-
Notifications
You must be signed in to change notification settings - Fork 0
/
main_segformer.py
525 lines (463 loc) · 22.6 KB
/
main_segformer.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
516
517
518
519
520
521
522
523
524
525
from tqdm import tqdm
import utils
import os
import random
import argparse
import numpy as np
from torch.utils import data
from datasets import VOCSegmentation, Cityscapes, INFRAPARIS, INFRAPARIS_IR
from utils import ext_transforms as et
from metrics import StreamSegMetrics
import torch
import torch.nn as nn
from utils.visualizer import Visualizer
from PIL import Image
import matplotlib
import matplotlib.pyplot as plt
from datasets import av
from segformer.B0.B0 import SegFormerB0 as B0
from segformer.B1.B1 import SegFormerB1 as B1
from segformer.B2.B2 import SegFormerB2 as B2
from segformer.B3.B3 import SegFormerB3 as B3
from segformer.B4.B4 import SegFormerB4 as B4
from segformer.B5.B5 import SegFormerB5 as B5
def get_argparser():
parser = argparse.ArgumentParser()
# Datset Options
parser.add_argument("--data_root", type=str, default='./datasets/data',
help="path to Dataset")
parser.add_argument("--odgt_root", type=str, default='./datasets/data',
help="path to odgt file")
parser.add_argument("--dataset", type=str, default='voc',
choices=['voc', 'cityscapes', 'av', 'infraPARIS', 'infraPARIS_IR'], help='Name of dataset')
parser.add_argument("--num_classes", type=int, default=None,
help="num classes (default: None)")
# Deeplab Options
parser.add_argument("--separable_conv", action='store_true', default=False,
help="apply separable conv to decoder and aspp")
parser.add_argument("--output_stride", type=int, default=16, choices=[8, 16])
parser.add_argument("--model", type=str, default='B0',
choices=['B0', 'B1','B2','B3','B4','B5'], help='Name of dataset')
# Train Options
parser.add_argument("--test_only", action='store_true', default=False)
parser.add_argument("--save_val_results", action='store_true', default=False,
help="save segmentation results to \"./results\"")
parser.add_argument("--total_itrs", type=int, default=160e3,
help="epoch number (default: 30k)")
parser.add_argument("--lr", type=float, default=0.01,
help="learning rate (default: 0.01)")
parser.add_argument("--lr_policy", type=str, default='poly', choices=['poly', 'step'],
help="learning rate scheduler policy")
parser.add_argument("--step_size", type=int, default=10000)
parser.add_argument("--crop_val", action='store_true', default=False,
help='crop validation (default: False)')
parser.add_argument("--batch_size", type=int, default=16,
help='batch size (default: 16)')
parser.add_argument("--val_batch_size", type=int, default=4,
help='batch size for validation (default: 4)')
parser.add_argument("--crop_size", type=int, default=513)
parser.add_argument("--crop_size_h", type=int, default=513)
parser.add_argument("--crop_size_w", type=int, default=513)
parser.add_argument("--ckpt", default=None, type=str,
help="restore from checkpoint")
parser.add_argument("--ckpt_segformer", default=None, type=str,
help="restore segformer from checkpoint")
parser.add_argument("--continue_training", action='store_true', default=False)
parser.add_argument("--loss_type", type=str, default='cross_entropy',
choices=['cross_entropy', 'focal_loss'], help="loss type (default: False)")
parser.add_argument("--gpu_id", type=str, default='0',
help="GPU ID")
parser.add_argument("--weight_decay", type=float, default=1e-4,
help='weight decay (default: 1e-4)')
parser.add_argument("--random_seed", type=int, default=1,
help="random seed (default: 1)")
parser.add_argument("--print_interval", type=int, default=10,
help="print interval of loss (default: 10)")
parser.add_argument("--val_interval", type=int, default=100,
help="epoch interval for eval (default: 100)")
