-
Notifications
You must be signed in to change notification settings - Fork 1.7k
/
Copy pathtrain.py
552 lines (462 loc) · 24.9 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
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import os
import time
import numpy as np
import shutil
import functools
from collections import deque
import paddle
import paddle.nn.functional as F
import paddleseg.transforms.functional as Func
from paddleseg.utils import TimeAverager, calculate_eta, resume, logger, worker_init_fn
from paddleseg.models import losses
from models import EMA
from script import val
from utils import augmentation, load_ema_model, save_edge
paddle.set_printoptions(precision=15)
class Trainer():
def __init__(self, model, cfg):
'''model（nn.Layer): A sementic segmentation model.'''
self.cfg = cfg
self.ema_decay = cfg['ema_decay']
self.edgeconstrain = cfg['edgeconstrain']
self.edgepullin = cfg['edgepullin']
self.src_only = cfg['src_only']
self.featurepullin = cfg['featurepullin']
self.model = model
self.ema = EMA(self.model, self.ema_decay)
self.celoss = losses.CrossEntropyLoss()
self.klloss = losses.KLLoss()
self.mseloss = losses.MSELoss()
self.bceloss_src = losses.BCELoss(weight='dynamic')
self.bceloss_tgt = losses.BCELoss(weight='dynamic')
self.src_centers = [paddle.zeros((1, 19)) for _ in range(19)]
self.tgt_centers = [paddle.zeros((1, 19)) for _ in range(19)]
if 'None' in cfg['resume_ema']:
self.resume_ema = None
else:
self.resume_ema = cfg['resume_ema']
def train(self,
train_dataset_src,
train_dataset_tgt,
val_dataset_tgt=None,
val_dataset_src=None,
optimizer=None,
save_dir='output',
iters=10000,
batch_size=2,
resume_model=None,
save_interval=1000,
log_iters=10,
num_workers=0,
use_vdl=False,
keep_checkpoint_max=5,
test_config=None):
"""
Launch training.
Args:
train_dataset (paddle.io.Dataset): Used to read and process training datasets.
val_dataset_tgt (paddle.io.Dataset, optional): Used to read and process validation datasets.
optimizer (paddle.optimizer.Optimizer): The optimizer.
save_dir (str, optional): The directory for saving the model snapshot. Default: 'output'.
iters (int, optional): How may iters to train the model. Defualt: 10000.
batch_size (int, optional): Mini batch size of one gpu or cpu. Default: 2.
resume_model (str, optional): The path of resume model.
save_interval (int, optional): How many iters to save a model snapshot once during training. Default: 1000.
log_iters (int, optional): Display logging information at every log_iters. Default: 10.
num_workers (int, optional): Num workers for data loader. Default: 0.
use_vdl (bool, optional): Whether to record the data to VisualDL during training. Default: False.
keep_checkpoint_max (int, optional): Maximum number of checkpoints to save. Default: 5.
test_config(dict, optional): Evaluation config.
