-
Notifications
You must be signed in to change notification settings - Fork 0
/
train.py
515 lines (414 loc) · 19 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
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
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
from __future__ import unicode_literals
import random
import numpy as np
from collections import OrderedDict
import torch
from datasets.pascalvoc import PascalVOC
import generators.deeplabv2 as deeplabv2
import discriminators.discriminator as dis
from torchvision import transforms
from torch.autograd import Variable
from torch.utils.data import DataLoader
from utils.transforms import RandomSizedCrop, IgnoreLabelClass, ToTensorLabel, NormalizeOwn,ZeroPadding, OneHotEncode
from utils.lr_scheduling import poly_lr_scheduler
import torch.nn.functional as F
import torch.nn as nn
from functools import reduce
import torch.optim as optim
import os
import argparse
from torchvision.transforms import ToTensor,Compose
from utils.validate import val
from utils.helpers import pascal_palette_invert
import torchvision.transforms as transforms
import PIL.Image as Image
from discriminators.discriminator import Dis
import torch.utils.model_zoo as model_zoo
home_dir = os.path.dirname(os.path.realpath(__file__))
def parse_args():
# Parse arguments
parser = argparse.ArgumentParser()
parser.add_argument("prefix",
help="Prefix to identify current experiment")
parser.add_argument("dataset_dir",
help="A directory containing img (Images) and cls (GT Segmentation) folder")
parser.add_argument("--mode", choices=('base','adv','semi'),default='base',
help="base (baseline),adv (adversarial), semi (semi-supervised)")
parser.add_argument("--lam_adv",default=0.01,
help="Weight for Adversarial loss for Segmentation Network training")
parser.add_argument("--lam_semi",default=0.1,
help="Weight for Semi-supervised loss")
parser.add_argument("--t_semi",default=0.1,type=float,
help="Threshold for self-taught learning")
parser.add_argument("--nogpu",action='store_true',
help="Train only on cpus. Helpful for debugging")
parser.add_argument("--max_epoch",default=20,type=int,
help="Maximum iterations.")
parser.add_argument("--start_epoch",default=1,type=int,
help="Resume training from this epoch")
parser.add_argument("--snapshot",
help="Snapshot to resume training")
parser.add_argument("--snapshot_dir",default=os.path.join(home_dir,'data','snapshots'),
help="Location to store the snapshot")
parser.add_argument("--batch_size",default=10,type=int,
help="Batch size for training")
parser.add_argument("--val_orig",action='store_true',
help="Do Inference on original size image. Otherwise, crop to 321x321 like in training ")
parser.add_argument("--d_label_smooth",default=0.1,type=float,
help="Label smoothing for real images in Discriminator")
parser.add_argument("--d_optim",choices=('sgd','adam'),default='sgd',
help="Discriminator Optimizer.")
parser.add_argument("--no_norm",action='store_true',
help="No Normalizaion on the Images")
parser.add_argument("--init_net",choices=('imagenet','mscoco'),default='mscoco',
help="Pretrained Net for Segmentation Network")
parser.add_argument("--d_lr",default=0.0001,type=float,
help="lr for discriminator")
parser.add_argument("--g_lr",default=0.00025,type=float,
help="lr for generator")
parser.add_argument("--seed",default=1,type=int,
help="Seed for random numbers used in semi-supervised training")
parser.add_argument("--wait_semi",default=0,type=int,
help="Number of Epochs to wait before using semi-supervised loss")
parser.add_argument("--split",default=1,type=float)
args = parser.parse_args()
return args
'''
Snapshot the Best Model
'''
def snapshot(model,valoader,epoch,best_miou,snapshot_dir,prefix):
miou = val(model,valoader)
