/
train_r.py
455 lines (363 loc) · 16.9 KB
/
train_r.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
"""Script to train an R-separate model on a previously trained
G-LIS (does not have to contain any LIS modules).
Example:
python train_r.py --dataset folder --dataroot /path/to/datasets/celeba \
--crop_size 160 --image_size 80 --code_size 256 --norm weight \
--r_iterations 1 --lr 0.00005 --niter 2500 --spatial_dropout_r 0.1 \
--save_path_r /path/to/checkpoints/r-exp01 \
--load_path /path/to/checkpoints/exp01
Trains an R-separate based on a previous training of a G-LIS with 1 LIS
modules which was saved to /path/to/checkpoints/exp01.
R-separate is trained for 2.5k batches at learning rate 0.00005 and
spatial dropout of 10%. It is saved to /path/to/checkpoints/r-exp01.
"""
from __future__ import print_function
import sys
import os
sys.path.append(os.path.join(os.path.dirname(__file__), '..'))
import argparse
import math
import torch
import torch.nn as nn
import torch.optim as optim
import torchvision
import torchvision.datasets as datasets
import torchvision.transforms as transforms
from torch.autograd import Variable
import os
import os.path
import numpy as np
import imgaug as ia
from scipy import misc
from common import plotting
from common.model import *
parser = argparse.ArgumentParser()
parser.add_argument('--dataset', required = True,
help = 'cifar10 | lsun | imagenet | folder | lfw')
parser.add_argument('--lsun_class', default = 'bedroom',
help = 'class of lsun dataset to use')
parser.add_argument('--dataroot', required = True,
help = 'path to dataset')
parser.add_argument('--batch_size', type = int, default = 32,
help = 'input batch size')
parser.add_argument('--image_size', type = int, default = -1,
help = 'image size')
parser.add_argument('--width', type = int, default = -1,
help = 'image width')
parser.add_argument('--height', type = int, default = -1,
help = 'image height')
parser.add_argument('--crop_size', type = int, default = -1,
help = 'crop size before scaling')
parser.add_argument('--crop_width', type = int, default = -1,
help = 'crop width before scaling')
parser.add_argument('--crop_height', type = int, default = -1,
help = 'crop height before scaling')
parser.add_argument('--code_size', type = int, default = 128,
help = 'size of latent code')
parser.add_argument('--nfeature', type = int, default = 64,
help = 'number of features of first conv layer')
parser.add_argument('--nlayer', type = int, default = -1,
help = 'number of down/up conv layers')
parser.add_argument('--norm', default = 'none',
help = 'type of normalization: none | batch | weight | weight-affine')
parser.add_argument('--load_path', required = True,
help = 'load to continue existing experiment')
parser.add_argument('--save_path_r', default = None,
help = 'path to save generated files')
parser.add_argument('--load_path_r', default = None,
help = 'load to continue existing experiment')
parser.add_argument('--lr', type = float, default = 0.0001,
help = 'learning rate')
parser.add_argument('--test_interval', type = int, default = 1000,
help = 'how often to test reconstruction')
parser.add_argument('--test_lr', type = float, default = 0.01,
help = 'learning rate for reconstruction test')
parser.add_argument('--test_steps', type = int, default = 50,
help = 'number of steps in running reconstruction test')
parser.add_argument('--vis_interval', type = int, default = 100,
help = 'how often to save generated samples')
parser.add_argument('--vis_size', type = int, default = 10,
help = 'size of visualization grid')
parser.add_argument('--vis_row', type = int, default = -1,
help = 'height of visualization grid')
parser.add_argument('--vis_col', type = int, default = -1,
help = 'width of visualization grid')
parser.add_argument('--save_interval', type = int, default = 500,
help = 'how often to save network')
parser.add_argument('--niter', type = int, default = 50000,
help = 'number of iterations to train')
parser.add_argument('--final_test', action = 'store_true', default = False,
help = 'do final test')
parser.add_argument('--ls', action = 'store_true', default = False,
