-
Notifications
You must be signed in to change notification settings - Fork 10
/
Copy pathtrain_videosumm.py
243 lines (182 loc) · 12.5 KB
/
train_videosumm.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
import logging
import time
import os
import numpy as np
import torch
import torch.nn as nn
import torch.utils.data
from torch.nn.utils.rnn import pad_sequence
from models import *
from losses import *
from datasets import *
from utils import *
from helpers.bbox_helper import nms
from helpers.vsumm_helper import bbox2summary, get_summ_f1score
logger = logging.getLogger()
def train_videosumm(args, split, split_idx):
batch_time = AverageMeter('time')
data_time = AverageMeter('time')
model = Model_VideoSumm(args=args)
model = model.to(args.device)
calc_contrastive_loss = Dual_Contrastive_Loss().to(args.device)
parameters = [p for p in model.parameters() if p.requires_grad] + \
[p for p in calc_contrastive_loss.parameters() if p.requires_grad]
optimizer = torch.optim.Adam(parameters, lr=args.lr, weight_decay=args.weight_decay)
os.makedirs('{}/checkpoint'.format(args.model_dir), exist_ok=True)
max_train_fscore = -1
max_val_fscore = -1
best_val_epoch = 0
# model testing, load from checkpoint
checkpoint_path = None
if args.checkpoint and args.test:
checkpoint_path = '{}/model_best_split{}.pt'.format(args.checkpoint, split_idx)
checkpoint = torch.load(checkpoint_path, map_location='cpu')
print("load checkpoint from {}".format(checkpoint_path))
model.load_state_dict(checkpoint['model_state_dict'])
train_set = VideoSummDataset(keys=split['train_keys'], args=args)
train_loader = torch.utils.data.DataLoader(train_set, batch_size=args.batch_size, shuffle=True, num_workers=args.num_workers,
drop_last=False, pin_memory=True,
worker_init_fn=worker_init_fn, collate_fn=my_collate_fn)
val_set = VideoSummDataset(keys=split['test_keys'], args=args)
val_loader = torch.utils.data.DataLoader(val_set, batch_size=args.batch_size, shuffle=False, num_workers=args.num_workers,
drop_last=False, pin_memory=True,
worker_init_fn=worker_init_fn, collate_fn=my_collate_fn)
if args.test:
val_fscore = evaluate_videosumm(model, val_loader, args, epoch=0)
logger.info(f'F-score: {val_fscore:.4f}')
return val_fscore, best_val_epoch, max_train_fscore
logger.info('\n' + str(model))
for epoch in range(args.start_epoch, args.max_epoch):
model.train()
stats = AverageMeter('loss', 'cls_loss', 'loc_loss', 'ctr_loss', 'inter_contrastive_loss', 'intra_contrastive_loss')
data_length = len(train_loader)
end = time.time()
for k, (video_list, text_list, mask_video_list, mask_text_list, \
video_cls_label_list, video_loc_label_list, video_ctr_label_list, \
text_cls_label_list, text_loc_label_list, text_ctr_label_list, \
user_summary_list, n_frames_list, ratio_list, n_frame_per_seg_list, picks_list, change_points_list, \
video_to_text_mask_list, text_to_video_mask_list) in enumerate(train_loader):
data_time.update(time=time.time() - end)
batch_size = len(video_list)
video = pad_sequence(video_list, batch_first=True)
text = pad_sequence(text_list, batch_first=True)
mask_video = pad_sequence(mask_video_list, batch_first=True)
mask_text = pad_sequence(mask_text_list, batch_first=True)
video_cls_label = pad_sequence(video_cls_label_list, batch_first=True)
video_loc_label = pad_sequence(video_loc_label_list, batch_first=True)
video_ctr_label = pad_sequence(video_ctr_label_list, batch_first=True)
text_cls_label = pad_sequence(text_cls_label_list, batch_first=True)
text_loc_label = pad_sequence(text_loc_label_list, batch_first=True)
