-
Notifications
You must be signed in to change notification settings - Fork 2
/
train_cma_recon.py
405 lines (337 loc) · 16.4 KB
/
train_cma_recon.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
import os
import os.path as osp
import math
import torch
import wandb
import shutil
import torch.nn as nn
import torch.nn.parallel
import torch.backends.cudnn as cudnn
import torch.nn.functional as F
import torch.cuda.amp
import torchvision.transforms as transforms
from einops import rearrange
from tqdm import tqdm
import data
from model.encoders import ImageTextEncodersRecon
from model.cross_modal_attention import CrossModalAttentionRecon
from loss.cma_loss import CMA_Loss, CMA_Loss_Fast
from utils import AverageMeter, set_seed
from option import parser, verify_input_args
from sync_batchnorm import convert_model, SynchronizedBatchNorm2d
from eval_cma_recon import encode_data, compute_mask_IU
def train(epoch, total_iter, data_loader, model, criterion, recon_criterion, recon_weight,
optimizer, scaler, recon_warm, args, scheduler = None, bertemb_dict = None):
# switch to train mode
model.train()
if args.bn_eval:
modules = model.module.modules() if args.multi_gpu else model.modules()
for m in modules:
if isinstance(m, nn.BatchNorm2d) or isinstance(m, SynchronizedBatchNorm2d):
m.eval()
# average meters to record the training statistics
losses = AverageMeter()
stat_dict = dict()
losses_dict = dict()
losses_dict['cm_loss'] = AverageMeter()
losses_dict['recon'] = AverageMeter()
for itr, data in enumerate(data_loader):
total_iter += 1
if args.fast_batch:
img, txt, txt_len, recovery, num_txts_per_img, ids = data
img, txt, txt_len, recovery, num_txts_per_img = \
img.cuda(), txt.cuda(), txt_len.cuda(), recovery.cuda(), num_txts_per_img.cuda()
else:
img, txt, txt_len, ids = data
img, txt, txt_len = img.cuda(), txt.cuda(), txt_len.cuda()
with torch.cuda.amp.autocast(enabled=args.amp):
cm_feat, img_emb, txt_emb, img_feat_recon, img_feat, txt_bert = model.forward(img, txt, txt_len)
# Use pre-extracted text embedding from external LM for sampling
if args.pre_bertemb:
bertemb_list = []
for i, idx in enumerate(ids):
sentence, _, _, _, _ = data_loader.dataset.get_raw_item(idx)
bertemb_list.append(bertemb_dict[sentence])
txt_bert = torch.stack(bertemb_list)
if recon_warm:
loss, loss_dict = 0, {}
else:
if args.fast_batch:
txt_emb, cm_feat = txt_emb[recovery], cm_feat[:, recovery, :]
loss, loss_dict = criterion(cm_feat, txt_emb, num_txts_per_img, txt_bert=txt_bert)
else:
loss, loss_dict = criterion(cm_feat, txt_emb, img_emb, txt_bert=txt_bert)
recon_loss = recon_criterion(img_feat_recon, img_feat.detach())
loss_dict['recon'] = recon_loss
loss = loss + recon_weight * recon_loss
if total_iter < args.lr_warmup_iter:
loss *= float(total_iter) / args.lr_warmup_iter
if torch.isnan(loss).any():
print("!! NaN loss detected !!")
