-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathpublic_worker.py
366 lines (325 loc) · 17.6 KB
/
public_worker.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
import warnings
warnings.filterwarnings("ignore")
from torch.optim import lr_scheduler
import logging
from torch import optim
# from option import get_args
from options.Base_option import Base_options
# from datasets.data_util import *
from utils.util import *
import os
import torch.distributed as dist
import torch.multiprocessing as mp
from torch.cuda.amp import GradScaler
from utils.data_aug import random_crop
from utils.eval import *
from tqdm import tqdm
from tensorboardX import SummaryWriter
from torch.utils.data import Dataset, DataLoader
# load yaml file by mode
base_option = Base_options()
args = base_option.get_args()
yamls_dict = get_yaml_data('./config/' + args.model + '_config.yaml')
set_yaml_to_args(args, yamls_dict)
os.environ['CUDA_VISIBLE_DEVICES'] = args.gpu
os.environ['PYTHONWARNINGS'] = 'ignore:semaphore_tracker:UserWarning'
# os.environ['CUDA_LAUNCH_BLOCKING'] = '1'
# os.environ['CUDA_LAUNCH_BLOCKING'] = '1'
# os.environ['CUDA_CACHE_PATH'] = '~/.cudacache'
def main_worker(rank, n_pros):
if not os.path.exists('./checkSave/'):
os.makedirs('./checkSave/')
logging.basicConfig(filename=args.log_path + 'log.txt', level=logging.INFO,
format='[%(asctime)s.%(msecs)03d] %(message)s', datefmt='%H:%M:%S')
writer = SummaryWriter(args.log_path)
torch.cuda.set_device(rank)
dist.init_process_group(backend='nccl', init_method='tcp://127.0.0.1:{}'.format(args.port), world_size=n_pros,
rank=rank)
# get the dataset dynamically
dataset = getPackByNameUtil(py_name='datasets.' + args.loader_mode + '_dataset',
object_name=args.loader_mode + '_Dataset')
# get the trainer dynamically
train_iter = getPackByNameUtil(py_name='trainers.' + args.model + '_trainer',
object_name=args.model + '_Trainer')
# get the evaluaters dynamically
try:
evaluater_iter = getPackByNameUtil(py_name='evaluaters.' + args.model + '_evaluater',
object_name=args.model + '_Evaluater')
shower_iter = getPackByNameUtil(py_name='evaluaters.' + args.model + '_evaluater',
object_name=args.model + '_Shower')
except:
try:
evaluater_iter = getPackByNameUtil(py_name='evaluaters.' + args.loader_mode + '_evaluater',
object_name=args.loader_mode + '_Evaluater')
shower_iter = getPackByNameUtil(py_name='evaluaters.' + args.loader_mode + '_evaluater',
object_name=args.loader_mode + '_Shower')
except:
evaluater_iter = getPackByNameUtil(py_name='evaluaters.Base_evaluater',
object_name='Base_Evaluater')
shower_iter = getPackByNameUtil(py_name='evaluaters.Base_evaluater',
object_name='Base_Shower')
# get the model dynamically
model = getPackByNameUtil(py_name='models.' + args.model + '_net',
object_name=args.model + '_Net')
label_dataSet = dataset(args, mode='train')
val_dataset = dataset(args, mode='val')
show_dataset = dataset(args, mode='show')
label_train_sampler = torch.utils.data.distributed.DistributedSampler(label_dataSet)
val_sampler = torch.utils.data.distributed.DistributedSampler(val_dataset)
temp_show_sampler = torch.utils.data.distributed.DistributedSampler(show_dataset)
label_train_loader = DataLoader(label_dataSet,
shuffle=False,
num_workers=args.num_worker,
batch_size=args.batch_size,
pin_memory=True,
sampler=label_train_sampler
)
val_loader = DataLoader(val_dataset,
shuffle=False,
num_workers=1,
batch_size=1,
pin_memory=True,
sampler=val_sampler)
temp_show_loader = DataLoader(show_dataset,
shuffle=False,
num_workers=1,
batch_size=1,
pin_memory=True,
sampler=temp_show_sampler)
net = model(args).to(rank)
net = torch.nn.SyncBatchNorm.convert_sync_batchnorm(net)
net = torch.nn.parallel.DistributedDataParallel(net,
device_ids=[rank],
output_device=rank,
find_unused_parameters=True)
# torch.autograd.set_detect_anomaly(True)
# load pretrained parameters
if args.pretrain_path != '' and args.pretrain_path is not None:
print("loading from {}".format(args.pretrain_path))
map_location = {'cuda:%d' % 0: 'cuda:%d' % rank}
saved_state_dict = torch.load(args.pretrain_path, map_location='cpu')#['state_dict']
new_params = net.state_dict().copy()
for name, param in new_params.items():
if (name in saved_state_dict and param.size() == saved_state_dict[name].size()):
new_params[name].copy_(saved_state_dict[name])
# print(name)
elif name[7:] in saved_state_dict and param.size() == saved_state_dict[name[7:]].size():
new_params[name].copy_(saved_state_dict[name[7:]])
# print(name[7:])
elif 'module.' + name in saved_state_dict and param.size() == saved_state_dict['module.' + name].size():
new_params[name].copy_(saved_state_dict['module.' + name])
# print('module.' + name)
else:
print(name)
net.load_state_dict(new_params)
optimizer = optim.Adam(net.parameters(), lr=args.lr, betas=(0.5, 0.999))
"""Build cosine learning rate scheduler."""