snapshot = {
'epoch': epoch,
'state_dict': model.state_dict(),
'miou': miou
}
if miou > best_miou:
best_miou = miou
torch.save(snapshot,os.path.join(snapshot_dir,'{}.pth.tar'.format(prefix)))
print("[{}] Curr mIoU: {:0.4f} Best mIoU: {}".format(epoch,miou,best_miou))
return best_miou
'''
Use PreTrained Model for Initial Weights
'''
def init_weights(model,init_net):
if init_net == 'imagenet':
# Pretrain on ImageNet
inet_weights = model_zoo.load_url('https://download.pytorch.org/models/resnet101-5d3b4d8f.pth')
del inet_weights['fc.weight']
del inet_weights['fc.bias']
state = model.state_dict()
state.update(inet_weights)
model.load_state_dict(state)
elif init_net == 'mscoco':
# TODO: Upload the weights somewhere to use load.url()
filename = os.path.join(home_dir,'data','MS_DeepLab_resnet_pretrained_COCO_init.pth')
assert(os.path.isfile(filename))
saved_net = torch.load(filename)
new_state = model.state_dict()
saved_net = {k.partition('Scale.')[2]: v for i, (k,v) in enumerate(saved_net.items())}
new_state.update(saved_net)
model.load_state_dict(new_state)
'''
Baseline Training
'''
def train_base(generator,optimG,trainloader,valoader,args):
best_miou = -1
for epoch in range(args.start_epoch,args.max_epoch+1):
generator.train()
for batch_id, (img,mask,_) in enumerate(trainloader):
if args.nogpu:
img,mask = Variable(img),Variable(mask)
else:
img,mask = Variable(img.cuda()),Variable(mask.cuda())
itr = len(trainloader)*(epoch-1) + batch_id
cprob = generator(img)
cprob = nn.LogSoftmax()(cprob)
Lseg = nn.NLLLoss2d()(cprob,mask)
optimG = poly_lr_scheduler(optimG, args.g_lr, itr)
optimG.zero_grad()
Lseg.backward()
optimG.step()
print("[{}][{}]Loss: {:0.4f}".format(epoch,itr,Lseg.data[0]))
best_miou = snapshot(generator,valoader,epoch,best_miou,args.snapshot_dir,args.prefix)
'''
Adversarial Training
'''
def train_adv(generator,discriminator,optimG,optimD,trainloader,valoader,args):
best_miou = -1
for epoch in range(args.start_epoch,args.max_epoch+1):
generator.train()
for batch_id, (img,mask,ohmask) in enumerate(trainloader):
if args.nogpu:
img,mask,ohmask = Variable(img),Variable(mask),Variable(ohmask)
else:
img,mask,ohmask = Variable(img.cuda()),Variable(mask.cuda()),Variable(ohmask.cuda())
itr = len(trainloader)*(epoch-1) + batch_id
cpmap = generator(Variable(img.data,volatile=True))
cpmap = nn.Softmax2d()(cpmap)
N = cpmap.size()[0]
H = cpmap.size()[2]
W = cpmap.size()[3]
# Generate the Real and Fake Labels
targetf = Variable(torch.zeros((N,H,W)).long(),requires_grad=False)
targetr = Variable(torch.ones((N,H,W)).long(),requires_grad=False)
if not args.nogpu:
targetf = targetf.cuda()
targetr = targetr.cuda()
##########################
# DISCRIMINATOR TRAINING #
##########################
optimD.zero_grad()
# Train on Real
confr = nn.LogSoftmax()(discriminator(ohmask.float()))
if args.d_label_smooth != 0:
LDr = (1 - args.d_label_smooth)*nn.NLLLoss2d()(confr,targetr)
LDr += args.d_label_smooth * nn.NLLLoss2d()(confr,targetf)
else:
LDr = nn.NLLLoss2d()(confr,targetr)
LDr.backward()
# Train on Fake
conff = nn.LogSoftmax()(discriminator(Variable(cpmap.data)))
LDf = nn.NLLLoss2d()(conff,targetf)
LDf.backward()
optimD = poly_lr_scheduler(optimD, args.d_lr, itr)
optimD.step()
######################
# GENERATOR TRAINING #
#####################
optimG.zero_grad()
optimD.zero_grad()
cmap = generator(img)
cpmapsmax = nn.Softmax2d()(cmap)
cpmaplsmax = nn.LogSoftmax()(cmap)
conff = nn.LogSoftmax()(discriminator(cpmapsmax))
LGce = nn.NLLLoss2d()(cpmaplsmax,mask)
LGadv = nn.NLLLoss2d()(conff,targetr)
LGseg = LGce + args.lam_adv *LGadv
LGseg.backward()
poly_lr_scheduler(optimG, args.g_lr, itr)
optimG.step()
print("[{}][{}] LD: {:.4f} LDfake: {:.4f} LD_real: {:.4f} LG: {:.4f} LG_ce: {:.4f} LG_adv: {:.4f}" \