"""
start_iter = 0
self.model.train()
nranks = paddle.distributed.ParallelEnv().nranks
local_rank = paddle.distributed.ParallelEnv().local_rank
if resume_model is not None:
logger.info(resume_model)
start_iter = resume(self.model, optimizer, resume_model)
load_ema_model(self.model, self.resume_ema)
if not os.path.isdir(save_dir):
if os.path.exists(save_dir):
os.remove(save_dir)
os.makedirs(save_dir)
if nranks > 1:
paddle.distributed.fleet.init(is_collective=True)
optimizer = paddle.distributed.fleet.distributed_optimizer(
optimizer) # The return is Fleet object
ddp_model = paddle.distributed.fleet.distributed_model(self.model)
batch_sampler_src = paddle.io.DistributedBatchSampler(
train_dataset_src,
batch_size=batch_size,
shuffle=True,
drop_last=True)
loader_src = paddle.io.DataLoader(
train_dataset_src,
batch_sampler=batch_sampler_src,
num_workers=num_workers,
return_list=True,
worker_init_fn=worker_init_fn, )
batch_sampler_tgt = paddle.io.DistributedBatchSampler(
train_dataset_tgt,
batch_size=batch_size,
shuffle=True,
drop_last=True)
loader_tgt = paddle.io.DataLoader(
train_dataset_tgt,
batch_sampler=batch_sampler_tgt,
num_workers=num_workers,
return_list=True,
worker_init_fn=worker_init_fn, )
if use_vdl:
from visualdl import LogWriter
log_writer = LogWriter(save_dir)
iters_per_epoch = len(batch_sampler_tgt)
best_mean_iou = -1.0
best_model_iter = -1
reader_cost_averager = TimeAverager()
batch_cost_averager = TimeAverager()
save_models = deque()
batch_start = time.time()
iter = start_iter
while iter < iters:
for _, (data_src,
data_tgt) in enumerate(zip(loader_src, loader_tgt)):
reader_cost_averager.record(time.time() - batch_start)
loss_dict = {}
#### training #####
images_tgt = data_tgt[0]
labels_tgt = data_tgt[1].astype('int64')
images_src = data_src[0]
labels_src = data_src[1].astype('int64')
edges_src = data_src[2].astype('int64')
edges_tgt = data_tgt[2].astype('int64')
if nranks > 1:
logits_list_src = ddp_model(images_src)
else:
logits_list_src = self.model(images_src)
##### source seg & edge loss ####
loss_src_seg_main = self.celoss(logits_list_src[0], labels_src)
loss_src_seg_aux = 0.1 * self.celoss(logits_list_src[1],
labels_src)
loss_src_seg = loss_src_seg_main + loss_src_seg_aux
loss_dict["source_main"] = float(loss_src_seg_main)
loss_dict["source_aux"] = float(loss_src_seg_aux)
loss = loss_src_seg
del loss_src_seg, loss_src_seg_aux, loss_src_seg_main
#### generate target pseudo label ####
with paddle.no_grad():
if nranks > 1:
logits_list_tgt = ddp_model(images_tgt)
else:
logits_list_tgt = self.model(images_tgt)
pred_P_1 = F.softmax(logits_list_tgt[0], axis=1)
labels_tgt_psu = paddle.argmax(pred_P_1.detach(), axis=1)
# aux label
pred_P_2 = F.softmax(logits_list_tgt[1], axis=1)
pred_c = (pred_P_1 + pred_P_2) / 2
labels_tgt_psu_aux = paddle.argmax(pred_c.detach(), axis=1)
if self.edgeconstrain:
loss_src_edge = self.bceloss_src(
logits_list_src[2], edges_src) # 1, 2 640, 1280
src_edge = paddle.argmax(
logits_list_src[2].detach().clone(),
axis=1) # 1, 1, 640,1280
src_edge_acc = ((src_edge == edges_src).numpy().sum().astype('float32')\
/functools.reduce(lambda a, b: a * b, src_edge.shape))*100
if (not self.src_only) and (iter > 200000):
#### target seg & edge loss ####
logger.info("Add target edege loss")
edges_tgt = Func.mask_to_binary_edge(
labels_tgt_psu.detach().clone().numpy(),
radius=2,
num_classes=train_dataset_tgt.NUM_CLASSES)