scheduler = lr_scheduler.CosineAnnealingLR(optimizer,
T_max=(args.epoch - args.warmup_step) * len(label_train_loader),
eta_min=1e-6)
scaler = GradScaler()
max_val_loss = 100000
step = 0
for epoch in range(args.epoch):
label_train_sampler.set_epoch(epoch)
loop = tqdm(enumerate(label_train_loader), total=len(label_train_loader), position=rank)
loss_epoch = 0
for (i, label_data) in loop:
if epoch < args.warmup_step:
cur_lr = args.lr * step / (args.warmup_step * len(label_train_loader))
for param_group in optimizer.param_groups:
param_group['lr'] = cur_lr
else:
scheduler.step()
merge_img = label_data[0]
merge_alpha = label_data[1]
merge_trimap = label_data[2]
merge_fg = label_data[3]
merge_bg = label_data[4]
user_map = label_data[5]
idx = label_data[-1]
if args.aug_crop:
merge_img, merge_alpha, merge_trimap, merge_fg, merge_bg = \
random_crop(merge_img, merge_alpha, merge_trimap, merge_fg, merge_bg)
# mixup
if args.aug_mixup:
lam = np.random.beta(0.4, 0.4)
index = torch.randperm(merge_img.shape[0]).cuda()
merge_img = lam * merge_img + (1 - lam) * merge_img[index]
merge_alpha = lam * merge_alpha + (1 - lam) * merge_alpha[index]
merge_trimap = lam * merge_trimap + (1 - lam) * merge_trimap[index]
merge_fg = lam * merge_fg + (1 - lam) * merge_fg[index]
merge_bg = lam * merge_bg + (1 - lam) * merge_bg[index]
net.train()
optimizer.zero_grad()
loss_dict = train_iter(net,
merge_img.cuda().float(),
merge_trimap.cuda().float(),
merge_alpha.cuda().float(),
user_map=user_map.cuda().float(),
mode=args.model,
fg=merge_fg.cuda().float(),
bg=merge_bg.cuda().float(),
args=args,
epoch=epoch,
cur_step=step,
total_step=args.epoch * len(label_train_loader))
assert 'loss' in loss_dict.keys(), 'The keys of loss dict must include [loss].'
loss = loss_dict['loss']
if args.amp == True:
scaler.scale(loss).backward()
scaler.step(optimizer)
scaler.update()
else:
loss.backward()
optimizer.step()
step += 1
# construct the str of loss items
log_str_pre = args.model + ' epoch:{}/item:{} '.format(epoch, i)
loss_items = loss_dict.items()
loss_epoch += loss.item()
log_str_loss = 'lr:{:.4f} l_e:{:.4f} '.format(optimizer.state_dict()['param_groups'][0]['lr'],
loss_epoch / (i + 1))
# log_str_loss = 'le:{:.4f} '.format(loss_epoch / (i + 1))
for k, l in loss_items:
log_str_loss += k + ':{:.3f} '.format(l.item())
if rank == 0:
logging.info(log_str_pre + log_str_loss)
# tensorboard
writer.add_scalar('training_loss', loss.item(), global_step=step)
loop.set_description(args.model + '|epoch:{}'.format(epoch))
loop.set_postfix_str(log_str_loss)
# scheduler.step()
# validation
if (epoch+1) % args.val_per_epoch == 0 and epoch >= 0:
net.eval()
with torch.no_grad():
error_sad_sum = 0
error_mad_sum = 0
error_mse_sum = 0
error_grad_sum = 0
sad_fg_sum = 0
sad_bg_sum = 0
sad_tran_sum = 0
index = 0
val_loop = tqdm(enumerate(val_loader), total=len(val_loader), position=rank)
val_loop.set_description('val|')
for (i, label_data) in val_loop:
label_img = label_data[0].cuda().float()
label_alpha = label_data[1].cuda().float() # .unsqueeze(1)
trimap = label_data[2].cuda().float().unsqueeze(1)
instance_map = label_data[3].cuda().float()
eval_out = evaluater_iter(net,
label_img,
label_alpha,
trimap,
fusion=args.fusion,
interac=args.inter_num)
error_sad, error_mad, error_mse, error_grad, sad_fg, sad_bg, sad_tran = eval_out[0], eval_out[1], \
eval_out[2], eval_out[3], \
eval_out[4], eval_out[5], \
eval_out[6]
# out = net(label_img)
# matte = out[-1]
# error_sad, error_mad, error_mse, error_grad, sad_fg, sad_bg, sad_tran = computeAllMatrix(matte,
# label_alpha,
# trimap)
index += error_sad - error_sad + 1
