forked from JinshuChen/MOGAN
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathrun.py
495 lines (452 loc) · 21.6 KB
/
run.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
#!/usr/bin/env python
# coding: utf-8
# import torch.nn.functional as F
# import torch.optim as optim
# from apex import amp
# from tqdm import tqdm
# from skimage import io as img
# import torchvision.models as models
# from torchsummaryX import summary as modelsummary
# import nni
import time
import os
import warnings
import shutil
import argparse
import logging
import math
import torch
import torch.nn as nn
import numpy as np
import scipy.misc
import matplotlib
import matplotlib.pyplot as plt
import colorama
from utils_functions import ran_gen, seconds2time, get_logger, data_augmenter, plot_sinloss, make_padder, get_reals, Generate_noise, convert_image_np, draw_concat, calc_gradient_penalty, reset_grads, modify_scales, get_seg, calc_local_rec
from training_functions import organise_models, recon_loss, VGGPerceptualLoss
from configs import get_configs, LogbookFormatter
from progress_bar import create_progressbar
from evaluation import calculate_sifid_given_paths, calculate_cs
warnings.filterwarnings("ignore")
matplotlib.use('Agg')
# np.set_printoptions(threshold = np.inf)
# @profile
def main():
# params = nni.get_next_parameter()
parser = argparse.ArgumentParser()
parser.add_argument('--mode', type=str, required=True)
parser.add_argument('--config_file', type=str, required=True)
parser.add_argument('--log', type=bool, default=True)
parser.add_argument('--logbook', type=str, default='log.txt')
logger = get_logger()
parser = get_configs(parser, logger)
opts = parser.parse_args()
red = colorama.Fore.RED
green = colorama.Fore.GREEN
white = colorama.Fore.WHITE
cyan = colorama.Fore.CYAN
reset = colorama.Style.RESET_ALL
bright = colorama.Style.BRIGHT
dim = colorama.Style.DIM
seed = opts.seed
# seed = params['seed']
np.random.seed(seed)
torch.manual_seed(seed)
torch.cuda.manual_seed_all(seed)
img_path = opts.img_input_dir
anot_path = opts.anot_input_dir
output_dir = opts.output_dir
local_time = time.strftime("%m%d_%H%M%S", time.localtime())
dir2save = '%s/%s_seed%s/' % (output_dir, local_time, seed)
gen_num = opts.gen_num
dir2gen = '%sran_gen/' % (dir2save)
try:
os.makedirs(dir2save)
os.makedirs(dir2gen)
shutil.copyfile('%s/%s' % (os.getcwd(), opts.config_file), '%s/%s' % (dir2save, opts.config_file))
shutil.copyfile('%s/my_models.py' % os.getcwd(), '%s/my_models.py' % dir2save)
shutil.copyfile('%s/run.py' % os.getcwd(), '%s/run.py' % dir2save)
shutil.copyfile('%s/training_functions.py' % os.getcwd(), '%s/training_functions.py' % dir2save)
shutil.copyfile('%s/utils_functions.py' % os.getcwd(), '%s/utils_functions.py' % dir2save)
except OSError:
raise Exception("Files ERROR!")
