-
Notifications
You must be signed in to change notification settings - Fork 3
/
train.py
360 lines (310 loc) · 14.7 KB
/
train.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
#! python3
# -*- coding: utf-8 -*-
"""
################################################################################################
Implementation of 'PROGRESSIVE GROWING OF GANS FOR IMPROVED QUALITY, STABILITY, AND VARIATION'##
https://arxiv.org/pdf/1710.10196.pdf ##
################################################################################################
https://github.com/shanexn
Created: 2018-06-18
################################################################################################
"""
import os
import random
import warnings
from PIL import Image
from model import *
import torch.optim as optim
import torch.backends.cudnn as cudnn
import torchvision.utils as vutils
import torch.utils.data as udata
import torchvision.datasets as vdatasets
import torchvision.transforms as transforms
################################################################################
MANUAL_SEED = 0.5
CUDNN_BENCHMARK = True
TO_GPU = True # For now, all run in GPU mode
DEVICE = torch.device("cuda:0")
G_LR = 0.0001
D_LR = 0.0001
ADAM_BETA = (0.0, 0.99)
ADAM_EPS = 1e-8
LOD_STABLE_IMGS = 2 * 1e5 # Thousands of real images to show before doubling the resolution.
LOD_FADEIN_IMGS = 2 * 1e5 # Thousands of real images to show when fading in new layers.
INIT_RESOLUTION = 4 # Image resolution used at the beginning training.
SAVE_IMG_STEP = 500
LAMBDA_FOR_WGANGP = 10
CRITIC_FOR_WGANGP = 1
MINIBATCH_MAPPING = {
4: 16,
8: 16,
16: 16,
32: 16,
64: 16,
128: 8,
256: 4,
512: 2,
1024: 1
}
#################################################################################
# Construct Generator and Discriminator #########################################
#################################################################################
class PGGAN(object):
def __init__(self,
resolution, # Resolution.
latent_size, # Dimensionality of the latent vectors.
criterion_type="GAN", # ["GAN", "WGAN-GP"]
minibatch_mapping=MINIBATCH_MAPPING, # Batch size of each resolution
rgb_channel=3, # Output channel size, for rgb always 3
fmap_base=2 ** 13, # Overall multiplier for the number of feature maps.
fmap_decay=1.0, # log2 feature map reduction when doubling the resolution.
fmap_max=2 ** 9, # Maximum number of feature maps in any layer.
is_tanh=True,
is_sigmoid=True
):
self.latent_size_ = latent_size
self.rgb_channel_ = rgb_channel
self.fmap_base_ = fmap_base
self.fmap_decay_ = fmap_decay
self.fmap_max_ = fmap_max
self.image_pyramid_ = int(np.log2(resolution)) # max level of the Image Pyramid
self.resolution_ = 2 ** self.image_pyramid_ # correct resolution
self.lod_init_res_ = INIT_RESOLUTION # Image resolution used at the beginning.
self.criterion_type_ = criterion_type
self.is_tanh_ = is_tanh
self.is_sigmoid_ = False if self.criterion_type_ in ["WGAN-GP"] else is_sigmoid
self.gradient_weight_real_ = torch.FloatTensor([-1]).cuda()
self.gradient_weight_fake_ = torch.FloatTensor([1]).cuda()
# Mentioned on page13 of the Paper (Appendix A.1)
# 4x4 - 128x128 -> 16
# 256x256 -> 14
# 512x512 -> 6
# 1024x1024 -> 3
# self.minibatch_mapping_ = {
# 4: 16, 8: 16, 16: 16, 32: 16, 64: 16, 128: 16,
# 256: 16, 512: 6, 1024: 3}
self.minibatch_mapping_ = minibatch_mapping
self._init_network()
def gen_image_by_existed_models(self, model_path, file_name, fake_seed=None):
import re
regx = "(?<=x)\d+"
if isinstance(model_path, (tuple, list)):
g_model_path, d_model_path = model_path
else:
g_model_path = model_path
if not os.path.exists(g_model_path):
raise FileExistsError("Models not exist: {}".format(model_path))
res = re.findall(regx, g_model_path)
if res is None:
raise TypeError("Models is not inappropriate: {}".format(model_path))
res = int(res[0])
# batch_size = self.minibatch_mapping_[res]
batch_size = 1
self.g_net.net_config = [int(np.log2(res))-2, "stable", 1.0]
self.g_net.load_state_dict(torch.load(g_model_path))
# self.d_net.load_state_dict(torch.load(d_model_path))
if fake_seed is None:
