-
Notifications
You must be signed in to change notification settings - Fork 0
/
train.py
427 lines (361 loc) · 20.6 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
import os
import argparse
import json
from lib.config import conf, cfg_from_file
"""------------------------------------some settings----------------------------------------"""
parser = argparse.ArgumentParser()
parser.add_argument('--cfg', dest='cfg_file', help='optional config file', default='configs/demo.yml', type=str)
args = parser.parse_args()
if args.cfg_file is not None:
cfg_from_file(args.cfg_file)
print('The CKPT saved here:', conf.save_path)
if not os.path.exists(conf.save_path):
os.mkdir(conf.save_path)
print('spatial encoder layer num: {} / temporal decoder layer num: {}'.format(conf.enc_layer, conf.dec_layer))
print('-------------student model setting------------------')
print(args.cfg_file)
print(conf)
with open(os.path.join(conf.save_path, "configs.json"), 'w') as f:
json.dump(conf, f)
"""-----------------------------------------------------------------------------------------"""
os.environ['CUDA_VISIBLE_DEVICES'] = str(conf.gpu_id)
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch import log_softmax, optim
from torch.optim.lr_scheduler import ReduceLROnPlateau
import numpy as np
np.set_printoptions(precision=3)
import time
import pandas as pd
import copy
from tqdm import tqdm
torch.set_num_threads(4)
from tensorboardX import SummaryWriter
from dataloader.action_genome import AG, cuda_collate_fn
from lib.object_detector import detector
from lib.evaluation_recall import BasicSceneGraphEvaluator
from lib.AdamW import AdamW
from lib.sttran import STTran
from lib.assign_pseudo_label import prepare_func
from lib.transition_module import transition_module
from lib.ults.track import track, track_diff, track_iou, track_diff_iou
from lib.ults.init_teacher_model import init_teacher_model
AG_dataset_train = AG(mode="train", datasize=conf.datasize, data_path=conf.data_path, ws_object_bbox_path=conf.ws_object_bbox_path, remove_one_frame_video=conf.remove_one_frame_video,
filter_nonperson_box_frame=True, filter_small_box=False if conf.mode == 'predcls' else True)
dataloader_train = torch.utils.data.DataLoader(AG_dataset_train, shuffle=True, num_workers=conf.num_workers,
collate_fn=cuda_collate_fn, pin_memory=False)
AG_dataset_test = AG(mode="test", datasize=conf.datasize, data_path=conf.data_path, ws_object_bbox_path=None, remove_one_frame_video=True,
filter_nonperson_box_frame=True, filter_small_box=False if conf.mode == 'predcls' else True)
dataloader_test = torch.utils.data.DataLoader(AG_dataset_test, shuffle=False, num_workers=conf.num_workers,
collate_fn=cuda_collate_fn, pin_memory=False)
gpu_device = torch.device("cuda")
# freeze the detection backbone
object_detector = detector(train=True, object_classes=AG_dataset_train.object_classes, use_SUPPLY=True, conf=conf).to(device=gpu_device)
object_detector.eval()
if conf.union_box_feature:
faset_rcnn_model, transforms = prepare_func()
else:
faset_rcnn_model = None
transforms = None
model = STTran(mode=conf.mode,
attention_class_num=len(AG_dataset_train.attention_relationships),
spatial_class_num=len(AG_dataset_train.spatial_relationships),
contact_class_num=len(AG_dataset_train.contacting_relationships),
obj_classes=AG_dataset_train.object_classes,
enc_layer_num=conf.enc_layer,
dec_layer_num=conf.dec_layer,
transformer_mode=conf.transformer_mode,
is_wks=conf.is_wks,
feat_dim=conf.feat_dim).to(device=gpu_device)
print("create student model successfully")
print('*'*50)
if conf.teacher_mode_cfg is None:
print("Do not need to create teacher model")
print('*'*50)
else:
t_model = init_teacher_model(conf.teacher_mode_cfg, AG_dataset_train, gpu_device)
print("create teacher model successfully")
print('*'*50)
if conf.ckpt is None:
print('Do not need to load CKPT')
start_epoch = 0
else:
ckpt = torch.load(conf.ckpt, map_location=gpu_device)
model.load_state_dict(ckpt['state_dict'], strict=False)
print('CKPT {} is loaded'.format(conf.ckpt))
left_pos = conf.ckpt.rfind('_')
right_pos = conf.ckpt.rfind('.')