if opts.log:
logbook = opts.logbook
logpath = dir2save + logbook
loghandler = logging.FileHandler(filename=logpath, mode="a", encoding="utf-8")
loghandler.setLevel(logging.INFO)
logbook_formatter = LogbookFormatter(fmt="[%(asctime)s] %(message)s", datefmt="%Y-%m-%d %H:%M:%S")
loghandler.setFormatter(logbook_formatter)
logger.addHandler(loghandler)
### can't Loading Weights at present
weights2load = 0
G_weights2load = ''
D_weights2load = ''
###
Gs = []
Zs = []
Ds = []
reals = []
masks = []
noises = []
NoiseWeight = []
errD2plot = []
errG2plot = []
mode = opts.mode
channels = opts.channels
kernel_size = opts.kernel_size
stride = opts.stride
if_padding = opts.if_padding
if_lazy = opts.if_lazy
G_num_layer = opts.G_num_layer
D_num_layer = opts.D_num_layer
if mode == 'f':
weight4style = opts.weight4style
scale_base = opts.scale_base
# scales = opts.scales
scales = modify_scales(anot_path, scale_base)
logger.info('-' * 80)
logger.info(green + '[INFO]: scales are set to %s' % scales + reset)
out_channels = opts.out_channels
lr_g = opts.lr_g
lr_d = opts.lr_d
iters_list = [int(i) for i in opts.iters_list]
D_steps = opts.D_steps
G_steps = opts.G_steps
lambda_grad = opts.lambda_grad
n_segments = opts.n_segments
compactness = opts.compactness
sigma = opts.sigma
start_label = opts.start_label
device = torch.device("cuda:0")
alpha4rec_ini = opts.alpha4rec
alpha4cos_ini = opts.alpha4cos
alpha4vgg_ini = opts.alpha4vgg
p_loss = VGGPerceptualLoss(resize=False, device=device) if alpha4vgg_ini != 0 else 0
###
# factor4rec = calc_factor('rec', scales)
# factor4cos = calc_factor('cos', scales)
# factor4vgg = calc_factor('vgg', scales)
###
noise_weight = opts.noise_weight
noise_weight_ini = noise_weight
p4flip = opts.p4flip
torch.backends.cudnn.benchmark = True
# amp.register_float_function(torch, 'sigmoid')
reals, masks = get_reals(mode, img_path, anot_path, scales, scale_base, reals, channels, masks)
reals, masks = reals[::-1], masks[::-1]
reals_b, reals_fa, masks_b, masks_f = [], [], [], []
for _ in reals:
reals_b.append(_[1])
reals_fa.append(_[3])
for _ in masks:
masks_b.append(_[0])
masks_f.append(_[1])
for scale_num in range(scales):
outfile_dir = '%s%s/' % (dir2save, scale_num)
try:
os.makedirs(outfile_dir)
except OSError:
raise Exception("Files ERROR!")
_, __, ___, ____, _____ = reals[scale_num][0], reals[scale_num][1], reals[scale_num][2], reals[scale_num][3], masks_f[scale_num]
plt.imsave('%s/real_original.png' % (outfile_dir), convert_image_np(_), vmin=0, vmax=1)
plt.imsave('%s/real_background.png' % (outfile_dir), convert_image_np(__), vmin=0, vmax=1)
plt.imsave('%s/real_foregrounds.png' % (outfile_dir), convert_image_np(___), vmin=0, vmax=1)
plt.imsave('%s/real_foreground_a.png' % (outfile_dir), convert_image_np(____), vmin=0, vmax=1)
scipy.misc.toimage(convert_image_np(_____[:, 0, :, :][None, :, :, :])).save('%s/mask_f.png' % (outfile_dir))
torch.save(reals_fa, dir2save+'reals_f.pth')
torch.save(reals_b, dir2save+'reals_b.pth')
torch.save(masks_f, dir2save+'masks_f.pth')
logger.info('-' * 80)
logger.info(green + '[INFO]: data prepared!' + reset)
logger.info('-' * 80)
torch.cuda.synchronize()
start_time = time.time()
logger.info(green + '[INFO]: training starts at %s' % time.strftime("%H:%M:%S", time.localtime()) + reset)
logger.info('-' * 80)
for scale_num in range(scales):
iters = iters_list[scale_num]
outfile_dir = '%s%s/' % (dir2save, scale_num)
real_curr = reals[scale_num]
x = np.random.choice(iters, int(iters*p4flip), replace=False)
# real_seg = get_seg(real_curr[3], n_segments=n_segments, compactness=compactness, sigma=sigma, start_label=start_label)
zeros = torch.zeros_like(real_curr[3]).to(device)
edge_w, edge_h = math.ceil(0.1*real_curr[3].shape[3]), math.ceil(0.1*real_curr[3].shape[2])
for i in range(edge_w):
zeros[:,:,:,i] = 1.
for i in range(real_curr[3].shape[3]-edge_w, real_curr[3].shape[3]):
zeros[:,:,:,i] = 1.
for i in range(edge_h):
zeros[:,:,i,:] = 1.
for i in range(real_curr[3].shape[2]-edge_h, real_curr[3].shape[2]):
zeros[:,:,i,:] = 1.