help = 'use LSGAN')
parser.add_argument('--output_scale', action = 'store_true', default = False,
help = 'save x*2-1 instead of x when saving image')
parser.add_argument('--net', default = 'last',
help = 'network prefix to use for loading G')
parser.add_argument('--spatial_dropout_r', type = float, default = 0,
help = 'spatial dropout applied to R')
parser.add_argument('--r_iterations', type = int, default = 3,
help = 'how many LIS modules to use in G')
parser.add_argument('--g_upscaling', default='fractional',
help = 'upscaling method to use in G: fractional|nearest|bilinear')
opt = parser.parse_args()
print(opt)
transform_list = []
if (opt.crop_height > 0) and (opt.crop_width > 0):
transform_list.append(transforms.CenterCrop(opt.crop_height, crop_width))
elif opt.crop_size > 0:
transform_list.append(transforms.CenterCrop(opt.crop_size))
if (opt.height > 0) and (opt.width > 0):
transform_list.append(transforms.Scale(opt.height, opt.width))
elif opt.image_size > 0:
transform_list.append(transforms.Scale(opt.image_size))
transform_list.append(transforms.CenterCrop(opt.image_size))
opt.height = opt.image_size
opt.width = opt.image_size
else:
raise ValueError('must specify valid image size')
transform_list.append(transforms.ToTensor())
if (opt.vis_row <= 0) or (opt.vis_col <= 0):
opt.vis_row = opt.vis_size
opt.vis_col = opt.vis_size
if opt.nlayer < 0:
opt.nlayer = 0
s = max(opt.width, opt.height)
while s >= 8:
s = (s + 1) // 2
opt.nlayer = opt.nlayer + 1
if opt.dataset == 'cifar10':
dataset1 = datasets.CIFAR10(root = opt.dataroot, download = True,
transform = transforms.Compose(transform_list))
dataset2 = datasets.CIFAR10(root = opt.dataroot, train = False,
transform = transforms.Compose(transform_list))
def get_data(k):
if k < len(dataset1):
return dataset1[k][0]
else:
return dataset2[k - len(dataset1)][0]
else:
if opt.dataset in ['imagenet', 'folder', 'lfw']:
dataset = datasets.ImageFolder(root = opt.dataroot,
transform = transforms.Compose(transform_list))
elif opt.dataset == 'lsun':
dataset = datasets.LSUN(db_path = opt.dataroot, classes = [opt.lsun_class + '_train'],
transform = transforms.Compose(transform_list))
def get_data(k):
return dataset[k][0]
data_index = torch.load(os.path.join(opt.dataroot, 'data_index.pt'))
train_index = data_index['train']
if opt.final_test:
test_index = data_index['final_test']
else:
test_index = data_index['running_test']
gen = GeneratorLearnedInputSpace(opt.width, opt.height, opt.nfeature, opt.nlayer, opt.code_size, opt.norm, n_lis_layers=opt.r_iterations, upscaling=opt.g_upscaling)
print(gen)
gen.cuda()
testfunc = nn.MSELoss()
r = build_reverser(opt.width, opt.height, opt.nfeature//2, opt.nlayer, opt.code_size, opt.norm, opt.spatial_dropout_r)
print(r)
r.cuda()
if not opt.final_test:
dis = build_discriminator(opt.width, opt.height, opt.nfeature, opt.nlayer, opt.norm)
print(dis)
dis.cuda()
for param in dis.parameters():
param.requires_grad = False
if opt.ls:
lossfunc = nn.MSELoss()
else:
lossfunc = nn.BCELoss()
lossfunc_r = nn.MSELoss()
r_opt = optim.RMSprop(r.parameters(), lr = opt.lr, eps = 1e-6, alpha = 0.9)
history = plotting.History()
history.add_group("loss-r", ["train"], increasing=False)
history.add_group("loss-stage1", ["train"], increasing=False)
history.add_group("loss-stage2", ["train"], increasing=True)
history.add_group("loss-stage-mix", ["train-stage1", "train-stage2"], increasing=True)
state = {}
def load_state(path, prefix, gen_only = False):
gen.load_state_dict(torch.load(os.path.join(opt.load_path, 'net_archive', '{0}_gen.pt'.format(prefix))))
if not gen_only:
dis.load_state_dict(torch.load(os.path.join(opt.load_path, 'net_archive', '{0}_dis.pt'.format(prefix))))
def load_state_r(path, prefix):
global history
saved_state = torch.load(os.path.join(opt.load_path_r, 'net_archive', '{0}_state.pt'.format(prefix)))
state.update(saved_state)
history = plotting.History.from_string(saved_state['history'])
r_fp = os.path.join(opt.load_path_r, 'net_archive', '{0}_r.pt'.format(prefix))
r_opt_fp = os.path.join(opt.load_path_r, 'net_archive', '{0}_r_opt.pt'.format(prefix))
if os.path.isfile(r_fp):