edges_tgt = paddle.to_tensor(edges_tgt, dtype='int64')
loss_tgt_edge = self.bceloss_tgt(logits_list_tgt[2],
edges_tgt)
loss_edge = loss_tgt_edge + loss_src_edge
else:
loss_tgt_edge = paddle.zeros([1])
loss_edge = loss_src_edge
loss += loss_edge
loss_dict['target_edge'] = float(loss_tgt_edge)
loss_dict['source_edge'] = float(loss_src_edge)
del loss_edge, loss_tgt_edge, loss_src_edge
#### target aug loss #######
augs = augmentation.get_augmentation()
images_tgt_aug, labels_tgt_aug = augmentation.augment(
images=images_tgt.cpu(),
labels=labels_tgt_psu.detach().cpu(),
aug=augs,
iters="{}_1".format(iter))
images_tgt_aug = images_tgt_aug.cuda()
labels_tgt_aug = labels_tgt_aug.cuda()
_, labels_tgt_aug_aux = augmentation.augment(
images=images_tgt.cpu(),
labels=labels_tgt_psu_aux.detach().cpu(),
aug=augs,
iters="{}_2".format(iter))
labels_tgt_aug_aux = labels_tgt_aug_aux.cuda()
if nranks > 1:
logits_list_tgt_aug = ddp_model(images_tgt_aug)
else:
logits_list_tgt_aug = self.model(images_tgt_aug)
loss_tgt_aug_main = 0.1 * (self.celoss(logits_list_tgt_aug[0],
labels_tgt_aug))
loss_tgt_aug_aux = 0.1 * (0.1 * self.celoss(
logits_list_tgt_aug[1], labels_tgt_aug_aux))
loss_tgt_aug = loss_tgt_aug_aux + loss_tgt_aug_main
loss += loss_tgt_aug
loss_dict['target_aug_main'] = float(loss_tgt_aug_main)
loss_dict['target_aug_aux'] = float(loss_tgt_aug_aux)
del images_tgt_aug, labels_tgt_aug_aux, images_tgt, \
loss_tgt_aug, loss_tgt_aug_aux, loss_tgt_aug_main
#### edge input seg; src & tgt edge pull in ######
if self.edgepullin:
src_edge_logit = logits_list_src[2]
feat_src = paddle.concat(
[logits_list_src[0], src_edge_logit], axis=1).detach()
out_src = self.model.fusion(feat_src)
loss_src_edge_rec = self.celoss(out_src, labels_src)
tgt_edge_logit = logits_list_tgt_aug[2]
# tgt_edge_logit = paddle.to_tensor(
# Func.mask_to_onehot(edges_tgt.squeeze().numpy(), 2)
# ).unsqueeze(0).astype('float32')
feat_tgt = paddle.concat(
[logits_list_tgt[0], tgt_edge_logit], axis=1).detach()
out_tgt = self.model.fusion(feat_tgt)
loss_tgt_edge_rec = self.celoss(out_tgt, labels_tgt)
loss_edge_rec = loss_tgt_edge_rec + loss_src_edge_rec
loss += loss_edge_rec
loss_dict['src_edge_rec'] = float(loss_src_edge_rec)
loss_dict['tgt_edge_rec'] = float(loss_tgt_edge_rec)
del loss_tgt_edge_rec, loss_src_edge_rec
#### mask input feature & pullin ######
if self.featurepullin:
# inner-class loss
feat_src = logits_list_src[0]
feat_tgt = logits_list_tgt_aug[0]
center_src_s, center_tgt_s = [], []
total_pixs = logits_list_src[0].shape[2] * \
logits_list_src[0].shape[3]
for i in range(train_dataset_tgt.NUM_CLASSES):
pred = paddle.argmax(
logits_list_src[0].detach().clone(),
axis=1).unsqueeze(0) # 1, 1, 640, 1280
sel_num = paddle.sum((pred == i).astype('float32'))
# ignore tensor that do not have features in this img
if sel_num > 0:
feat_sel_src = paddle.where(
(pred == i).expand_as(feat_src), feat_src,
paddle.zeros(feat_src.shape))
center_src = paddle.mean(
feat_sel_src,
axis=[2, 3]) / (sel_num / total_pixs) # 1, C
self.src_centers[i] = 0.99 * self.src_centers[i] + (
1 - 0.99) * center_src
pred = labels_tgt_aug.unsqueeze(0) # 1, 1, 512, 512
sel_num = paddle.sum((pred == i).astype('float32'))
if sel_num > 0:
feat_sel_tgt = paddle.where(
(pred == i).expand_as(feat_tgt), feat_tgt,
paddle.zeros(feat_tgt.shape))
center_tgt = paddle.mean(
feat_sel_tgt, axis=[2,
3]) / (sel_num / total_pixs)
self.tgt_centers[i] = 0.99 * self.tgt_centers[i] + (