assert zeros[0,0,0,0] == 1
if mode == 'f':
alpha4cos = alpha4cos_ini
if scale_num >= scales: # 4 5
alpha4rec = alpha4rec_ini * 10
else: # 0 1 2 3
alpha4rec = alpha4rec_ini
real_curr[3] = real_curr[3].to(device)
h, w = real_curr[3].shape[2], real_curr[3].shape[3]
D, G, optimizerD, optimizerG, schedulerD, schedulerG = organise_models(
mode, device, weights2load, lr_g, lr_d, channels, kernel_size, stride, if_padding,
G_num_layer, D_num_layer, out_channels, factor=0.01+weight4style*(scales-scale_num-1)/scales
)
elif mode == 'b':
# if scale_num <= 0:
# lr_g = 0.0001
# lr_d = 0.0001
alpha4rec = alpha4rec_ini
alpha4cos = alpha4cos_ini
real_curr[1] = real_curr[1].to(device)
h, w = real_curr[1].shape[2], real_curr[1].shape[3]
D, G, optimizerD, optimizerG, schedulerD, schedulerG = organise_models(
mode, device, weights2load, lr_g, lr_d, channels, kernel_size, stride, if_padding, G_num_layer, D_num_layer, out_channels
)
# [D, G], [optimizerD, optimizerG] = amp.initialize([D, G], [optimizerD, optimizerG], opt_level='O1', num_losses=14)
# p_loss = 0
#p_loss = p_loss.to(device)
r_loss = recon_loss(False)
r_loss = r_loss.to(device)
if if_padding:
padder = make_padder(0)
else:
padder = make_padder((G_num_layer-1)*1+2+1)
# if opts.ani==True:
# fpadder = make_padder(0)
# h_f = h_f + (1+2+G_num_layer*2)*2
# w_f = w_f + (1+2+G_num_layer*2)*2
noise_1 = padder(Generate_noise([channels, h, w], device=device, if_0=True, if_c_same=False))
epoch_iterator = create_progressbar(
iterable=range(iters),
desc="Training scale [{}/{}]".format(scale_num, scales-1),
offset=0, leave=True, logging_on_update=False, logging_on_close=True, postfix=True
)
for i in epoch_iterator:
epoch_iterator.set_description('Scale [{}/{}], Iteration [{}/{}]'.format(scale_num+1, scales, i+1, iters))
if mode == 'f':
if i >= 1600 and scale_num > 0:
alpha4rec = alpha4rec_ini
styles_ref = []
_tmp = real_curr[3].squeeze(0).cpu()
for cnt in range(G_num_layer*2+2):
if if_padding:
_padder = make_padder(0)
else:
_padder = make_padder(2+2*3+(G_num_layer-1-cnt)*1)
_augment = data_augmenter(_tmp, device=device)
_augment_ = _padder(_augment)
styles_ref.append(_augment_.detach())
del _augment, _augment_
if Gs == []:
noise_1 = padder(Generate_noise([1, h, w], device=device, if_0=False, if_c_same=True))
noise_2 = padder(Generate_noise([1, h, w], device=device, if_0=False, if_c_same=True))
# noise_2_f = padder(get_slerp_interp([1, h_f, w_f], device=device, iters=iters, iter_curr=i, if_c_same=True, start=noise_2_f_s, end=noise_2_f_e))
else:
noise_2 = padder(Generate_noise([channels, h, w], device=device, if_0=False, if_c_same=False))
# noise_2_f = padder(get_slerp_interp([channels, h_f, w_f], device=device, iters=iters, iter_curr=i, if_c_same=False, start=noise_2_f_s, end=noise_2_f_e))
for j in range(D_steps):
if (j == 0) & (i == 0):
if Gs == []:
noise_3 = padder(Generate_noise([channels, h, w], device=device, if_0=True, if_c_same=False))
prev = torch.full([1, channels, h, w], 0, device=device)
_ = prev
prev = padder(prev)
noise_weight = 1
else:
criterion = nn.MSELoss()
if mode == 'f':
prev = padder(draw_concat(Gs, Zs, reals_fa, NoiseWeight, _, 'rand', kernel_size, channels, device, padder, G_num_layer, mode))
noise_3 = draw_concat(Gs, Zs, reals_fa, NoiseWeight, _, 'rec', kernel_size, channels, device, padder, G_num_layer, mode)
RMSE = torch.sqrt(criterion(real_curr[3], noise_3))
elif mode == 'b':
prev = padder(draw_concat(Gs, Zs, reals_b, NoiseWeight, _, 'rand', kernel_size, channels, device, padder, G_num_layer, mode))