parser.add_argument("--download", action='store_true', default=False,
help="download datasets")
# PASCAL VOC Options
parser.add_argument("--year", type=str, default='2012',
choices=['2012_aug', '2012', '2011', '2009', '2008', '2007'], help='year of VOC')
# Visdom options
parser.add_argument("--enable_vis", action='store_true', default=False,
help="use visdom for visualization")
parser.add_argument("--vis_port", type=str, default='13570',
help='port for visdom')
parser.add_argument("--vis_env", type=str, default='main',
help='env for visdom')
parser.add_argument("--vis_num_samples", type=int, default=8,
help='number of samples for visualization (default: 8)')
return parser
def get_dataset(opts):
""" Dataset And Augmentation
"""
if opts.dataset == 'voc':
train_transform = et.ExtCompose([
# et.ExtResize(size=opts.crop_size),
et.ExtRandomScale((0.5, 2.0)),
et.ExtRandomCrop(size=(opts.crop_size, opts.crop_size), pad_if_needed=True),
et.ExtRandomHorizontalFlip(),
et.ExtToTensor(),
et.ExtNormalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225]),
])
if opts.crop_val:
val_transform = et.ExtCompose([
et.ExtResize(opts.crop_size),
et.ExtCenterCrop(opts.crop_size),
et.ExtToTensor(),
et.ExtNormalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225]),
])
else:
val_transform = et.ExtCompose([
et.ExtToTensor(),
et.ExtNormalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225]),
])
train_dst = VOCSegmentation(root=opts.data_root, year=opts.year,
image_set='train', download=opts.download, transform=train_transform)
val_dst = VOCSegmentation(root=opts.data_root, year=opts.year,
image_set='val', download=False, transform=val_transform)
return train_dst, val_dst
if opts.dataset == 'cityscapes':
train_transform = et.ExtCompose([
# et.ExtResize( 512 ),
et.ExtRandomCrop(size=(opts.crop_size_h, opts.crop_size_w)),
et.ExtColorJitter(brightness=0.5, contrast=0.5, saturation=0.5),
et.ExtRandomHorizontalFlip(),
et.ExtToTensor(),
et.ExtNormalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225]),
])
val_transform = et.ExtCompose([
# et.ExtResize( 512 ),
et.ExtToTensor(),
et.ExtNormalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225]),
])
train_dst = Cityscapes(root=opts.data_root,
split='train', transform=train_transform)
val_dst = Cityscapes(root=opts.data_root,
split='val', transform=val_transform)
return train_dst, val_dst
if opts.dataset == 'infraPARIS':
train_transform = et.ExtCompose([
# et.ExtResize( 512 ),
et.ExtRandomCrop(size=(opts.crop_size_h, opts.crop_size_w)),
et.ExtColorJitter(brightness=0.5, contrast=0.5, saturation=0.5),
et.ExtRandomHorizontalFlip(),
et.ExtToTensor(),
et.ExtNormalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225]),
])
val_transform = et.ExtCompose([
#et.ExtResize( 512 ),
et.ExtToTensor(),
et.ExtNormalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225]),
])
train_dst = INFRAPARIS(root=opts.data_root,
split='train', transform=train_transform)
val_dst = INFRAPARIS(root=opts.data_root,
split='val', transform=val_transform)
test_dst = INFRAPARIS(root=opts.data_root,
split='test', transform=val_transform)
return train_dst, val_dst, test_dst
if opts.dataset == 'infraPARIS_IR':
train_transform = et.ExtCompose([
# et.ExtResize( 512 ),
et.ExtRandomCrop(size=(opts.crop_size_h, opts.crop_size_w)),
#et.ExtColorJitter(brightness=0.5, contrast=0.5, saturation=0.5),
et.ExtRandomHorizontalFlip(),
et.ExtToTensor(),
et.ExtNormalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225]),
])
val_transform = et.ExtCompose([
# et.ExtResize( 512 ),
et.ExtToTensor(),
et.ExtNormalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225]),
])
train_dst = INFRAPARIS_IR(root=opts.data_root,
split='train', transform=train_transform)
val_dst = INFRAPARIS_IR(root=opts.data_root,
split='val', transform=val_transform)
test_dst = INFRAPARIS_IR(root=opts.data_root,