start_epoch = int(conf.ckpt[left_pos+1:right_pos]) + 1
evaluator1 = BasicSceneGraphEvaluator(
mode=conf.mode,
AG_object_classes=AG_dataset_train.object_classes,
AG_all_predicates=AG_dataset_train.relationship_classes,
AG_attention_predicates=AG_dataset_train.attention_relationships,
AG_spatial_predicates=AG_dataset_train.spatial_relationships,
AG_contacting_predicates=AG_dataset_train.contacting_relationships,
iou_threshold=0.5,
constraint='with')
evaluator2 = BasicSceneGraphEvaluator(
mode=conf.mode,
AG_object_classes=AG_dataset_train.object_classes,
AG_all_predicates=AG_dataset_train.relationship_classes,
AG_attention_predicates=AG_dataset_train.attention_relationships,
AG_spatial_predicates=AG_dataset_train.spatial_relationships,
AG_contacting_predicates=AG_dataset_train.contacting_relationships,
iou_threshold=0.5,
constraint='semi', semithreshold=0.9)
evaluator3 = BasicSceneGraphEvaluator(
mode=conf.mode,
AG_object_classes=AG_dataset_train.object_classes,
AG_all_predicates=AG_dataset_train.relationship_classes,
AG_attention_predicates=AG_dataset_train.attention_relationships,
AG_spatial_predicates=AG_dataset_train.spatial_relationships,
AG_contacting_predicates=AG_dataset_train.contacting_relationships,
iou_threshold=0.5,
constraint='no')
# loss function, default Multi-label margin loss
if conf.bce_loss:
ce_loss = nn.CrossEntropyLoss()
# bce_loss = nn.BCEWithLogitsLoss()
bce_loss = nn.BCELoss()
if conf.loss == 'KL':
# kl_loss = nn.KLDivLoss(reduction="batchmean")
kl_loss = nn.KLDivLoss(reduction="sum")
elif conf.loss == 'L1':
L1_loss = nn.L1Loss()
elif conf.loss == 'L2':
L2_loss = nn.MSELoss()
else:
ce_loss = nn.CrossEntropyLoss()
mlm_loss = nn.MultiLabelMarginLoss()
softmax = nn.Softmax(dim=1)
sigmoid = nn.Sigmoid()
if conf.transition_module:
trans_module = transition_module().to(device=gpu_device)
else:
trans_module = None
# optimizer
if conf.optimizer == 'adamw':
# optimizer = AdamW(model.parameters(), lr=conf.lr)
if trans_module is None:
optimizer = AdamW(model.parameters(), lr=conf.lr)
elif trans_module is not None:
optimizer = AdamW([{'params': model.parameters()}, {'params': trans_module.parameters(), 'lr': conf.t_lr}], lr=conf.lr)
# optimizer = AdamW([{'params': model.parameters()}, {'params': trans_module.parameters(), 'lr': conf.lr}], lr=conf.lr)
elif conf.optimizer == 'adam':
optimizer = optim.Adam(model.parameters(), lr=conf.lr)
elif conf.optimizer == 'sgd':
optimizer = optim.SGD(model.parameters(), lr=conf.lr, momentum=0.9, weight_decay=0.01)
scheduler = ReduceLROnPlateau(optimizer, "max", patience=1, factor=0.5, verbose=True, threshold=1e-4, threshold_mode="abs", min_lr=1e-7)
writer = SummaryWriter(conf.tensorboard_name)
# some parameters
tr = []
test_res = {}
save_loss = {}
save_epoch = 1000
for epoch in range(start_epoch, conf.nepoch):
model.train()
object_detector.is_train = True
start = time.time()
train_iter = iter(dataloader_train)
test_iter = iter(dataloader_test)
alpha1_cnt = 0
alpha2_cnt = 0
loss_cnt_tr = []
with tqdm(total=len(dataloader_train)) as t:
for b in range(len(dataloader_train)):
data = next(train_iter)
if conf.is_wks:
im_data = None
im_info = None
gt_boxes = None