noise_3 = draw_concat(Gs, Zs, reals_b, NoiseWeight, _, 'rec', kernel_size, channels, device, padder, G_num_layer, mode)
RMSE = torch.sqrt(criterion(real_curr[1], noise_3))
noise_weight = noise_weight_ini*RMSE
noise_3 = padder(noise_3)
else:
if mode == 'f':
prev = padder(draw_concat(Gs, Zs, reals_fa, NoiseWeight, _, 'rand', kernel_size, channels, device, padder, G_num_layer, mode))
elif mode == 'b':
prev = padder(draw_concat(Gs, Zs, reals_b, NoiseWeight, _, 'rand', kernel_size, channels, device, padder, G_num_layer, mode))
if Gs == []:
noise = noise_2
else:
noise = noise_weight * noise_2 + prev
D.zero_grad()
if mode == 'f':
output = D(real_curr[3])
elif mode == 'b':
output = D(real_curr[1])
if i in x:
errD_real = output.mean()
else:
errD_real = -output.mean()
# with amp.scale_loss(errD_real, optimizerD, loss_id=0) as errD_real:
# errD_real.backward(retain_graph=True)
errD_real.backward(retain_graph=True)
if i in x:
errD_real = -errD_real
if mode == 'f':
fake = G(noise.detach(), styles_ref, prev)
elif mode == 'b':
fake = G(noise.detach(), prev)
output = D(fake.detach())
if i in x:
errD_fake = -output.mean()
else:
errD_fake = output.mean()
# with amp.scale_loss(errD_fake, optimizerD, loss_id=1) as errD_fake:
# errD_fake.backward(retain_graph=True)
errD_fake.backward(retain_graph=True)
if i in x:
errD_fake = -errD_fake
if mode == 'f':
gradient_penalty = calc_gradient_penalty(D, real_curr[3], fake, lambda_grad, device)
elif mode == 'b':
gradient_penalty = calc_gradient_penalty(D, real_curr[1], fake, lambda_grad, device)
# with amp.scale_loss(gradient_penalty, optimizerD, loss_id=2) as gradient_penalty:
# gradient_penalty.backward()
gradient_penalty.backward()
optimizerD.step()
D.zero_grad()
optimizerD.zero_grad()
_errD_real = errD_real.item()
_errD_fake = errD_fake.item()
_gradient_penalty = gradient_penalty.item()
del errD_real, errD_fake, gradient_penalty
_errD = _errD_real + _errD_fake + _gradient_penalty
errD2plot.append([_errD_real, _errD_fake, _gradient_penalty])
schedulerD.step(_errD)
for j in range(G_steps):
G.zero_grad()
###
output = D(fake)
###
errG = -output.mean()
# with amp.scale_loss(errG, optimizerG, loss_id=3) as errG:
# errG.backward(retain_graph=True)
errG.backward(retain_graph=True)
Z_opt = noise_weight * noise_1 + noise_3
if mode == 'f':
_tmp = G(Z_opt.detach(), styles_ref, noise_3)
elif mode == 'b':
_tmp = G(Z_opt.detach(), noise_3)
if alpha4rec != 0:
# loss = r_loss
loss = nn.L1Loss()
Z_opt = noise_weight * noise_1 + noise_3
if mode == 'f':
_loss = loss(_tmp*zeros, real_curr[3]*zeros)
# _loss = calc_local_rec(loss, _tmp, real_seg)
elif mode == 'b':
_loss = loss(_tmp, real_curr[1])
rec_loss = alpha4rec * _loss
del _loss
# with amp.scale_loss(rec_loss, optimizerG, loss_id=4) as rec_loss:
# rec_loss.backward(retain_graph=True)
rec_loss.backward(retain_graph=True)
rec_loss = rec_loss.detach()
else:
Z_opt = noise_1
rec_loss = torch.Tensor([0])
if alpha4cos != 0:
loss = nn.CosineEmbeddingLoss()
Z_opt = noise_weight * noise_1 + noise_3
if mode == 'f':
_loss = loss(_tmp, real_curr[3], torch.ones_like(real_curr[3]))
elif mode == 'b':
_loss = loss(_tmp, real_curr[1], torch.ones_like(real_curr[1]))
cos_loss = alpha4cos * _loss
del _loss
# with amp.scale_loss(cos_loss, optimizerG, loss_id=5) as cos_loss:
# cos_loss.backward(retain_graph=True)
cos_loss.backward(retain_graph=True)
cos_loss = cos_loss.detach()
else:
