img_set)
fullpath = '%s/G_%s_%d_%d.png' \
% (self.image_dir, name, gen_iterations, i)
im.save(fullpath)
i = -1
img = fake_imgs[i].detach()
region_features, _ = image_encoder(img)
att_sze = region_features.size(2)
_, _, att_maps = words_loss(region_features.detach(),
words_embs.detach(),
None, cap_lens,
None, self.batch_size)
img_set, _ = \
build_super_images(fake_imgs[i].detach().cpu(),
captions, self.ixtoword, att_maps, att_sze)
if img_set is not None:
im = Image.fromarray(img_set)
fullpath = '%s/D_%s_%d.png' \
% (self.image_dir, name, gen_iterations)
im.save(fullpath)
def train(self):
text_encoder, image_encoder, caption_cnn, caption_rnn, netG, netsD, start_epoch = self.build_models()
avg_param_G = copy_G_params(netG)
optimizerG, optimizersD = self.define_optimizers(netG, netsD)
real_labels, fake_labels, match_labels = self.prepare_labels()
batch_size = self.batch_size
nz = cfg.GAN.Z_DIM
noise = Variable(torch.FloatTensor(batch_size, nz))
fixed_noise = Variable(torch.FloatTensor(batch_size, nz).normal_(0, 1))
if cfg.CUDA:
noise, fixed_noise = noise.cuda(), fixed_noise.cuda()
gen_iterations = 0
for epoch in range(start_epoch, self.max_epoch):
start_t = time.time()
data_iter = iter(self.data_loader)
step = 0
while step < self.num_batches:
# (1) Prepare training data and Compute text embeddings
data = data_iter.next()
imgs, captions, cap_lens, class_ids, keys = prepare_data(data)
hidden = text_encoder.init_hidden(batch_size)
# words_embs: batch_size x nef x seq_len
# sent_emb: batch_size x nef
words_embs, sent_emb = text_encoder(captions, cap_lens, hidden)
words_embs, sent_emb = words_embs.detach(), sent_emb.detach()
mask = (captions == 0)
num_words = words_embs.size(2)
if mask.size(1) > num_words:
mask = mask[:, :num_words]
# (2) Generate fake images
noise.data.normal_(0, 1)
fake_imgs, _, mu, logvar = netG(noise, sent_emb, words_embs, mask)
# (3) Update D network
errD_total = 0
D_logs = ''
for i in range(len(netsD)):
netsD[i].zero_grad()
errD = discriminator_loss(netsD[i], imgs[i], fake_imgs[i],
sent_emb, real_labels, fake_labels)
# backward and update parameters
errD.backward()
optimizersD[i].step()
errD_total += errD
D_logs += 'errD%d: %.2f ' % (i, errD.data[0])
# (4) Update G network: maximize log(D(G(z)))
# compute total loss for training G
step += 1
gen_iterations += 1
netG.zero_grad()
errG_total, G_logs = \
generator_loss(netsD, image_encoder, caption_cnn, caption_rnn, captions, fake_imgs, real_labels,
words_embs, sent_emb, match_labels, cap_lens, class_ids)
kl_loss = KL_loss(mu, logvar)
errG_total += kl_loss
G_logs += 'kl_loss: %.2f ' % kl_loss.data[0]
# backward and update parameters
errG_total.backward()
optimizerG.step()
for p, avg_p in zip(netG.parameters(), avg_param_G):
avg_p.mul_(0.999).add_(0.001, p.data)
if gen_iterations % 100 == 0:
print(D_logs + '\n' + G_logs)
# save images
if gen_iterations % 1000 == 0:
backup_para = copy_G_params(netG)
load_params(netG, avg_param_G)
self.save_img_results(netG, fixed_noise, sent_emb,
words_embs, mask, image_encoder,
captions, cap_lens, epoch, name='average')
load_params(netG, backup_para)
end_t = time.time()
print('''[%d/%d][%d]
Loss_D: %.2f Loss_G: %.2f Time: %.2fs'''
% (epoch, self.max_epoch, self.num_batches,
errD_total.data[0], errG_total.data[0],
end_t - start_t))
if epoch % cfg.TRAIN.SNAPSHOT_INTERVAL == 0: # and epoch != 0:
self.save_model(netG, avg_param_G, netsD, epoch)
self.save_model(netG, avg_param_G, netsD, self.max_epoch)
def save_singleimages(self, images, filenames, save_dir,
split_dir, sentenceID=0):
for i in range(images.size(0)):
s_tmp = '%s/single_samples/%s/%s' % \
(save_dir, split_dir, filenames[i])
folder = s_tmp[:s_tmp.rfind('/')]
if not os.path.isdir(folder):
print('Make a new folder: ', folder)
mkdir_p(folder)
fullpath = '%s_%d.jpg' % (s_tmp, sentenceID)
# range from [-1, 1] to [0, 1]
# img = (images[i] + 1.0) / 2
img = images[i].add(1).div(2).mul(255).clamp(0, 255).byte()
# range from [0, 1] to [0, 255]
ndarr = img.permute(1, 2, 0).data.cpu().numpy()
im = Image.fromarray(ndarr)
im.save(fullpath)
def sampling(self, split_dir):
if cfg.TRAIN.NET_G == '':
print('Error: the path for model is not found!')