.format(epoch,itr,(LDr + LDf).data[0],LDr.data[0],LDf.data[0],LGseg.data[0],LGce.data[0],LGadv.data[0]))
best_miou = snapshot(generator,valoader,epoch,best_miou,args.snapshot_dir,args.prefix)
'''
Semi supervised training
'''
def train_semi(generator,discriminator,optimG,optimD,trainloader_l,trainloader_u,valoader,args):
best_miou = -1
for epoch in range(args.start_epoch,args.max_epoch+1):
generator.train()
trainloader_l_iter = iter(trainloader_l)
trainloader_u_iter = iter(trainloader_u)
print("Epoch: {}".format(epoch))
batch_id = 0
# Randomly pick labeled or unlabeled data for training
while(True):
if random.random() <0.5:
loader = trainloader_l_iter
labeled = True
else:
loader = trainloader_u_iter
labeled = False
# Check if the loader has a batch available
try:
img,mask,ohmask = next(loader)
except:
# Curr loader doesn't have data
if labeled:
loader = trainloader_u_iter
labeled = False
else:
loader = trainloader_l_iter
labeled = True
# Check if the new loader has data
try:
img,mask,ohmask = next(loader)
except:
# Boith loaders exhausted
break
batch_id += 1
if args.nogpu:
img,mask,ohmask = Variable(img),Variable(mask),Variable(ohmask)
else:
img,mask,ohmask = Variable(img.cuda()),Variable(mask.cuda()),Variable(ohmask.cuda())
itr = (len(trainloader_u) + len(trainloader_l))*(epoch-1) + batch_id
if labeled:
################################################
# Labelled data for Discriminator Training #
################################################
cpmap = generator(Variable(img.data,volatile=True))
cpmap = nn.Softmax2d()(cpmap)
N = cpmap.size()[0]
H = cpmap.size()[2]
W = cpmap.size()[3]
# Generate the Real and Fake Labels
targetf = Variable(torch.zeros((N,H,W)).long())
targetr = Variable(torch.ones((N,H,W)).long())
if not args.nogpu:
targetf = targetf.cuda()
targetr = targetr.cuda()
# Train on Real
confr = nn.LogSoftmax()(discriminator(ohmask.float()))
optimD.zero_grad()
if args.d_label_smooth != 0:
LDr = (1 - args.d_label_smooth)*nn.NLLLoss2d()(confr,targetr)
LDr += args.d_label_smooth * nn.NLLLoss2d()(confr,targetf)
else:
LDr = nn.NLLLoss2d()(confr,targetr)
LDr.backward()
# Train on Fake
conff = nn.LogSoftmax()(discriminator(Variable(cpmap.data)))
LDf = nn.NLLLoss2d()(conff,targetf)
LDf.backward()
LDr_d = LDr.data[0]
LDf_d = LDf.data[0]
LD_d = LDr_d + LDf_d
optimD = poly_lr_scheduler(optimD, args.d_lr, itr)
optimD.step()
#####################################
# labelled data Generator Training #
#####################################
optimG.zero_grad()
optimD.zero_grad()
cpmap = generator(img)
cpmapsmax = nn.Softmax2d()(cpmap)
cpmaplsmax = nn.LogSoftmax()(cpmap)
conff = nn.LogSoftmax()(discriminator(cpmapsmax))
LGce = nn.NLLLoss2d()(cpmaplsmax,mask)
LGadv = nn.NLLLoss2d()(conff,targetr)
LGadv_d = LGadv.data[0]
LGce_d = LGce.data[0]
LGsemi_d = 0 # No semi-supervised training
LGadv = args.lam_adv*LGadv
(LGce + LGadv).backward()
optimG = poly_lr_scheduler(optimG, args.g_lr, itr)
optimG.step()
else:
#####################################
# Use unlabelled data to get L_semi #
#####################################
# No discriminator training
LD_d = 0
LDr_d = 0
LDf_d = 0
# Init all loss to 0 for logging ease
LGsemi_d = 0
LGce_d = 0
LGadv_d = 0
optimG.zero_grad()
if epoch > args.wait_semi:
cpmap = generator(img)
cpmapsmax = nn.Softmax2d()(cpmap)
conf = discriminator(cpmapsmax)
confsmax = nn.Softmax2d()(conf)
conflsmax = nn.LogSoftmax()(conf)
N = cpmap.size()[0]
H = cpmap.size()[2]
W = cpmap.size()[3]
# Adversarial Loss
targetr = Variable(torch.ones((N,H,W)).long())
if not args.nogpu:
targetr = targetr.cuda()
LGadv = nn.NLLLoss2d()(conflsmax,targetr)
LGadv_d = LGadv.data[0]
# Semi-Supervised Loss
hardpred = torch.max(cpmapsmax,1)[1].squeeze(1)
idx = np.zeros(cpmap.data.cpu().numpy().shape,dtype=np.uint8)