num_boxes = None
else:
im_data = copy.deepcopy(data[0].cuda(0))
im_info = copy.deepcopy(data[1].cuda(0))
gt_boxes = copy.deepcopy(data[2].cuda(0))
num_boxes = copy.deepcopy(data[3].cuda(0))
gt_annotation = AG_dataset_train.gt_annotations[data[4]]
frame_names = AG_dataset_train.video_list[data[4]]
# prevent gradients to FasterRCNN
with torch.no_grad():
entry = object_detector(im_data, im_info, gt_boxes, num_boxes, gt_annotation, frame_names, faset_rcnn_model, transforms)
t_entry = copy.deepcopy(entry)
# pass when no relation
# if entry['im_idx'].shape[0] == 0:
if entry != None:
if conf.teacher_mode_cfg is not None:
with torch.no_grad():
t_pred = t_model(t_entry)
pred = model(entry)
# rel_num*3/6/17
attention_distribution = pred["attention_distribution"]
spatial_distribution = pred["spatial_distribution"]
contact_distribution = pred["contacting_distribution"]
object_label = pred['labels']
attention_label = torch.tensor(pred["attention_gt"], dtype=torch.long).to(device=attention_distribution.device)
# 对于[[1]]这种形状,只squeeze一维
if attention_label.shape[0] > 1:
attention_label.squeeze_()
else:
attention_label.squeeze_(1)
# attention_label: 一维tensor,每个值对应每对people-object的关系类别,整体like tensor([2, 1, 1, 2, 0, 0], device='cuda:0')
if not conf.bce_loss:
# multi-label margin loss or adaptive loss
# spatial_label/contact_label: 二维tensor,其中每个一维tensor对应每对people-object的关系类别
# 每个一维tensor like tensor([ 2, 4, -1, -1, -1, -1], device='cuda:0'),-1之前的表示gt的关系
spatial_label = -torch.ones([len(pred["spatial_gt"]), 6], dtype=torch.long).to(device=attention_distribution.device)
contact_label = -torch.ones([len(pred["contacting_gt"]), 17], dtype=torch.long).to(device=attention_distribution.device)
for i in range(len(pred["spatial_gt"])):
spatial_label[i, : len(pred["spatial_gt"][i])] = torch.tensor(pred["spatial_gt"][i])
contact_label[i, : len(pred["contacting_gt"][i])] = torch.tensor(pred["contacting_gt"][i])
else:
# bce loss
# spatial_label/contact_label: 二维tensor,其中每个一维tensor对应每对people-object的关系类别
# 每个一维tensor like tensor([ 0, 0, 1, 0, 1, 0], device='cuda:0'),1表示gt的关系,其他表示没有gt关系
spatial_label = torch.zeros([len(pred["spatial_gt"]), 6], dtype=torch.float32).to(device=attention_distribution.device)
contact_label = torch.zeros([len(pred["contacting_gt"]), 17], dtype=torch.float32).to(device=attention_distribution.device)
for i in range(len(pred["spatial_gt"])):
spatial_label[i, pred["spatial_gt"][i]] = 1
contact_label[i, pred["contacting_gt"][i]] = 1
# 确定soft target和hard target的比例
# 只需要soft target或者只需要hard target时,设定alpha=0或1即可
if conf.temperature is None or conf.temperature == 1:
temperature = 1
else:
temperature = conf.temperature
losses = {}
if conf.alpha is None or conf.alpha == 0:
alpha = 0
else:
alpha = conf.alpha
gt_rel = pred['rel_gt']
if alpha != 0:
# 需要soft label计算蒸馏损失的情况,否则只需要hard label
if conf.teacher_mode_cfg is not None:
student_object_distribution = pred['distribution']
student_attention_distribution = pred["attention_distribution"]
student_spatial_distribution = pred["spatial_distribution"]
student_contact_distribution = pred["contacting_distribution"]
teacher_object_distribution = t_pred['distribution']