r.load_state_dict(r_fp)
if os.path.isfile(r_opt_fp):
r_opt.load_state_dict(r_opt_fp)
def save_state(path, prefix):
torch.save(r.state_dict(), os.path.join(opt.save_path_r, 'net_archive', '{0}_r.pt'.format(prefix)))
torch.save(r_opt.state_dict(), os.path.join(opt.save_path_r, 'net_archive', '{0}_r_opt.pt'.format(prefix)))
state.update({
'index_shuffle' : index_shuffle,
'current_iter' : current_iter,
'best_iter' : best_iter,
'min_loss' : min_loss,
'current_sample' : current_sample,
'history' : history.to_string()
})
torch.save(state, os.path.join(opt.save_path_r, 'net_archive', '{0}_state.pt'.format(prefix)))
loss_plotter = plotting.LossPlotter(
history.get_group_names(),
history.get_groups_increasing(),
save_to_fp=os.path.join(opt.save_path_r, "loss.jpg")
)
loss_plotter.start_batch_idx = 100
def visualize(code, filename, filename_r, filename_all):
gen.eval()
generated = torch.Tensor(code.size(0), 3, opt.height, opt.width)
generated_r = [torch.Tensor(code.size(0), 3, opt.height, opt.width) for _ in range(opt.r_iterations)]
generated_all = []
for i in range((code.size(0) - 1) // opt.batch_size + 1):
batch_size = min(opt.batch_size, code.size(0) - i * opt.batch_size)
batch_code = Variable(code[i * opt.batch_size : i * opt.batch_size + batch_size])
generated_images, _ = gen(batch_code, n_execute_lis_layers=0)
generated[i * opt.batch_size : i * opt.batch_size + batch_size].copy_(generated_images.data)
generated_images_r_by_iter = []
for r_iter in range(opt.r_iterations):
generated_images_r, _ = gen(batch_code, n_execute_lis_layers=r_iter+1)
generated_r[r_iter][i * opt.batch_size : i * opt.batch_size + batch_size].copy_(generated_images_r.data)
generated_images_r_by_iter.append(generated_images_r)
for imgidx in range(generated_images.size(0)):
image = generated_images[imgidx].data.cpu().numpy()
image = (image * 255).astype(np.uint8).transpose((1, 2, 0))
generated_all.append(image)
for r_iter in range(opt.r_iterations):
image_r = generated_images_r_by_iter[r_iter][imgidx].data.cpu().numpy()
image_r = (image_r * 255).astype(np.uint8).transpose((1, 2, 0))
generated_all.append(image_r)
if opt.output_scale:
torchvision.utils.save_image(generated * 2 - 1, filename, opt.vis_row)
for r_iter in range(opt.r_iterations):
torchvision.utils.save_image(generated_r[r_iter] * 2 - 1, filename_r.format(r_iter), opt.vis_row)
else:
torchvision.utils.save_image(generated, filename, opt.vis_row)
for r_iter in range(opt.r_iterations):
torchvision.utils.save_image(generated_r[r_iter], filename_r.format(r_iter), opt.vis_row)
misc.imsave(filename_all, ia.draw_grid(generated_all, cols=opt.vis_col*(1+opt.r_iterations)))
gen.train()
def visualize(code, filename, filename_r, filename_both):
gen.eval()
r.eval()
generated = torch.Tensor(code.size(0), 3, opt.height, opt.width)
generated_r = torch.Tensor(code.size(0), 3, opt.height, opt.width)
generated_both = []
for i in range((code.size(0) - 1) // opt.batch_size + 1):
batch_size = min(opt.batch_size, code.size(0) - i * opt.batch_size)
batch_code = Variable(code[i * opt.batch_size : i * opt.batch_size + batch_size])
generated_images, _ = gen(batch_code, n_execute_lis_layers=opt.r_iterations)
generated[i * opt.batch_size : i * opt.batch_size + batch_size].copy_(generated_images.data)
code_reversed = r(generated_images)
generated_images_r, _ = gen(code_reversed, n_execute_lis_layers=opt.r_iterations)
generated_r[i * opt.batch_size : i * opt.batch_size + batch_size].copy_(generated_images_r.data)
for j in range(generated_images.size(0)):
image = generated_images[j].data.cpu().numpy()
image_r = generated_images_r[j].data.cpu().numpy()
image = (image * 255).astype(np.uint8).transpose((1, 2, 0))
image_r = (image_r * 255).astype(np.uint8).transpose((1, 2, 0))
generated_both.append(image)
generated_both.append(image_r)
if opt.output_scale:
torchvision.utils.save_image(generated * 2 - 1, filename, opt.vis_row)
torchvision.utils.save_image(generated_r * 2 - 1, filename_r, opt.vis_row)
else:
torchvision.utils.save_image(generated, filename, opt.vis_row)
torchvision.utils.save_image(generated_r, filename_r, opt.vis_row)