1 - 0.99) * center_tgt
center_src_s.append(center_src)
center_tgt_s.append(center_tgt)
if iter >= 3000: # average center structure alignment
src_centers = paddle.concat(self.src_centers, axis=0)
tgt_centers = paddle.concat(
self.tgt_centers, axis=0) # 19， 2048
relatmat_src = paddle.matmul(
src_centers, src_centers, transpose_y=True) # 19，19
relatmat_tgt = paddle.matmul(
tgt_centers, tgt_centers, transpose_y=True)
loss_intra_relate = self.klloss(relatmat_src, (relatmat_tgt+relatmat_src)/2) \
+ self.klloss(relatmat_tgt, (relatmat_tgt+relatmat_src)/2)
loss_pix_align_src = self.mseloss(
paddle.to_tensor(center_src_s),
paddle.to_tensor(self.src_centers).detach().clone())
loss_pix_align_tgt = self.mseloss(
paddle.to_tensor(center_tgt_s),
paddle.to_tensor(self.tgt_centers).detach().clone())
loss_feat_align = loss_pix_align_src + loss_pix_align_tgt + loss_intra_relate
loss += loss_feat_align
loss_dict['loss_pix_align_src'] = float(
loss_pix_align_src)
loss_dict['loss_pix_align_tgt'] = float(
loss_pix_align_tgt)
loss_dict['loss_intra_relate'] = float(
loss_intra_relate)
del loss_pix_align_tgt, loss_pix_align_src, loss_intra_relate,
self.tgt_centers = [
item.detach().clone() for item in self.tgt_centers
]
self.src_centers = [
item.detach().clone() for item in self.src_centers
]
loss.backward()
del loss
loss = sum(loss_dict.values())
optimizer.step()
self.ema.update_params()
with paddle.no_grad():
##### log & save #####
lr = optimizer.get_lr()
# update lr
if isinstance(optimizer, paddle.distributed.fleet.Fleet):
lr_sche = optimizer.user_defined_optimizer._learning_rate
else:
lr_sche = optimizer._learning_rate
if isinstance(lr_sche, paddle.optimizer.lr.LRScheduler):
lr_sche.step()
if self.cfg['save_edge']:
tgt_edge = paddle.argmax(
logits_list_tgt_aug[2].detach().clone(),
axis=1) # 1, 1, 640,1280
src_feed_gt = paddle.argmax(
src_edge_logit.astype('float32'), axis=1)
tgt_feed_gt = paddle.argmax(
tgt_edge_logit.astype('float32'), axis=1)
logger.info('src_feed_gt_{}_{}_{}'.format(
src_feed_gt.shape,
src_feed_gt.max(), src_feed_gt.min()))
logger.info('tgt_feed_gt_{}_{}_{}'.format(
tgt_feed_gt.shape,
max(tgt_feed_gt), min(tgt_feed_gt)))
save_edge(src_feed_gt, 'src_feed_gt_{}'.format(iter))
save_edge(tgt_feed_gt, 'tgt_feed_gt_{}'.format(iter))
save_edge(tgt_edge, 'tgt_pred_{}'.format(iter))
save_edge(src_edge,
'src_pred_{}_{}'.format(iter, src_edge_acc))
save_edge(edges_src, 'src_gt_{}'.format(iter))
save_edge(edges_tgt, 'tgt_gt_{}'.format(iter))
self.model.clear_gradients()
batch_cost_averager.record(
time.time() - batch_start, num_samples=batch_size)
iter += 1
if (iter) % log_iters == 0 and local_rank == 0:
label_tgt_acc = ((labels_tgt == labels_tgt_psu).numpy().sum().astype('float32')\
/functools.reduce(lambda a, b: a * b, labels_tgt_psu.shape))*100
remain_iters = iters - iter
avg_train_batch_cost = batch_cost_averager.get_average()
avg_train_reader_cost = reader_cost_averager.get_average(
)
eta = calculate_eta(remain_iters, avg_train_batch_cost)
logger.info(
"[TRAIN] epoch: {}, iter: {}/{}, loss: {:.4f}, tgt_pix_acc: {:.4f}, lr: {:.6f}, batch_cost: {:.4f}, reader_cost: {:.5f}, ips: {:.4f} samples/sec | ETA {}"
.format((iter - 1) // iters_per_epoch + 1, iter,
iters, loss, label_tgt_acc, lr,
avg_train_batch_cost, avg_train_reader_cost,
batch_cost_averager.get_ips_average(), eta))
if use_vdl:
log_writer.add_scalar('Train/loss', loss, iter)
