-
Notifications
You must be signed in to change notification settings - Fork 5
/
train.py
351 lines (292 loc) · 13.1 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
import datetime
import os
import time
import torch
import torch.utils.data
import wandb
import cv2
import random
import transforms as T
import utils
import numpy as np
import gc
import operator
from functools import reduce
from bert.modeling_bert import BertModel
from lib import segmentation
from loss.loss import Loss
def seed_everything(seed=2401):
random.seed(seed)
os.environ['PYTHONHASHSEED'] = str(seed)
np.random.seed(seed)
torch.manual_seed(seed)
torch.cuda.manual_seed(seed)
torch.cuda.manual_seed_all(seed)
torch.backends.cudnn.benchmark = False
torch.backends.cudnn.deterministic = True
def get_dataset(image_set, transform, args):
from data.dataset_refer_bert import ReferDataset
ds = ReferDataset(args,
split=image_set,
image_transforms=transform,
target_transforms=None
)
num_classes = 2
return ds, num_classes
def IoU(pred, gt):
pred = pred.argmax(1)
intersection = torch.sum(torch.mul(pred, gt))
union = torch.sum(torch.add(pred, gt)) - intersection
if intersection == 0 or union == 0:
iou = 0
else:
iou = float(intersection) / float(union)
return iou, intersection, union
def get_transform(args):
transforms = [
T.Resize(args.img_size, args.img_size),
T.ToTensor(),
T.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
]
return T.Compose(transforms)
def criterion(input, target, weight=0.1):
return Loss(weight=weight)(input, target)
def evaluate(model, data_loader, bert_model, epoch):
model.eval()
metric_logger = utils.MetricLogger(delimiter=" ")
header = "Test: "
total_its = 0
acc_ious = 0
# evaluation variables
cum_I, cum_U = 0, 0
eval_seg_iou_list = [.5, .6, .7, .8, .9]
seg_correct = np.zeros(len(eval_seg_iou_list), dtype=np.int32)
seg_total = 0
mean_IoU = []
total_loss = 0
with torch.no_grad():
for data in metric_logger.log_every(data_loader, 100, header):
total_its += 1
image, target, sentences, attentions = data
pixels = cv2.countNonZero(target.data.numpy()[0]) / 230400.
image, target, sentences, attentions = image.cuda(non_blocking=True),\
target.cuda(non_blocking=True),\
sentences.cuda(non_blocking=True),\
attentions.cuda(non_blocking=True)
sentences = sentences.squeeze(1)
attentions = attentions.squeeze(1)
if bert_model is not None:
last_hidden_states = bert_model(sentences, attention_mask=attentions)[0]
embedding = last_hidden_states.permute(0, 2, 1) # (B, 768, N_l) to make Conv1d happy
attentions = attentions.unsqueeze(dim=-1) # (B, N_l, 1)
output = model(image, embedding, l_mask=attentions)
else:
output = model(image, sentences, l_mask=attentions)
iou, I, U = IoU(output, target)
loss = criterion(output, target)
total_loss += loss.item()
acc_ious += iou
mean_IoU.append(iou)
cum_I += I
cum_U += U
for n_eval_iou in range(len(eval_seg_iou_list)):
eval_seg_iou = eval_seg_iou_list[n_eval_iou]
seg_correct[n_eval_iou] += (iou >= eval_seg_iou)
seg_total += 1
iou = acc_ious / total_its
mean_IoU = np.array(mean_IoU)
mIoU = np.mean(mean_IoU)
print('Final results:')
print('Mean IoU is %.2f\n' % (mIoU * 100.))