idx = idx.transpose(0, 2, 3, 1)
confnp = confsmax[:,1,...].data.cpu().numpy()
hardprednp = hardpred.data.cpu().numpy()
idx[confnp > args.t_semi] = np.identity(21, dtype=idx.dtype)[hardprednp[ confnp > args.t_semi]]
LG = args.lam_adv*LGadv
if np.count_nonzero(idx) != 0:
cpmaplsmax = nn.LogSoftmax()(cpmap)
idx = Variable(torch.from_numpy(idx.transpose(0,3,1,2)).byte().cuda())
LGsemi_arr = cpmaplsmax.masked_select(idx)
LGsemi = -1*LGsemi_arr.mean()
LGsemi_d = LGsemi.data[0]
LGsemi = args.lam_semi*LGsemi
LG += LGsemi
LG.backward()
optimG = poly_lr_scheduler(optimG, args.g_lr, itr)
optimG.step()
# Manually free all variables. Look into details of how variables are freed
del idx
del confnp
del confsmax
del hardpred
del hardprednp
del cpmapsmax
del cpmap
LGseg_d = LGce_d + LGadv_d + LGsemi_d
# Manually free memory! Later, really understand how computation graphs free variables
print("[{}][{}] LD: {:.4f} LD_fake: {:.4f} LD_real: {:.4f} LG: {:.4f} LG_ce: {:.4f} LG_adv: {:.4f} LG_semi: {:.4f}"\
.format(epoch,itr,LD_d,LDr_d,LDf_d,LGseg_d,LGce_d,LGadv_d,LGsemi_d))
best_miou = snapshot(generator,valoader,epoch,best_miou,args.snapshot_dir,args.prefix)
def main():
args = parse_args()
random.seed(0)
torch.manual_seed(0)
if not args.nogpu:
torch.cuda.manual_seed_all(0)
if args.no_norm:
imgtr = [ToTensor()]
else:
imgtr = [ToTensor(),NormalizeOwn()]
labtr = [IgnoreLabelClass(),ToTensorLabel()]
cotr = [RandomSizedCrop((321,321))]
trainset_l = PascalVOC(home_dir,args.dataset_dir,img_transform=Compose(imgtr), label_transform=Compose(labtr),co_transform=Compose(cotr),split=args.split,labeled=True)
trainset_u = PascalVOC(home_dir,args.dataset_dir,img_transform=Compose(imgtr), label_transform=Compose(labtr),co_transform=Compose(cotr),split=args.split,labeled=False)
trainloader_l = DataLoader(trainset_l,batch_size=args.batch_size,shuffle=True,num_workers=2,drop_last=True)
trainloader_u = DataLoader(trainset_u,batch_size=args.batch_size,shuffle=True,num_workers=2,drop_last=True)
#########################
# Validation Dataloader #
########################
if args.val_orig:
if args.no_norm:
imgtr = [ZeroPadding(),ToTensor()]
else:
imgtr = [ZeroPadding(),ToTensor(),NormalizeOwn()]
labtr = [IgnoreLabelClass(),ToTensorLabel()]
cotr = []
else:
if args.no_norm:
imgtr = [ToTensor()]
else:
imgtr = [ToTensor(),NormalizeOwn()]
labtr = [IgnoreLabelClass(),ToTensorLabel()]
cotr = [RandomSizedCrop((321,321))]
valset = PascalVOC(home_dir,args.dataset_dir,img_transform=Compose(imgtr), \
label_transform = Compose(labtr),co_transform=Compose(cotr),train_phase=False)
valoader = DataLoader(valset,batch_size=1)
#############
# GENERATOR #
#############
generator = deeplabv2.ResDeeplab()
init_weights(generator,args.init_net)
optimG = optim.SGD(filter(lambda p: p.requires_grad, \
generator.parameters()),lr=args.g_lr,momentum=0.9,\
weight_decay=0.0001,nesterov=True)
if not args.nogpu:
generator = nn.DataParallel(generator).cuda()
#################
# DISCRIMINATOR #
################
if args.mode != "base":
discriminator = Dis(in_channels=21)
if args.d_optim == 'adam':
optimD = optim.Adam(filter(lambda p: p.requires_grad, \
discriminator.parameters()),lr = args.d_lr,weight_decay=0.0001)
else:
optimD = optim.SGD(filter(lambda p: p.requires_grad, \
discriminator.parameters()),lr=args.d_lr,weight_decay=0.0001,momentum=0.5,nesterov=True)
if not args.nogpu:
discriminator = nn.DataParallel(discriminator).cuda()
if args.mode == 'base':
train_base(generator,optimG,trainloader_l,valoader,args)
elif args.mode == 'adv':
train_adv(generator,discriminator,optimG,optimD,trainloader_l,valoader,args)
else:
print("Semi-Supervised training")
train_semi(generator,discriminator,optimG,optimD,trainloader_l,trainloader_u,valoader,args)
if __name__ == '__main__':
main()