fake_seed = torch.randn(batch_size,
self.latent_size_, 1, 1)
fake_work = self.g_net(fake_seed)
vutils.save_image(fake_work.detach(),
file_name, nrow=4, normalize=True, padding=0)
def train(self, dataset_root):
# init settings
random.seed = MANUAL_SEED
torch.manual_seed(MANUAL_SEED)
cudnn.benchmark = CUDNN_BENCHMARK
# self._init_network()
self._init_optimizer()
self._init_criterion()
net_level = 0
net_status = "stable"
net_alpha = 1.0
pyramid_from = int(np.log2(self.lod_init_res_))
pyramid_to = self.image_pyramid_
for cursor_pyramid in range(pyramid_from-1, pyramid_to):
minibatch = self.minibatch_mapping_[2**(cursor_pyramid+1)] # if not in the map, raise error directly
net_level = cursor_pyramid - 1
stable_steps = int(LOD_STABLE_IMGS // minibatch)
fadein_steps = int(LOD_FADEIN_IMGS // minibatch)
current_level_res = 2 ** (net_level + 2)
d_datasets = collect_image_data(dataset_root, minibatch, current_level_res, 2)
fake_seed = torch.randn(minibatch, self.latent_size_, 1, 1).cuda()
# The first step should be keep "stable" status
# The rest steps would run as two phases, "fadein" -> "stable"
if cursor_pyramid == pyramid_from-1:
net_status == "stable"
for cursor_step in range(stable_steps):
self._train(net_level, net_status, net_alpha, minibatch, cursor_step, d_datasets, fake_seed)
else:
net_status = "fadein"
for cursor_step in range(fadein_steps):
net_alpha = 1.0 - (cursor_step + 1) / fadein_steps
self._train(net_level, net_status, net_alpha, minibatch, cursor_step, d_datasets, fake_seed)
for cursor_step in range(stable_steps):
net_alpha = 1.0
self._train(net_level, "stable", net_alpha, minibatch, cursor_step, d_datasets, fake_seed)
torch.save(self.g_net.state_dict(), './output/Gnet_%dx%d.pth' % (current_level_res, current_level_res))
torch.save(self.d_net.state_dict(), './output/Dnet_%dx%d.pth' % (current_level_res, current_level_res))
def _train(self, net_level, net_status, net_alpha, batch_size, cur_step, d_datasets, fake_seed):
# prepare data
current_level_res = 2 ** (net_level + 2)
g_data = fake_seed
work_real = random.choice(d_datasets)[0].cuda()
# init nets
self.g_net.net_config = [net_level, net_status, net_alpha]
self.d_net.net_config = [net_level, net_status, net_alpha]
work_fake = self.g_net(g_data)
d_real = self.d_net(work_real)
d_fake = self.d_net(work_fake.detach())
if self.criterion_type_ == "WGAN-GP":
# update d_net
self.d_net.zero_grad()
# # for i in range(CRITIC_FOR_WGANGP):
mean_real = d_real.mean()
mean_real.backward(self.gradient_weight_real_)
mean_fake = d_fake.mean()
mean_fake.backward(self.gradient_weight_fake_)
gradient_penalty = self._gradient_penalty(work_real.data, work_fake.data, batch_size, current_level_res)
gradient_penalty.backward()
self.d_optim.step()
# update g_net
self.g_net.zero_grad()
d_fake = self.d_net(work_fake)
mean_fake = d_fake.mean()
mean_fake.backward(self.gradient_weight_real_)
self.g_optim.step()
elif self.criterion_type_ == "GAN":
target_real = torch.full((batch_size, 1, 1, 1), 1).cuda()
target_fake = torch.full((batch_size, 1, 1, 1), 0).cuda()
# update d_net
self.d_net.zero_grad()
loss_d_real = self.criterion(d_real, target_real)
loss_d_real.backward()
loss_d_fake = self.criterion(d_fake, target_fake)
loss_d_fake.backward()
# loss_d = loss_d_real + loss_d_fake * 0.1
# loss_d.backward()
self.d_optim.step()
# update g_net
self.g_net.zero_grad()
d_fake = self.d_net(work_fake)
loss_g = self.criterion(d_fake, target_real)
loss_g.backward()
self.g_optim.step()
if cur_step % SAVE_IMG_STEP == 0:
vutils.save_image(work_real,
"./output/real_netlevel%02d.jpg" % net_level,
nrow=4, normalize=True, padding=0)
fake_work = self.g_net(g_data)
vutils.save_image(fake_work.detach(),
'./output/fake_%s_netlevel%02d_%07d.jpg' % (net_status, net_level, cur_step),
nrow=4, normalize=True, padding=0)
def _init_network(self):
# Init Generator and Discriminator
self.g_net = Generator(self.resolution_, self.latent_size_, self.rgb_channel_,
self.fmap_base_, self.fmap_decay_, self.fmap_max_, is_tanh=self.is_tanh_)
self.d_net = Discriminator(self.resolution_, self.rgb_channel_,
self.fmap_base_, self.fmap_decay_, self.fmap_max_, is_sigmoid=self.is_sigmoid_)