Z_opt = noise_1
cos_loss = torch.Tensor([0])
if alpha4vgg_ini != 0:
loss = p_loss
Z_opt = noise_weight * noise_1 + noise_3
if mode == 'f':
# _loss = alpha4vgg_ini * loss(_tmp, real_curr[3], device)
_loss = loss(_tmp, real_curr[3], device)
elif mode == 'b':
_loss = alpha4vgg_ini * loss(_tmp, real_curr[1], device)
perceptual_loss = _loss
# perceptual_loss1 = _loss1
# perceptual_loss2 = _loss2
del _loss
# perceptual_loss = factor4vgg[scale_num] * alpha4vgg * p_loss(G(Z_opt.detach(), styles_ref, noise_3), real_curr[3], device)
# perceptual_loss = factor4vgg[scale_num] * alpha4vgg * p_loss(G(Z_opt.detach(), noise_3), real_curr[1], device)
# with amp.scale_loss(perceptual_loss_f, optimizerG_f, loss_id=6) as perceptual_loss_f:
# perceptual_loss_f.backward(retain_graph=True)
# with amp.scale_loss(perceptual_loss, optimizerG, loss_id=5) as perceptual_loss:
# perceptual_loss.backward(retain_graph=True)
# perceptual_loss1.backward(retain_graph=True)
perceptual_loss.backward(retain_graph=True)
# perceptual_loss = perceptual_loss1.detach() + perceptual_loss2.detach()
perceptual_loss = perceptual_loss.detach()
else:
Z_opt = noise_1
perceptual_loss = torch.Tensor([0])
optimizerG.step()
G.zero_grad()
optimizerG.zero_grad()
_errG = errG.item()
_rec_loss = rec_loss.item()
_cos_loss = cos_loss.item()
_perceptual_loss = perceptual_loss.item()
del errG, rec_loss, cos_loss, perceptual_loss
errG2plot.append([_errG, _rec_loss, _cos_loss, _perceptual_loss])
_errG = _errG + _rec_loss + _cos_loss + _perceptual_loss
schedulerG.step(_errG)
del noise_2
if i % 200 == 0 or i == (iters-1):
if mode == 'b':
_fake = fake.cpu()
_fake = _fake * masks_b[scale_num]
_fake = _fake + masks_b[scale_num] - torch.ones_like(masks_b[scale_num])
plt.imsave('%s/fake_%s_%s.png' % (outfile_dir, mode, str(i)), convert_image_np(_fake.detach()), vmin=0, vmax=1)
elif mode == 'f':
plt.imsave('%s/fake_%s_%s.png' % (outfile_dir, mode, str(i)), convert_image_np(fake.detach()), vmin=0, vmax=1)
if i % 500 == 0 or i == (iters-1):
plot_sinloss(errG2plot, errD2plot, scale_num, iters_list, outfile_dir, mode, i)
epoch_iterator.close()
torch.save(G.state_dict(), '%s/G_%s.pth' % (outfile_dir, mode))
torch.save(D.state_dict(), '%s/D_%s.pth' % (outfile_dir, mode))
G = reset_grads(G, False)
G.eval()
D = reset_grads(D, False)
D.eval()
Gs.append(G)
Ds.append(D)
NoiseWeight.append(noise_weight)
Zs.append(noise_1)
# torch.save(Gs, '%s/Gs.pth' % (dir2save))
torch.save(Zs, '%s/Zs.pth' % (dir2save))
torch.save(NoiseWeight, '%s/noiseweight_%s.pth' % (dir2save, mode))
del D, G
torch.cuda.synchronize()
end_time = time.time()
logger.info('-' * 80)
logger.info(green + '[INFO]: training time cost : %s' % seconds2time(end_time - start_time) + reset)
logger.info('-' * 80)
logger.info(green + '[INFO]: randomly generating %s samples...' %(opts.gen_num) + reset)
logger.info('-' * 80)
if mode == 'f':
ran_gen(Gs, Zs, NoiseWeight, reals_fa, opts, dir2gen, padder)
elif mode == 'b':
ran_gen(Gs, Zs, NoiseWeight, reals_b, opts, dir2gen, padder)
logger.info('-' * 80)
logger.info(green + '[INFO]: calculating eval metrics...' + reset)
logger.info('-' * 80)
sifid = calculate_sifid_given_paths(dir2gen+'real.png', dir2gen, batch_size=1, dims=64, suffix='png')
diversity = calculate_cs(dir2gen, suffix='png')
logger.info(green + '[INFO]: SIFID : %6f DIVERSITY : %6f GQI : %6f ' % (sifid, diversity, diversity/sifid)+ reset)
# nni.report_final_result(diversity/sifid)
if __name__ == "__main__":
main()