error_sad_sum += error_sad
error_mad_sum += error_mad
error_mse_sum += error_mse
error_grad_sum += error_grad
sad_fg_sum += sad_fg
sad_bg_sum += sad_bg
sad_tran_sum += sad_tran
dist.barrier()
dist.all_reduce(error_sad_sum, op=dist.ReduceOp.SUM)
dist.all_reduce(error_mad_sum, op=dist.ReduceOp.SUM)
dist.all_reduce(error_mse_sum, op=dist.ReduceOp.SUM)
dist.all_reduce(error_grad_sum, op=dist.ReduceOp.SUM)
dist.all_reduce(sad_fg_sum, op=dist.ReduceOp.SUM)
dist.all_reduce(sad_bg_sum, op=dist.ReduceOp.SUM)
dist.all_reduce(sad_tran_sum, op=dist.ReduceOp.SUM)
dist.all_reduce(index, op=dist.ReduceOp.SUM)
ave_val_loss = error_mad_sum / index
ave_error_sad_sum = error_sad_sum / index
ave_error_mad_sum = error_mad_sum / index
ave_error_mse_sum = error_mse_sum / index
ave_error_grad_sum = error_grad_sum / index
ave_sad_fg_sum = sad_fg_sum / index
ave_sad_bg_sum = sad_bg_sum / index
ave_sad_tran_sum = sad_tran_sum / index
if rank == 0:
sum_loss = ave_val_loss
if sum_loss < max_val_loss:
max_val_loss = sum_loss
torch.save(net.state_dict(), args.save_path_model + 'model_best')
writer.add_scalar('val_SAD', ave_error_sad_sum, global_step=epoch)
writer.add_scalar('val_MAD', ave_error_mad_sum, global_step=epoch)
writer.add_scalar('val_MSE', ave_error_mse_sum, global_step=epoch)
writer.add_scalar('val_Grad', ave_error_grad_sum, global_step=epoch)
writer.add_scalar('val_SAD_F', ave_sad_fg_sum, global_step=epoch)
writer.add_scalar('val_SAD_B', ave_sad_bg_sum, global_step=epoch)
writer.add_scalar('val_SAD_T', ave_sad_tran_sum, global_step=epoch)
logging.info(
args.model + "|epoch:{} "
"val best_mad_loss:{:.5f} "
"error_sad:{:.5f} "
"error_mad:{:.5f} "
"error_mse:{:.5f} "
"error_grad:{:.5f} "
"sad_fg:{:.5f} "
"sad_bg:{:.5f} "
"sad_tran:{:.5f}"
.format(
epoch,
max_val_loss,
ave_error_sad_sum,
ave_error_mad_sum,
ave_error_mse_sum,
ave_error_grad_sum,
ave_sad_fg_sum,
ave_sad_bg_sum,
ave_sad_tran_sum)
)
metrix_str = '{:20}\t{:20}\t{:20}\n' \
'{:20}\t{:20}\t{:20}\n' \
'{:20}\t{:20}\t{:20}\n' \
.format('Val|epoch: {}'.format(epoch),
'Best_mad: {:.5f}'.format(max_val_loss),
'Grad: {:.5f}'.format(ave_error_grad_sum),
'Sad: {:.5f}'.format(ave_error_sad_sum),
'Mad: {:.5f}'.format(ave_error_mad_sum),
'Mse: {:.5f}'.format(ave_error_mse_sum),
'Sad_fg: {:.5f}'.format(ave_sad_fg_sum),
'Sad_bg: {:.5f}'.format(ave_sad_bg_sum),
'Sad_tran: {:.5f}'.format(ave_sad_tran_sum)
)
print(metrix_str)
if (epoch+1) % args.show_per_epoch == 0 and epoch > 0:
net.eval()
with torch.no_grad():
for (i, label_data) in enumerate(temp_show_loader):
if i == 100:
break
label_img = label_data[0].cuda().float()
label_alpha = label_data[1].cuda().float()
matte, source = shower_iter(net, label_img, label_alpha)
source = (source.cpu().numpy()[0].transpose([1, 2, 0])) * 255
# out = net(label_img.cuda().float())
# matte = out[-1]
# source = (label_img.cuda().float().cpu().numpy()[0].transpose([1, 2, 0])) * 255
source = np.clip(source, 0, 255)[..., :3]
matte = matte[0][0].data.cpu().numpy()
Image.fromarray(((matte * 255).astype('uint8')), mode='L').save(
args.save_path_img + '{}_mask.png'.format(i))
bg = np.zeros(shape=source.shape)
bg[:, :, 1] = 255
merge = source * matte[:, :, np.newaxis] + bg * (1 - matte[:, :, np.newaxis])
merge = np.array(merge, dtype='uint8')
cv2.imwrite(args.save_path_img + '{}.png'.format(i), merge)
if epoch % args.save_per_epoch == 0 and rank == 0 and epoch > 0:
torch.save(net.state_dict(), args.save_path_model + 'model_{}'.format(epoch))
if __name__ == '__main__':
DDP = True
mp.spawn(main_worker, nprocs=len(args.gpu.split(',')), args=(len(args.gpu.split(',')),))