split='test', transform=val_transform)
return train_dst, val_dst, test_dst
if opts.dataset == 'av':
train_transform = et.ExtCompose([
# et.ExtResize( 512 ),
et.ExtRandomCrop(size=(opts.crop_size, opts.crop_size)),
et.ExtColorJitter(brightness=0.5, contrast=0.5, saturation=0.5),
et.ExtRandomHorizontalFlip(),
et.ExtToTensor(),
et.ExtNormalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225]),
])
val_transform = et.ExtCompose([
# et.ExtResize( 512 ),
et.ExtToTensor(),
et.ExtNormalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225]),
])
train_dst = av.dataset(root_dataset=opts.data_root, root_odgt=opts.odgt_root,
split = 'train', transform=train_transform)
val_dst = av.dataset(root_dataset=opts.data_root, root_odgt=opts.odgt_root,
split = 'val', transform=val_transform)
return train_dst, val_dst
def validate(opts, model, loader, device, metrics, ret_samples_ids=None):
"""Do validation and return specified samples"""
metrics.reset()
ret_samples = []
if opts.save_val_results:
if not os.path.exists('results'):
os.mkdir('results')
denorm = utils.Denormalize(mean=[123.675, 116.28, 103.53], std=[58.395, 57.12, 57.375])
img_id = 0
with torch.no_grad():
for i, (images, labels) in tqdm(enumerate(loader)):
images = images.to(device, dtype=torch.float32)
labels = labels.to(device, dtype=torch.long)
outputs = model(images)
preds = outputs.detach().max(dim=1)[1].cpu().numpy()
targets = labels.cpu().numpy()
metrics.update(targets, preds)
if ret_samples_ids is not None and i in ret_samples_ids: # get vis samples
ret_samples.append(
(images[0].detach().cpu().numpy(), targets[0], preds[0]))
if opts.save_val_results:
for i in range(len(images)):
image = images[i].detach().cpu().numpy()
target = targets[i]
pred = preds[i]
image = (denorm(image) * 255).transpose(1, 2, 0).astype(np.uint8)
target = loader.dataset.decode_target(target).astype(np.uint8)
pred = loader.dataset.decode_target(pred).astype(np.uint8)
Image.fromarray(image).save('results/%d_image.png' % img_id)
Image.fromarray(target).save('results/%d_target.png' % img_id)
Image.fromarray(pred).save('results/%d_pred.png' % img_id)
fig = plt.figure()
plt.imshow(image)
plt.axis('off')
plt.imshow(pred, alpha=0.7)
ax = plt.gca()
ax.xaxis.set_major_locator(matplotlib.ticker.NullLocator())
ax.yaxis.set_major_locator(matplotlib.ticker.NullLocator())
plt.savefig('results/%d_overlay.png' % img_id, bbox_inches='tight', pad_inches=0)
plt.close()
img_id += 1
score = metrics.get_results()
return score, ret_samples
def main():
opts = get_argparser().parse_args()
if opts.dataset.lower() == 'voc':
opts.num_classes = 21
elif opts.dataset.lower() == 'cityscapes':
opts.num_classes = 19
elif opts.dataset.lower() == 'infraparis':
opts.num_classes = 19
elif opts.dataset.lower() == 'infraparis_ir':
opts.num_classes = 19
elif opts.dataset.lower() == 'av':
opts.num_classes = 19
# Setup visualization
vis = Visualizer(port=opts.vis_port,
env=opts.vis_env) if opts.enable_vis else None
if vis is not None: # display options
vis.vis_table("Options", vars(opts))
os.environ['CUDA_VISIBLE_DEVICES'] = opts.gpu_id
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
print("Device: %s" % device)
# Setup random seed
torch.manual_seed(opts.random_seed)
np.random.seed(opts.random_seed)
random.seed(opts.random_seed)
# Setup dataloader
if opts.dataset=='voc' and not opts.crop_val:
opts.val_batch_size = 1
if (opts.dataset.lower() == 'infraparis') or ( opts.dataset.lower() == 'infraparis_ir'):
train_dst, val_dst,test_dst = get_dataset(opts)
test_loader = data.DataLoader(
test_dst, batch_size=opts.val_batch_size, shuffle=True, num_workers=2, drop_last=True)
else:
train_dst, val_dst = get_dataset(opts)
train_loader = data.DataLoader(
train_dst, batch_size=opts.batch_size, shuffle=True, num_workers=2,
drop_last=True) # drop_last=True to ignore single-image batches.