# Record all losses if there are more than 2 losses.
if len(loss_dict) > 1:
for name, loss in loss_dict.items():
log_writer.add_scalar('Train/loss_' + name,
loss, iter)
log_writer.add_scalar('Train/lr', lr, iter)
log_writer.add_scalar('Train/batch_cost',
avg_train_batch_cost, iter)
log_writer.add_scalar('Train/reader_cost',
avg_train_reader_cost, iter)
log_writer.add_scalar('Train/tgt_label_acc',
label_tgt_acc, iter)
reader_cost_averager.reset()
batch_cost_averager.reset()
if (iter % save_interval == 0 or
iter == iters) and (val_dataset_tgt is not None):
num_workers = 4 if num_workers > 0 else 0 # adjust num_worker=4
if test_config is None:
test_config = {}
self.ema.apply_shadow()
self.ema.model.eval()
PA_tgt, _, MIoU_tgt, _ = val.evaluate(
self.model,
val_dataset_tgt,
num_workers=num_workers,
**test_config)
if (iter % (save_interval * 30)) == 0 \
and self.cfg['eval_src']: # add evaluate on src
PA_src, _, MIoU_src, _ = val.evaluate(
self.model,
val_dataset_src,
num_workers=num_workers,
**test_config)
logger.info(
'[EVAL] The source mIoU is ({:.4f}) at iter {}.'
.format(MIoU_src, iter))
self.ema.restore()
self.model.train()
if (iter % save_interval == 0 or
iter == iters) and local_rank == 0:
current_save_dir = os.path.join(save_dir,
"iter_{}".format(iter))
if not os.path.isdir(current_save_dir):
os.makedirs(current_save_dir)
paddle.save(
self.model.state_dict(),
os.path.join(current_save_dir, 'model.pdparams'))
paddle.save(self.ema.shadow,
os.path.join(current_save_dir,
'model_ema.pdparams'))
paddle.save(
optimizer.state_dict(),
os.path.join(current_save_dir, 'model.pdopt'))
save_models.append(current_save_dir)
if len(save_models) > keep_checkpoint_max > 0:
model_to_remove = save_models.popleft()
shutil.rmtree(model_to_remove)
if val_dataset_tgt is not None:
if MIoU_tgt > best_mean_iou:
best_mean_iou = MIoU_tgt
best_model_iter = iter
best_model_dir = os.path.join(save_dir,
"best_model")
paddle.save(self.model.state_dict(),
os.path.join(best_model_dir,
'model.pdparams'))
logger.info(
'[EVAL] The model with the best validation mIoU ({:.4f}) was saved at iter {}.'
.format(best_mean_iou, best_model_iter))
if use_vdl:
log_writer.add_scalar('Evaluate/mIoU', MIoU_tgt,
iter)
log_writer.add_scalar('Evaluate/PA', PA_tgt, iter)
if self.cfg['eval_src']:
log_writer.add_scalar('Evaluate/mIoU_src',
MIoU_src, iter)
log_writer.add_scalar('Evaluate/PA_src', PA_src,
iter)
batch_start = time.time()
self.ema.update_buffer()
# # Calculate flops.
if local_rank == 0:
def count_syncbn(m, x, y):
x = x[0]
nelements = x.numel()
m.total_ops += int(2 * nelements)
_, c, h, w = images_src.shape
flops = paddle.flops(
self.model, [1, c, h, w],
custom_ops={paddle.nn.SyncBatchNorm: count_syncbn})
# Sleep for half a second to let dataloader release resources.
time.sleep(0.5)
if use_vdl:
log_writer.close()