text_ctr_label = pad_sequence(text_ctr_label_list, batch_first=True)
for i in range(len(video_to_text_mask_list)):
video_to_text_mask_list[i] = video_to_text_mask_list[i].to(args.device)
text_to_video_mask_list[i] = text_to_video_mask_list[i].to(args.device)
video, text = video.to(args.device), text.to(args.device)
mask_video, mask_text = mask_video.to(args.device), mask_text.to(args.device)
video_cls_label = video_cls_label.to(args.device) #[B, T]
video_loc_label = video_loc_label.to(args.device) #[B, T, 2]
video_ctr_label = video_ctr_label.to(args.device) #[B, T]
text_cls_label = text_cls_label.to(args.device) #[B, T]
text_loc_label = text_loc_label.to(args.device) #[B, T, 2]
text_ctr_label = text_ctr_label.to(args.device) #[B, T]
video_pred_cls, video_pred_loc, video_pred_ctr, text_pred_cls, text_pred_loc, text_pred_ctr, contrastive_pairs = \
model(video=video, text=text, mask_video=mask_video, mask_text=mask_text,
video_label=video_cls_label, text_label=text_cls_label,
video_to_text_mask_list=video_to_text_mask_list, text_to_video_mask_list=text_to_video_mask_list)
cls_loss = calc_cls_loss(video_pred_cls, video_cls_label.to(torch.long), mask=mask_video) + \
calc_cls_loss(text_pred_cls, text_cls_label.to(torch.long), mask=mask_text)
loc_loss = calc_loc_loss(video_pred_loc, video_loc_label, video_cls_label) + \
calc_loc_loss(text_pred_loc, text_loc_label, text_cls_label)
ctr_loss = calc_ctr_loss(video_pred_ctr, video_ctr_label, video_cls_label) + \
calc_ctr_loss(text_pred_ctr, text_ctr_label, text_cls_label)
inter_contrastive_loss, intra_contrastive_loss = calc_contrastive_loss(contrastive_pairs)
inter_contrastive_loss = inter_contrastive_loss * args.lambda_contrastive_inter
intra_contrastive_loss = intra_contrastive_loss * args.lambda_contrastive_intra
loss = cls_loss + loc_loss + ctr_loss + inter_contrastive_loss + intra_contrastive_loss
optimizer.zero_grad()
loss.backward()
optimizer.step()
stats.update(loss=loss.item(), cls_loss=cls_loss.item(),
loc_loss=loc_loss.item(), ctr_loss=ctr_loss.item(),
inter_contrastive_loss=inter_contrastive_loss.item(),
intra_contrastive_loss=intra_contrastive_loss.item())
batch_time.update(time=time.time() - end)
end = time.time()
if (k + 1) % args.print_freq == 0:
logger.info(f'[Train] Epoch: {epoch+1}/{args.max_epoch} Iter: {k+1}/{data_length} '
f'Time: {batch_time.time:.3f} Data: {data_time.time:.3f} '
f'Loss: {stats.cls_loss:.4f}/{stats.loc_loss:.4f}/{stats.ctr_loss:.4f}/{stats.inter_contrastive_loss:.4f}/{stats.intra_contrastive_loss:.4f}/{stats.loss:.4f}')
save_checkpoint = {
'epoch': epoch+1,
'model_state_dict': model.state_dict(),
'max_val_fscore': max_val_fscore,
'max_train_fscore': max_train_fscore,
}
if (epoch + 1) % args.eval_freq == 0:
val_fscore = evaluate_videosumm(model, val_loader, args, epoch=epoch)
if max_val_fscore < val_fscore:
max_val_fscore = val_fscore
best_val_epoch = epoch + 1
torch.save(save_checkpoint, '{}/checkpoint/model_best_split{}.pt'.format(args.model_dir, split_idx))
logger.info(f'[Eval] Epoch: {epoch+1}/{args.max_epoch} F-score: {val_fscore:.4f}/{max_val_fscore:.4f}\n\n')
args.writer.add_scalar(f'Split{split_idx}/Val/max_fscore', max_val_fscore, epoch+1)
args.writer.add_scalar(f'Split{split_idx}/Val/fscore', val_fscore, epoch+1)
args.writer.add_scalar(f'Split{split_idx}/Train/loss', stats.loss, epoch+1)
args.writer.add_scalar(f'Split{split_idx}/Train/cls_loss', stats.cls_loss, epoch+1)
args.writer.add_scalar(f'Split{split_idx}/Train/loc_loss', stats.loc_loss, epoch+1)