val_loader = data.DataLoader(
val_dst, batch_size=opts.val_batch_size, shuffle=True, num_workers=2)
print("Dataset: %s, Train set: %d, Val set: %d" %
(opts.dataset, len(train_dst), len(val_dst)))
# Set up model (all model are 'constructed at network.modeling)
if opts.model == 'B0':
model = B0()
elif opts.model == 'B1':
model = B1()
elif opts.model == 'B2':
model = B2()
elif opts.model == 'B3':
model = B3()
elif opts.model == 'B4':
model = B4()
elif opts.model == 'B5':
model = B5()
#utils.set_bn_momentum(model.backbone, momentum=0.01)
# Set up metrics
metrics = StreamSegMetrics(opts.num_classes)
optimizer = torch.optim.AdamW(params=[
{'params': model.encoder.parameters(), 'lr': opts.lr},
{'params': model.head.parameters(), 'lr': opts.lr * 10},
], lr=opts.lr, weight_decay=opts.weight_decay, betas=(0.9, 0.999), )
#optimizer = torch.optim.AdamW(
# model.parameters(), lr=opts.lr, weight_decay=opts.weight_decay, betas=(0.9, 0.999), )
# optimizer = torch.optim.SGD(params=model.parameters(), lr=opts.lr, momentum=0.9, weight_decay=opts.weight_decay)
# torch.optim.lr_scheduler.StepLR(optimizer, step_size=opts.lr_decay_step, gamma=opts.lr_decay_factor)
# if opts.lr_policy == 'poly':
# scheduler = utils.PolyLR(optimizer, opts.total_itrs, power=0.9)
if opts.lr_policy == 'poly':
scheduler = utils.PolyLR(optimizer, opts.total_itrs, power=1.0)
elif opts.lr_policy == 'step':
scheduler = torch.optim.lr_scheduler.StepLR(optimizer, step_size=opts.step_size, gamma=0.1)
# Set up criterion
# criterion = utils.get_loss(opts.loss_type)
if opts.loss_type == 'focal_loss':
criterion = utils.FocalLoss(ignore_index=255, size_average=True)
elif opts.loss_type == 'cross_entropy':
criterion = nn.CrossEntropyLoss(ignore_index=255, reduction='mean')
def save_ckpt(path):
""" save current model
"""
torch.save({
"cur_itrs": cur_itrs,
"model_state": model.module.state_dict(),
"optimizer_state": optimizer.state_dict(),
"scheduler_state": scheduler.state_dict(),
"best_score": best_score,
}, path)
print("Model saved as %s" % path)
utils.mkdir('checkpoints')
# Restore
best_score = 0.0
cur_itrs = 0
cur_epochs = 0
if opts.ckpt is not None and os.path.isfile(opts.ckpt):
# https://github.com/VainF/DeepLabV3Plus-Pytorch/issues/8#issuecomment-605601402, @PytaichukBohdan
checkpoint = torch.load(opts.ckpt, map_location=torch.device('cpu'))
print('t')
model.load_state_dict(checkpoint["model_state"])
model = nn.DataParallel(model)
print('t')
model.to(device)
if opts.continue_training:
optimizer.load_state_dict(checkpoint["optimizer_state"])
scheduler.load_state_dict(checkpoint["scheduler_state"])
cur_itrs = checkpoint["cur_itrs"]
best_score = checkpoint['best_score']
print("Training state restored from %s" % opts.ckpt)
print("Model restored from %s" % opts.ckpt)
del checkpoint # free memory
elif opts.ckpt_segformer is not None and os.path.isfile(opts.ckpt_segformer):
checkpoint = torch.load(opts.ckpt_segformer, map_location=torch.device('cpu'))
model_dict = model.state_dict()
state_dict = {'encoder.'+k:v for k,v in checkpoint.items()}
#state_dict['head.fuse.0.weight']=state_dict['encoder.head.weight']
#del state_dict['encoder.head.weight']
#state_dict['head.fuse.0.bias']=state_dict['encoder.head.bias']
#del state_dict['encoder.head.bias']
state_dict.popitem()
state_dict.popitem()
for k,v in state_dict.items():
print(k)