results_str = ''
for n_eval_iou in range(len(eval_seg_iou_list)):
results_str += ' precision@%s = %.2f\n' % \
(str(eval_seg_iou_list[n_eval_iou]), seg_correct[n_eval_iou] * 100. / seg_total)
results_str += ' overall IoU = %.2f\n' % (cum_I * 100. / cum_U)
print(results_str)
if args.local_rank == 0:
wandb.log({
"val mIoU": mIoU,
"val oiou": cum_I * 100. / cum_U,
"val Loss": total_loss / total_its})
return 100 * iou, 100 * cum_I / cum_U
def train_one_epoch(model, criterion, optimizer, data_loader, lr_scheduler, epoch, print_freq,
iterations, bert_model):
model.train()
metric_logger = utils.MetricLogger(delimiter=" ")
metric_logger.add_meter('lr', utils.SmoothedValue(window_size=1, fmt='{value}'))
header = 'Epoch: [{}]'.format(epoch)
train_loss = 0
total_its = 0
# for data in data_loader:
for i, data in enumerate(metric_logger.log_every(data_loader, print_freq, header)):
total_its += 1
image, target, sentences, attentions = data
image, target, sentences, attentions = image.cuda(non_blocking=True),\
target.cuda(non_blocking=True),\
sentences.cuda(non_blocking=True),\
attentions.cuda(non_blocking=True)
sentences = sentences.squeeze(1)
attentions = attentions.squeeze(1)
if bert_model is not None:
last_hidden_states = bert_model(sentences, attention_mask=attentions)[0] # (6, 10, 768)
embedding = last_hidden_states.permute(0, 2, 1) # (B, 768, N_l) to make Conv1d happy
attentions = attentions.unsqueeze(dim=-1) # (batch, N_l, 1)
output = model(image, embedding, attentions)#, sentences_hidden_state)# [4,2,120,120]
else:
output = model(image, sentences, attentions)#, sentences_hidden_state)
optimizer.zero_grad()
loss = criterion(output, target)
loss.backward()
optimizer.step()
lr_scheduler.step()
torch.cuda.synchronize()
train_loss += loss.item()
iterations += 1
metric_logger.update(loss=loss.item(), lr=optimizer.param_groups[0]["lr"])
del image, target, sentences, attentions, loss, output, data
if bert_model is not None:
del last_hidden_states, embedding
gc.collect()
torch.cuda.empty_cache()
torch.cuda.synchronize()
if args.local_rank == 0:
wandb.log({
"Train Loss": train_loss / total_its,})
def main(args):
dataset, num_classes = get_dataset("train",
get_transform(args=args),
args=args)
dataset_test, _ = get_dataset("val",
get_transform(args=args),
args=args)
# batch sampler
print(f"local rank {args.local_rank} / global rank {utils.get_rank()} successfully built train dataset.")
num_tasks = utils.get_world_size()
global_rank = utils.get_rank()
train_sampler = torch.utils.data.distributed.DistributedSampler(dataset, num_replicas=num_tasks, rank=global_rank,
shuffle=True)
test_sampler = torch.utils.data.SequentialSampler(dataset_test)
# data loader
data_loader = torch.utils.data.DataLoader(
dataset, batch_size=args.batch_size,
sampler=train_sampler, num_workers=args.workers, pin_memory=args.pin_mem, drop_last=True)
data_loader_test = torch.utils.data.DataLoader(
dataset_test, batch_size=1, sampler=test_sampler, num_workers=args.workers)
# model initialization
print(args.model)
model = segmentation.__dict__[args.model](pretrained=args.pretrained_swin_weights,
args=args)
model = torch.nn.SyncBatchNorm.convert_sync_batchnorm(model)
model.cuda()
model = torch.nn.parallel.DistributedDataParallel(model, device_ids=[args.local_rank], find_unused_parameters=True)
single_model = model.module #ddp
# print(model)
if args.model != 'lavt_one':
model_class = BertModel
bert_model = model_class.from_pretrained(args.ck_bert)
bert_model.pooler = None # a work-around for a bug in Transformers = 3.0.2 that appears for DistributedDataParallel
bert_model.cuda()
bert_model = torch.nn.SyncBatchNorm.convert_sync_batchnorm(bert_model)
bert_model = torch.nn.parallel.DistributedDataParallel(bert_model, device_ids=[args.local_rank])