args.writer.add_scalar(f'Split{split_idx}/Train/ctr_loss', stats.ctr_loss, epoch+1)
args.writer.add_scalar(f'Split{split_idx}/Train/inter_contrastive_loss', stats.inter_contrastive_loss, epoch+1)
args.writer.add_scalar(f'Split{split_idx}/Train/intra_contrastive_loss', stats.intra_contrastive_loss, epoch+1)
return max_val_fscore, best_val_epoch, max_train_fscore
@torch.no_grad()
def evaluate_videosumm(model, val_loader, args, epoch=None):
model.eval()
stats = AverageMeter('fscore')
data_length = len(val_loader)
with torch.no_grad():
for k, (video_list, text_list, mask_video_list, mask_text_list, \
video_cls_label_list, video_loc_label_list, video_ctr_label_list, \
text_cls_label_list, text_loc_label_list, text_ctr_label_list, \
user_summary_list, n_frames_list, ratio_list, n_frame_per_seg_list, picks_list, change_points_list, \
video_to_text_mask_list, text_to_video_mask_list) in enumerate(val_loader):
batch_size = len(video_list)
video = pad_sequence(video_list, batch_first=True)
text = pad_sequence(text_list, batch_first=True)
mask_video = pad_sequence(mask_video_list, batch_first=True)
mask_text = pad_sequence(mask_text_list, batch_first=True)
video_cls_label = pad_sequence(video_cls_label_list, batch_first=True)
video_loc_label = pad_sequence(video_loc_label_list, batch_first=True)
video_ctr_label = pad_sequence(video_ctr_label_list, batch_first=True)
text_cls_label = pad_sequence(text_cls_label_list, batch_first=True)
text_loc_label = pad_sequence(text_loc_label_list, batch_first=True)
text_ctr_label = pad_sequence(text_ctr_label_list, batch_first=True)
for i in range(len(video_to_text_mask_list)):
video_to_text_mask_list[i] = video_to_text_mask_list[i].to(args.device)
text_to_video_mask_list[i] = text_to_video_mask_list[i].to(args.device)
video, text = video.to(args.device), text.to(args.device)
mask_video, mask_text = mask_video.to(args.device), mask_text.to(args.device)
video_cls_label = video_cls_label.to(args.device) #[B, T]
video_loc_label = video_loc_label.to(args.device) #[B, T, 2]
video_ctr_label = video_ctr_label.to(args.device) #[B, T]
text_cls_label = text_cls_label.to(args.device) #[B, T]
text_loc_label = text_loc_label.to(args.device) #[B, T, 2]
text_ctr_label = text_ctr_label.to(args.device) #[B, T]
pred_cls_batch, pred_bboxes_batch = model.predict(video=video, text=text,
mask_video=mask_video, mask_text=mask_text,
video_label=video_cls_label, text_label=text_cls_label,
video_to_text_mask_list=video_to_text_mask_list,
text_to_video_mask_list=text_to_video_mask_list) #[B, T], [B, T, 2]
mask_video_bool = mask_video.cpu().numpy().astype(bool)
for i in range(batch_size):
video_length = np.sum(mask_video_bool[i])
pred_cls = pred_cls_batch[i, mask_video_bool[i]] #[T]
pred_bboxes = np.clip(pred_bboxes_batch[i, mask_video_bool[i]], 0, video_length).round().astype(np.int32) #[T, 2]
pred_cls, pred_bboxes = nms(pred_cls, pred_bboxes, args.nms_thresh)
pred_summ, pred_summ_upsampled, pred_score, pred_score_upsampled = bbox2summary(
video_length, pred_cls, pred_bboxes, change_points_list[i], n_frames_list[i], n_frame_per_seg_list[i], picks_list[i], proportion=ratio_list[i], seg_score_mode='mean')
eval_metric = 'max' if args.dataset == 'SumMe' else 'avg'
fscore = get_summ_f1score(pred_summ_upsampled, user_summary_list[i], eval_metric=eval_metric)
stats.update(fscore=fscore)
if (k + 1) % args.print_freq == 0:
logger.info(f'[Eval] Epoch: {epoch+1}/{args.max_epoch} Iter: {k+1}/{data_length} F-score: {stats.fscore:.4f}')
return stats.fscore