else:
if split_dir == 'test':
split_dir = 'valid'
# Build and load the generator
if cfg.GAN.B_DCGAN:
netG = G_DCGAN()
else:
netG = G_NET()
netG.apply(weights_init)
netG.cuda()
netG.eval()
#
text_encoder = RNN_ENCODER(self.n_words, nhidden=cfg.TEXT.EMBEDDING_DIM)
state_dict = \
torch.load(cfg.TRAIN.NET_E, map_location=lambda storage, loc: storage)
text_encoder.load_state_dict(state_dict)
print('Load text encoder from:', cfg.TRAIN.NET_E)
text_encoder = text_encoder.cuda()
text_encoder.eval()
batch_size = self.batch_size
nz = cfg.GAN.Z_DIM
noise = Variable(torch.FloatTensor(batch_size, nz), volatile=True)
noise = noise.cuda()
model_dir = cfg.TRAIN.NET_G
state_dict = \
torch.load(model_dir, map_location=lambda storage, loc: storage)
netG.load_state_dict(state_dict)
print('Load G from: ', model_dir)
# the path to save generated images
s_tmp = model_dir[:model_dir.rfind('.pth')]
save_dir = '%s/%s' % (s_tmp, split_dir)
mkdir_p(save_dir)
cnt = 0
for _ in range(1): # (cfg.TEXT.CAPTIONS_PER_IMAGE):
for step, data in enumerate(self.data_loader, 0):
cnt += batch_size
if step % 100 == 0:
print('step: ', step)
# if step > 50:
# break
imgs, captions, cap_lens, class_ids, keys = prepare_data(data)
hidden = text_encoder.init_hidden(batch_size)
# words_embs: batch_size x nef x seq_len
# sent_emb: batch_size x nef
words_embs, sent_emb = text_encoder(captions, cap_lens, hidden)
words_embs, sent_emb = words_embs.detach(), sent_emb.detach()
mask = (captions == 0)
num_words = words_embs.size(2)
if mask.size(1) > num_words:
mask = mask[:, :num_words]
# (2) Generate fake images
noise.data.normal_(0, 1)
fake_imgs, _, _, _ = netG(noise, sent_emb, words_embs, mask)
for j in range(batch_size):
s_tmp = '%s/single/%s' % (save_dir, keys[j])
folder = s_tmp[:s_tmp.rfind('/')]
if not os.path.isdir(folder):
print('Make a new folder: ', folder)
mkdir_p(folder)
k = -1
# for k in range(len(fake_imgs)):
im = fake_imgs[k][j].data.cpu().numpy()
# [-1, 1] --> [0, 255]
im = (im + 1.0) * 127.5
im = im.astype(np.uint8)
im = np.transpose(im, (1, 2, 0))
im = Image.fromarray(im)
fullpath = '%s_s%d.png' % (s_tmp, k)
im.save(fullpath)
def gen_example(self, data_dic):
if cfg.TRAIN.NET_G == '':
print('Error: the path for morels is not found!')
else:
# Build and load the generator
text_encoder = \
RNN_ENCODER(self.n_words, nhidden=cfg.TEXT.EMBEDDING_DIM)
state_dict = \
torch.load(cfg.TRAIN.NET_E, map_location=lambda storage, loc: storage)
text_encoder.load_state_dict(state_dict)
print('Load text encoder from:', cfg.TRAIN.NET_E)
text_encoder = text_encoder.cuda()
text_encoder.eval()
# the path to save generated images
if cfg.GAN.B_DCGAN:
netG = G_DCGAN()
else:
netG = G_NET()
s_tmp = cfg.TRAIN.NET_G[:cfg.TRAIN.NET_G.rfind('.pth')]
model_dir = cfg.TRAIN.NET_G
state_dict = \
torch.load(model_dir, map_location=lambda storage, loc: storage)
netG.load_state_dict(state_dict)
print('Load G from: ', model_dir)
netG.cuda()
netG.eval()
for key in data_dic:
save_dir = '%s/%s' % (s_tmp, key)
mkdir_p(save_dir)
captions, cap_lens, sorted_indices = data_dic[key]
batch_size = captions.shape[0]
nz = cfg.GAN.Z_DIM
captions = Variable(torch.from_numpy(captions), volatile=True)
cap_lens = Variable(torch.from_numpy(cap_lens), volatile=True)
captions = captions.cuda()
cap_lens = cap_lens.cuda()
for i in range(1): # 16
noise = Variable(torch.FloatTensor(batch_size, nz), volatile=True)
noise = noise.cuda()
# (1) Extract text embeddings
hidden = text_encoder.init_hidden(batch_size)
# words_embs: batch_size x nef x seq_len
# sent_emb: batch_size x nef
words_embs, sent_emb = text_encoder(captions, cap_lens, hidden)
mask = (captions == 0)
# (2) Generate fake images
noise.data.normal_(0, 1)
fake_imgs, attention_maps, _, _ = netG(noise, sent_emb, words_embs, mask)
# G attention
cap_lens_np = cap_lens.cpu().data.numpy()
for j in range(batch_size):
save_name = '%s/%d_s_%d' % (save_dir, i, sorted_indices[j])
for k in range(len(fake_imgs)):
im = fake_imgs[k][j].data.cpu().numpy()
im = (im + 1.0) * 127.5
im = im.astype(np.uint8)
# print('im', im.shape)
im = np.transpose(im, (1, 2, 0))
# print('im', im.shape)
im = Image.fromarray(im)
fullpath = '%s_g%d.png' % (save_name, k)
im.save(fullpath)
for k in range(len(attention_maps)):
if len(fake_imgs) > 1:
im = fake_imgs[k + 1].detach().cpu()
else:
im = fake_imgs[0].detach().cpu()
attn_maps = attention_maps[k]
att_sze = attn_maps.size(2)
img_set, sentences = \
build_super_images2(im[j].unsqueeze(0),
captions[j].unsqueeze(0),
[cap_lens_np[j]], self.ixtoword,
[attn_maps[j]], att_sze)
if img_set is not None:
im = Image.fromarray(img_set)
fullpath = '%s_a%d.png' % (save_name, k)
im.save(fullpath)