import ipdb; ipdb.set_trace()
losses.update(loss)
for key, val in loss_dict.items():
losses_dict[key].update(val)
# Backprop
optimizer.zero_grad()
scaler.scale(loss).backward()
if args.grad_clip > 0:
scaler.unscale_(optimizer)
nn.utils.clip_grad.clip_grad_norm_(model.parameters(), args.grad_clip)
scaler.step(optimizer)
scaler.update()
wandb.log({'iter':total_iter}) if not args.no_wandb else None
if scheduler is not None and total_iter >= args.lr_warmup_iter:
scheduler.step()
# Print log info
if itr > 0 and (itr % args.log_step == 0 or itr + 1 == len(data_loader)):
log_msg = 'loss: %.4f (%.4f)' %(losses.val, losses.avg)
for key, val in losses_dict.items():
log_msg += ', %s: %.4f, (%.4f)' %(key.replace('_loss',''), val.val, val.avg)
n = int(math.ceil(math.log(len(data_loader) + 1, 10)))
print('[%d][%*d/%d] %s' %(epoch, n, itr, len(data_loader), log_msg))
del cm_feat, txt_emb, loss
return losses.avg, losses_dict, stat_dict, total_iter
def validation(epoch, data_loader, model, criterion, recon_criterion, recon_weight, args):
with torch.no_grad():
#NOTE Compute loss on the validation split
losses = AverageMeter()
losses_dict = dict()
losses_dict['cm_loss'] = AverageMeter()
losses_dict['recon'] = AverageMeter()
for _, data in tqdm(enumerate(data_loader)):
img, txt, txt_len, _ = data
img, txt, txt_len = img.cuda(), txt.cuda(), txt_len.cuda()
with torch.cuda.amp.autocast(enabled=args.amp):
cm_feat, img_emb, txt_emb, img_feat_recon, img_feat, txt_bert = model.forward(img, txt, txt_len)
loss, loss_dict = criterion(cm_feat, txt_emb, img_emb, txt_bert=txt_bert)
recon_loss = recon_criterion(img_feat_recon, img_feat.clone().detach())
loss_dict['recon'] = recon_loss
loss = loss + recon_weight * recon_loss
if torch.isnan(loss).any():
print("!! NaN loss detected !!")
import ipdb; ipdb.set_trace()
losses.update(loss)
for key, val in loss_dict.items():
losses_dict[key].update(val)
del img, txt, txt_len, cm_feat, img_emb, txt_emb, img_feat_recon, img_feat
log_msg = 'loss: %.4f (%.4f)' %(losses.val, losses.avg)
for key, val in losses_dict.items():
log_msg += ', %s: %.4f, (%.4f)' %(key.replace('_loss',''), val.val, val.avg)
print('Epoch [%d] validation: %s' %(epoch, log_msg))
# Compute cumulative IoU on the validation split. Considering two different pseudo labelling policy
cum_max_I, cum_max_U = 0., 0.
cum_avg_I, cum_avg_U = 0., 0.
avg_mIoU, max_mIoU = 0., 0.
max_precision, avg_precision = 0., 0.
max_recall, avg_recall = 0., 0.
dataset = data_loader.dataset
_, _, _, img_a_map, cm_a_map, _ = encode_data(
model,
data_loader,
crop_size=args.crop_size,
img_num_embeds=args.img_num_embeds,
embed_dim=args.embed_dim,
args=args
)
t= tqdm(range(len(dataset)), desc='Evaluating', leave=True)
for i in t:
_, raw_img_id, _, raw_label, _ = dataset.get_raw_item(i)
feat_map_size = int(args.crop_size / 16)
cm_a = cm_a_map[i]
top_slot_idx = torch.argmax(cm_a)
# Pseudo label: Attention map of the closest slot
img_a_map_for_slot = img_a_map[i, -1, top_slot_idx]
a_map = img_a_map_for_slot.reshape(1, feat_map_size, feat_map_size)
a_map = transforms.functional.resize(a_map, list(raw_label.shape)).squeeze().cpu().numpy()
a_map = ((a_map - a_map.min()) / (a_map.max() - a_map.min() + 1e-9))
hard_pred = (a_map >= args.pseudo_threshold)
I, U = compute_mask_IU(hard_pred, raw_label)
cum_max_I += I
cum_max_U += U
max_precision += (I / hard_pred.sum() if I != 0 else 0)
max_recall += I / raw_label.sum()
max_mIoU += I / U
# Pseudo label: Weighted sum of the attention maps with similarity scores between slots
avg_a_map = img_a_map[i, -1, :]
avg_a_map = cm_a.unsqueeze(1) * avg_a_map