model_dict.update(state_dict)
model.load_state_dict(model_dict)
model = nn.DataParallel(model)
model.to(device)
print('[load pretrianed model from mit b0]')
if opts.continue_training:
optimizer.load_state_dict(checkpoint["optimizer_state"])
scheduler.load_state_dict(checkpoint["scheduler_state"])
cur_itrs = checkpoint["cur_itrs"]
best_score = checkpoint['best_score']
print("Training state restored from %s" % opts.ckpt)
print("Model restored from %s" % opts.ckpt)
del checkpoint # free memory
else:
print("[!] Retrain")
model = nn.DataParallel(model)
model.to(device)
# ========== Train Loop ==========#
vis_sample_id = np.random.randint(0, len(val_loader), opts.vis_num_samples,
np.int32) if opts.enable_vis else None # sample idxs for visualization
denorm = utils.Denormalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225]) # denormalization for ori images
if opts.test_only:
model.eval()
if (opts.dataset.lower() == 'infraparis') or ( opts.dataset.lower() == 'infraparis_ir'):
print("Dataset: %s, Test set: %d" % (opts.dataset, len(test_dst)))
val_score, ret_samples = validate(
opts=opts, model=model, loader=test_loader, device=device, metrics=metrics,
ret_samples_ids=vis_sample_id)
print(metrics.to_str(val_score))
else:
val_score, ret_samples = validate(
opts=opts, model=model, loader=val_loader, device=device, metrics=metrics,
ret_samples_ids=vis_sample_id)
print(metrics.to_str(val_score))
return
interval_loss = 0
while True: # cur_itrs < opts.total_itrs:
# ===== Train =====
model.train()
cur_epochs += 1
for (images, labels) in train_loader:
cur_itrs += 1
images = images.to(device, dtype=torch.float32)
labels = labels.to(device, dtype=torch.long)
optimizer.zero_grad()
outputs = model(images)
loss = criterion(outputs, labels)
loss.backward()
optimizer.step()
np_loss = loss.detach().cpu().numpy()
interval_loss += np_loss
if vis is not None:
vis.vis_scalar('Loss', cur_itrs, np_loss)
if (cur_itrs) % 10 == 0:
interval_loss = interval_loss / 10
print("Epoch %d, Itrs %d/%d, Loss=%f" %
(cur_epochs, cur_itrs, opts.total_itrs, interval_loss))
interval_loss = 0.0
if (cur_itrs) % opts.val_interval == 0:
save_ckpt('checkpoints/latest_%s_%s_os%d.pth' %
(opts.model, opts.dataset, opts.output_stride))
print("validation...")
model.eval()
val_score, ret_samples = validate(
opts=opts, model=model, loader=val_loader, device=device, metrics=metrics,
ret_samples_ids=vis_sample_id)
print(metrics.to_str(val_score))
if val_score['Mean IoU'] > best_score: # save best model
best_score = val_score['Mean IoU']
save_ckpt('checkpoints/best_%s_%s_os%d.pth' %
(opts.model, opts.dataset, opts.output_stride))
if vis is not None: # visualize validation score and samples
vis.vis_scalar("[Val] Overall Acc", cur_itrs, val_score['Overall Acc'])
vis.vis_scalar("[Val] Mean IoU", cur_itrs, val_score['Mean IoU'])
vis.vis_table("[Val] Class IoU", val_score['Class IoU'])
for k, (img, target, lbl) in enumerate(ret_samples):
img = (denorm(img) * 255).astype(np.uint8)
target = train_dst.decode_target(target).transpose(2, 0, 1).astype(np.uint8)
lbl = train_dst.decode_target(lbl).transpose(2, 0, 1).astype(np.uint8)
concat_img = np.concatenate((img, target, lbl), axis=2) # concat along width
vis.vis_image('Sample %d' % k, concat_img)
model.train()
scheduler.step()
if cur_itrs >= opts.total_itrs:
return
if __name__ == '__main__':
main()