single_bert_model = bert_model
else:
bert_model = None
single_bert_model = None
# resume training
if args.resume:
checkpoint = torch.load(args.resume, map_location='cpu')
single_model.load_state_dict(checkpoint['model'], strict=False)
if args.model != 'lavt_one':
single_bert_model.load_state_dict(checkpoint['bert_model'])
# parameters to optimize
backbone_no_decay = list()
backbone_decay = list()
for name, m in single_model.backbone.named_parameters():
if 'norm' in name or 'absolute_pos_embed' in name or 'relative_position_bias_table' in name:
backbone_no_decay.append(m)
else:
backbone_decay.append(m)
if args.model != 'lavt_one':
params_to_optimize = [
{'params': backbone_no_decay, 'weight_decay': 0.0},
{'params': backbone_decay},
{"params": [p for p in single_model.classifier.parameters() if p.requires_grad]},
# the following are the parameters of bert
{"params": reduce(operator.concat,
[[p for p in single_bert_model.module.encoder.layer[i].parameters()
if p.requires_grad] for i in range(10)])},
]
else:
params_to_optimize = [
{'params': backbone_no_decay, 'weight_decay': 0.0},
{'params': backbone_decay},
{"params": [p for p in single_model.classifier.parameters() if p.requires_grad]},
# the following are the parameters of bert
{"params": reduce(operator.concat,
[[p for p in single_model.text_encoder.encoder.layer[i].parameters()
if p.requires_grad] for i in range(10)])},
]
# optimizer
optimizer = torch.optim.AdamW(params_to_optimize,
lr=args.lr,
weight_decay=args.weight_decay,
amsgrad=args.amsgrad
)
# learning rate scheduler
lr_scheduler = torch.optim.lr_scheduler.LambdaLR(optimizer,
lambda x: (1 - x / (len(data_loader) * args.epochs)) ** 0.9)
# housekeeping
start_time = time.time()
iterations = 0
best_oIoU = -0.1
# resume training (optimizer, lr scheduler, and the epoch)
if args.resume:
optimizer.load_state_dict(checkpoint['optimizer'])
lr_scheduler.load_state_dict(checkpoint['lr_scheduler'])
resume_epoch = checkpoint['epoch']
else:
resume_epoch = -999
# training loops
if args.local_rank == 0:
wandb.watch(model, log="all")
for epoch in range(max(0, resume_epoch+1), args.epochs):
data_loader.sampler.set_epoch(epoch)
train_one_epoch(model, criterion, optimizer, data_loader, lr_scheduler, epoch, args.print_freq,
iterations, bert_model)
iou, overallIoU = evaluate(model, data_loader_test, bert_model, epoch)
print('Average object IoU {}'.format(iou))
print('Overall IoU {}'.format(overallIoU))
best = (best_oIoU < overallIoU)
if single_bert_model is not None:
dict_to_save = {'model': single_model.state_dict(), 'bert_model': single_bert_model.state_dict(),
'optimizer': optimizer.state_dict(), 'epoch': epoch, 'args': args,
'lr_scheduler': lr_scheduler.state_dict()}
else:
dict_to_save = {'model': single_model.state_dict(),
'optimizer': optimizer.state_dict(), 'epoch': epoch, 'args': args,
'lr_scheduler': lr_scheduler.state_dict()}
if best:
print('Better epoch: {}\n'.format(epoch))
utils.save_on_master(dict_to_save, os.path.join(args.output_dir,
'model_best_{}.pth'.format(args.model_id)))
best_oIoU = overallIoU
utils.save_on_master(dict_to_save, os.path.join(args.output_dir,
'model_last_{}.pth'.format(args.model_id)))
if args.local_rank == 0:
wandb.save('model.h5')
# summarize
total_time = time.time() - start_time
total_time_str = str(datetime.timedelta(seconds=int(total_time)))
print('Training time {}'.format(total_time_str))
if __name__ == "__main__":
from args import get_parser
seed_everything()
parser = get_parser()
args = parser.parse_args()
if args.local_rank == 0:
wandb.init(project="rmsin_2080")
# set up distributed learning
utils.init_distributed_mode(args)
print('Image size: {}'.format(str(args.img_size)))
main(args)