avg_a_map = avg_a_map.sum(dim=0)
avg_a_map = avg_a_map.reshape(1, feat_map_size, feat_map_size)
avg_a_map = transforms.functional.resize(avg_a_map, list(raw_label.shape)).squeeze().cpu().numpy()
avg_a_map = ((avg_a_map - avg_a_map.min()) / (avg_a_map.max() - avg_a_map.min() + 1e-9))
hard_avg_pred = (avg_a_map >= args.pseudo_threshold)
I, U = compute_mask_IU(hard_avg_pred, raw_label)
cum_avg_I += I
cum_avg_U += U
avg_precision += (I / hard_avg_pred.sum() if I != 0 else 0)
avg_recall += I / raw_label.sum()
avg_mIoU += I / U
t.set_postfix({
"max | avg":" %.3f%% | %.3f%% " \
% (100*(cum_max_I/cum_max_U), 100*(cum_avg_I/cum_avg_U))
})
val_dict = {
'max_cIoU': 100 * (cum_max_I/cum_max_U),
'max_mIoU': max_mIoU * (100/len(dataset)),
'max_precision': max_precision * (100/len(dataset)),
'max_recall': max_recall * (100/len(dataset)),
'avg_cIoU': 100 * (cum_avg_I/cum_avg_U),
'avg_mIoU': avg_mIoU * (100/len(dataset)),
'avg_precision': avg_precision * (100/len(dataset)),
'avg_recall': avg_recall * (100/len(dataset))
}
del img_a_map, avg_a_map, a_map
del hard_pred, img_a_map_for_slot, hard_avg_pred
return losses.avg, losses_dict, val_dict
def update_best_score(new_score, old_score, is_higher_better):
if not old_score:
score, updated = new_score, True
else:
if is_higher_better:
score = max(new_score, old_score)
updated = new_score > old_score
else:
score = min(new_score, old_score)
updated = new_score < old_score
return score, updated
def warmup(model, epoch, args, multi_gpu):
if args.img_finetune and args.txt_finetune:
warm = epoch >= args.warm_epoch
if args.warm_img:
for idx, param in enumerate((model.module if multi_gpu else model).encoders.img_enc.img_backbone.parameters()):
param.requires_grad = warm
if args.warm_txt:
for idx, param in enumerate((model.module if multi_gpu else model).encoders.txt_enc.bert.parameters()):
param.requires_grad = warm
def save_ckpt(state, is_best, filename='ckpt.pth.tar', prefix=''):
torch.save(state, prefix + filename)
if is_best:
shutil.copyfile(prefix + filename, prefix + 'model_best.pth.tar')
print('Updating the best model checkpoint: {}'.format(prefix + 'model_best.pth.tar'))
def main():
args = verify_input_args(parser.parse_args())
set_seed(args.seed)
if not args.no_wandb:
wandb.init(project='weak_ref_seg', name = args.remark, group=args.wandb_group, entity='ise', config=args)
wandb.config.update(args)
log_dir = osp.join('./logs', args.remark)
if not osp.exists(log_dir):
os.makedirs(log_dir)
# Dataloaders
trn_loader = data.get_train_loader(args)
test_loader = data.get_test_loader(args)
# Construct the model
# NOTE CMA model
model = CrossModalAttentionRecon(ImageTextEncodersRecon(args), args.embed_dim, args)
if torch.cuda.is_available():
if args.multi_gpu:
model = nn.DataParallel(model, output_device=1)
if args.sync_bn:
model = convert_model(model)
model = model.cuda()
cudnn.benchmark = True
wandb.watch(models=model, log_freq=1000, log='gradients') if not args.no_wandb else None
# Loss and optimizer
recon_criterion = nn.MSELoss()
recon_weight = args.recon_weight
criterion = (CMA_Loss_Fast if args.fast_batch else CMA_Loss)(
margin=args.margin,
criterion=args.cma_criterion,
mining=args.cma_mining,
detach_target=args.cma_detach_target,
detach_img_target=args.cma_detach_img_target,
i_t_loss=None,
i_t_weight=args.i_t_weight,
temperature=args.info_temperature,
cm_i_weight=args.cm_i_weight,
size_p_loss=None,
size_p_weight=args.size_p_weight,
)
val_criterion = CMA_Loss(
margin=args.margin,
criterion=args.cma_criterion,
mining=args.cma_mining,
detach_target=args.cma_detach_target,