# if TO_GPU:
self.g_net.cuda()
self.d_net.cuda()
def _init_optimizer(self):
self.g_optim = optim.Adam(self.g_net.parameters(), lr=G_LR, betas=ADAM_BETA, eps=ADAM_EPS)
self.d_optim = optim.Adam(self.d_net.parameters(), lr=D_LR, betas=ADAM_BETA, eps=ADAM_EPS)
def _init_criterion(self):
if self.criterion_type_ == "GAN":
self.criterion = nn.BCELoss()
elif self.criterion_type_ == "WGAN-GP":
self.criterion = self._gradient_penalty
else:
raise ValueError("INVALID TYPE: {0}.".format(self.criterion_type_))
def _gradient_penalty(self, real_data, fake_data, batch_size, res):
"""
This algorithm was mentioned on the Page4 of paper
'Improved Training of Wasserstein GANs'
This implementation was from 'https://github.com/caogang/wgan-gp'
"""
epsilon = torch.rand(batch_size, 1)
epsilon = epsilon.expand(batch_size, real_data.nelement() / batch_size).contiguous().view(batch_size, 3, res, res)
epsilon = epsilon.cuda()
median_x = epsilon * real_data + ((1 - epsilon) * fake_data)
# if TO_GPU:
median_x = median_x.cuda()
median_data = torch.autograd.Variable(median_x, requires_grad=True)
d_median_data = self.d_net(median_data)
gradients = torch.autograd.grad(outputs=d_median_data, inputs=median_data,
grad_outputs=torch.ones(d_median_data.size()).cuda(),
create_graph=True, retain_graph=True, only_inputs=True)[0]
gradients = gradients.view(gradients.size(0), -1)
gradient_penalty = ((gradients.norm(2, dim=1) - 1) ** 2).mean() * LAMBDA_FOR_WGANGP
return gradient_penalty
#################################################################################
# Image loading and fake data generating ########################################
#################################################################################
# Temp used, will be rewritten
def collect_image_data(dir, batch_size, resolution, num_workers, max_samplesize=150):
path = os.path.join(dir, str(resolution))
dset = vdatasets.ImageFolder(
root=path, transform=transforms.Compose([
transforms.ToTensor(),
transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5)),
]))
dataloader = udata.DataLoader(dset, batch_size=batch_size,
shuffle=True, num_workers=num_workers, drop_last=True)
output = []
for i, j in enumerate(dataloader):
if i == max_samplesize:
break
output.append(j)
return output
def collect_fake_data(tag_dir):
pass
#################################################################################
# Func for reorganising images ##################################################
#################################################################################
def gen_classified_images(dir_path, classes=None, centre_crop=True, save_to_local=False):
"""
dir_path
|_ img1
|_ img2
:return:
"""
classes = [4, 8, 16, 32, 64, 128, 256, 512, 1024] if classes is None else classes
image_dataset = {i: [] for i in classes}
for i in os.listdir(dir_path):
img_path = os.path.join(dir_path, i)
if not os.path.isfile(img_path):
continue
# im = _image_reader(os.path.join(dir_path, i))
im = Image.open(os.path.join(dir_path, i))
im_info = im.info
if im is None:
continue
o_width, o_height = im.size
if centre_crop is True:
min_dim = o_width if o_width < o_height else o_height
im = _image_centre_crop(im, (min_dim, min_dim))
for class_ in classes:
im_resized = _resize_image(im, (class_, class_))
image_dataset[class_].append(im_resized)
if save_to_local is True:
folder = os.path.join(os.path.dirname(dir_path),
"{0}_{1}".format(os.path.basename(dir_path), "classified"),
str(class_), str(class_))
os.makedirs(folder, exist_ok=True)
im_resized.save(os.path.join(folder, i), **im_info)
return image_dataset
def _resize_image(im, size, resample=Image.ANTIALIAS):
return im.resize(size, resample)
def _image_centre_crop(im, outsize):
im_width, im_height = im.size
out_width, out_height = outsize
left = int(round((im_width - out_width) / 2.))
upper = int(round((im_height - out_height) / 2.))
right = left + out_width
lower = upper + out_height
return im.crop((left, upper, right, lower))
if __name__ == "__main__":
p = PGGAN(1024, 512, "WGAN-GP")
p.train(r"E:\workspace\datasets\CelebA\Img\img_align_celeba_classified")