misc.imsave(filename_both, ia.draw_grid(generated_both, cols=opt.vis_col*2))
gen.train()
r.train()
def makedirs():
if not os.path.exists(opt.save_path_r):
os.makedirs(opt.save_path_r)
for sub_folder in ('samples', 'samples_r', 'samples_both', 'running_test', 'net_archive', 'log'):
if not os.path.exists(os.path.join(opt.save_path_r, sub_folder)):
os.mkdir(os.path.join(opt.save_path_r, sub_folder))
load_state(opt.load_path, opt.net)
if opt.load_path_r is not None:
if opt.save_path_r is None:
opt.save_path_r = opt.load_path_r
if opt.load_path_r != opt.save_path_r:
makedirs()
vis_code = torch.load(os.path.join(opt.load_path_r, 'samples', 'vis_code.pt')).cuda()
load_state_r(opt.load_path_r, 'last')
index_shuffle = state['index_shuffle']
current_iter = state['current_iter']
best_iter = state['best_iter']
min_loss = state['min_loss']
current_sample = state['current_sample']
else:
if opt.save_path_r is None:
raise ValueError('must specify save path if not continue training')
makedirs()
vis_code = torch.randn(opt.vis_row * opt.vis_col, opt.code_size).cuda()
torch.save(vis_code, os.path.join(opt.save_path_r, 'samples', 'vis_code.pt'))
index_shuffle = torch.randperm(train_index.size(0))
current_iter = 0
best_iter = 0
min_loss = 1e100
current_sample = 0
vis_target = torch.Tensor(min(test_index.size(0), opt.vis_row * opt.vis_col), 3, opt.height, opt.width)
for i in range(vis_target.size(0)):
vis_target[i].copy_(get_data(test_index[i]))
if opt.output_scale:
torchvision.utils.save_image(vis_target * 2 - 1, os.path.join(opt.save_path_r, 'running_test', 'target.jpg'), opt.vis_row)
else:
torchvision.utils.save_image(vis_target, os.path.join(opt.save_path_r, 'running_test', 'target.jpg'), opt.vis_row)
ones = Variable(torch.ones(opt.batch_size, 1).cuda())
zeros = Variable(torch.zeros(opt.batch_size, 1).cuda())
loss_record = torch.zeros(opt.test_interval, 3)
visualize(
vis_code,
filename=os.path.join(opt.save_path_r, 'samples', 'sample_{0}.jpg'.format(current_iter)),
filename_r=os.path.join(opt.save_path_r, 'samples_r', 'sample_{0}_r.jpg'.format(current_iter)),
filename_both=os.path.join(opt.save_path_r, 'samples_both', 'sample_{0}_both.jpg'.format(current_iter))
)
# train for --niter batches
while current_iter < opt.niter:
current_iter = current_iter + 1
current_loss_record = loss_record[(current_iter - 1) % opt.test_interval]
rand_code = Variable(torch.randn(opt.batch_size, opt.code_size).cuda())
generated, _ = gen(rand_code, n_execute_lis_layers=opt.r_iterations)
# loss D orig samples
loss = lossfunc(
dis(Variable(generated.data, volatile=True)),
zeros
)
current_loss_record[0] = loss.data[0]
generated.detach()
r.zero_grad()
# train R
rand_code_fixed = r(generated)
loss = lossfunc_r(rand_code_fixed, rand_code)
current_loss_record[1] = loss.data[0]
loss.backward()
r_opt.step()
# loss D error-fixed samples
generated_fixed, _ = gen(rand_code_fixed, n_execute_lis_layers=opt.r_iterations)
loss = lossfunc(
dis(Variable(generated_fixed.data, volatile=True)),
zeros
)
current_loss_record[2] = loss.data[0]
history.add_value("loss-r", "train", current_iter, current_loss_record[1], average=False)
history.add_value("loss-stage1", "train", current_iter, current_loss_record[0], average=False)
history.add_value("loss-stage2", "train", current_iter, current_loss_record[2], average=False)
history.add_value("loss-stage-mix", "train-stage1", current_iter, current_loss_record[0], average=False)
history.add_value("loss-stage-mix", "train-stage2", current_iter, current_loss_record[2], average=False)
print('{0} | loss: r:{1} dis-stage1:{2} dis-stage2:{3}'.format(current_iter, current_loss_record[1], current_loss_record[0], current_loss_record[2]))
if current_iter % opt.vis_interval == 0:
visualize(
vis_code,
filename=os.path.join(opt.save_path_r, 'samples', 'sample_{0}.jpg'.format(current_iter)),
filename_r=os.path.join(opt.save_path_r, 'samples_r', 'sample_{0}_r.jpg'.format(current_iter)),
filename_both=os.path.join(opt.save_path_r, 'samples_both', 'sample_{0}_both.jpg'.format(current_iter))
)
loss_plotter.plot(history)
if current_iter % opt.save_interval == 0:
save_state(opt.save_path_r, current_iter)