detach_img_target=args.cma_detach_img_target,
i_t_weight=args.i_t_weight,
temperature=args.info_temperature,
cm_i_weight=args.cm_i_weight,
size_p_weight=args.size_p_weight,
)
module = model.module if args.multi_gpu else model
param_groups = [
{'params': module.cma.parameters(), 'lr': args.lr},
{'params': module.decoder.parameters(), 'lr': args.lr},
{'params': list(set(module.encoders.img_enc.parameters()).difference(set(module.encoders.img_enc.img_backbone.parameters()))),
'lr': args.lr * args.img_spm_lr_scale},
{'params': module.encoders.img_enc.img_backbone.parameters(), 'lr': args.lr * args.img_lr_scale},
{'params': list(set(module.encoders.txt_enc.parameters())), 'lr': args.lr * args.txt_lr_scale}
]
if args.optimizer == 'adam':
optimizer = torch.optim.Adam(param_groups, lr=args.lr, weight_decay=args.weight_decay, amsgrad=True)
elif args.optimizer == 'adamw':
optimizer = torch.optim.AdamW(param_groups, lr=args.lr, weight_decay=args.weight_decay)
elif args.optimizer == 'adamp':
from adamp import AdamP
optimizer = AdamP(param_groups, lr=args.lr, weight_decay=args.weight_decay)
if args.lr_scheduler == 'cosine':
lr_scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(optimizer, T_max=len(trn_loader)*args.num_epochs)
elif args.lr_scheduler == 'multi_step':
lr_scheduler = torch.optim.lr_scheduler.MultiStepLR(optimizer, milestones=args.lr_milestones, gamma = args.lr_step_gamma)
elif args.lr_scheduler == 'step':
lr_scheduler = torch.optim.lr_scheduler.StepLR(optimizer, step_size=args.lr_step_size, gamma = args.lr_step_gamma)
# AMP
scaler = torch.cuda.amp.GradScaler(enabled=args.amp)
total_iter = 0
best_loss = 10000
best_iou = 0
# Pre-compute bertemb
bertemb_dict = None
if args.pre_bertemb:
bertemb_dict = torch.load(osp.join(osp.dirname(args.data_path), osp.basename(args.data_path)+'_bertemb.pt'))
for epoch in range(args.num_epochs):
#warm up training data
warmup(model, epoch, args, args.multi_gpu)
recon_weight = args.recon_weight if epoch >= args.wo_recon_epoch else .0
# train for one epoch
loss, losses_dict, stat_dict, total_iter = train(
epoch, total_iter, trn_loader, model, criterion, recon_criterion, recon_weight, optimizer, scaler,
epoch < args.recon_warm_epoch,
args,
scheduler=lr_scheduler if args.lr_scheduler == 'cosine' else None,
bertemb_dict = bertemb_dict
)
# Compute validation loss
val_loss, val_losses_dict, val_dict = \
validation(epoch, test_loader, model, val_criterion, recon_criterion, recon_weight, args)
print(val_dict)
if not args.no_wandb:
wandb.log({"epoch": epoch}, step=total_iter)
wandb.log({"loss": loss}, step=total_iter)
for key, val in losses_dict.items():
wandb.log({key: val.avg}, step=total_iter)
for key, val in stat_dict.items():
wandb.log({key: val.avg}, step=total_iter)
wandb.log({"LR" : optimizer.param_groups[0]['lr']}, step=total_iter)
wandb.log({"val loss": val_loss}, step=total_iter)
for key, val in val_losses_dict.items():
wandb.log({"val "+key: val.avg}, step=total_iter)
for key, val in val_dict.items():
wandb.log({key: val}, step=total_iter)
# evaluate on validation set
with torch.no_grad():
# remember best rsum and save ckpt
# best_loss, updated = update_best_score(loss, best_loss, is_higher_better=False)
best_iou, best_updated = update_best_score(val_dict['avg_mIoU'], best_iou, is_higher_better=True)
save_ckpt({
'args': args,
'epoch': epoch,
'iou': val_dict['avg_mIoU'],
'loss': loss,
'model': model.state_dict(),
}, best_updated, prefix=log_dir + '/')
# adjust learning rate if rsum stagnates
if args.lr_scheduler != 'cosine':
lr_scheduler.step()
if __name__ == '__main__':
main()