teacher_attention_distribution = t_pred["attention_distribution"]
teacher_spatial_distribution = t_pred["spatial_distribution"]
teacher_contact_distribution = t_pred["contacting_distribution"]
if conf.label_fusion_strategy == 0:
fusion_spatial_distribution = teacher_spatial_distribution * alpha + spatial_label * (1-alpha)
fusion_contact_distribution = teacher_contact_distribution * alpha + contact_label * (1-alpha)
elif conf.label_fusion_strategy == 1:
pred_spatial_label, pred_contact_label = trans_module(teacher_spatial_distribution, teacher_contact_distribution, pred['obj_labels'])
pseudo_id, transition_id = track_iou(t_pred['labels'], t_pred['im_idx'], t_pred['boxes'][:, 1:5], 0.5)
spatial_relation_loss_pred = kl_loss(F.log_softmax(teacher_spatial_distribution[pseudo_id], dim=1), F.softmax(pred_spatial_label[transition_id], dim=1))
contact_relation_loss_pred = kl_loss(F.log_softmax(teacher_contact_distribution[pseudo_id], dim=1), F.softmax(pred_contact_label[transition_id], dim=1))
losses['spatial_relation_loss_pred'] = spatial_relation_loss_pred
losses['contact_relation_loss_pred'] = contact_relation_loss_pred
alpha1 = 2 - 2 * F.sigmoid(spatial_relation_loss_pred).detach()
alpha2 = 2 - 2 * F.sigmoid(contact_relation_loss_pred).detach()
fusion_spatial_distribution = teacher_spatial_distribution * alpha1 + spatial_label * (1-alpha1)
fusion_contact_distribution = teacher_contact_distribution * alpha2 + contact_label * (1-alpha2)
alpha1_cnt += alpha1.item()
alpha2_cnt += alpha2.item()
fusion_spatial_distribution[gt_rel] = spatial_label[gt_rel]
fusion_contact_distribution[gt_rel] = contact_label[gt_rel]
spatial_label = fusion_spatial_distribution
contact_label = fusion_contact_distribution
if conf.mode == 'sgcls' or conf.mode == 'sgdet':
losses['object_loss'] = ce_loss(pred['distribution'], object_label)
losses["attention_relation_loss"] = ce_loss(attention_distribution, attention_label)
if not conf.bce_loss:
losses["spatial_relation_loss"] = mlm_loss(spatial_distribution, spatial_label)
losses["contact_relation_loss"] = mlm_loss(contact_distribution, contact_label)
else:
if conf.loss == 'KL':
# print(spatial_distribution, spatial_label)
losses["spatial_relation_loss"] = kl_loss(F.log_softmax(spatial_distribution, dim=1), F.softmax(spatial_label, dim=1))
losses["contact_relation_loss"] = kl_loss(F.log_softmax(contact_distribution, dim=1), F.softmax(contact_label, dim=1))
elif conf.loss == 'L1':
losses["spatial_relation_loss"] = L1_loss(spatial_distribution, spatial_label)
losses["contact_relation_loss"] = L1_loss(contact_distribution, contact_label)
elif conf.loss == 'L2':
losses["spatial_relation_loss"] = L2_loss(spatial_distribution, spatial_label)
losses["contact_relation_loss"] = L2_loss(contact_distribution, contact_label)
elif conf.loss == 'BCE':
losses["spatial_relation_loss_BCE"] = bce_loss(spatial_distribution, spatial_label)
losses["contact_relation_loss_BCE"] = bce_loss(contact_distribution, contact_label)
optimizer.zero_grad()
loss = sum(losses.values())
loss.backward()
torch.nn.utils.clip_grad_norm_(model.parameters(), max_norm=5, norm_type=2)
optimizer.step()
for k in losses.keys():
writer.add_scalar(k, losses[k], epoch * len(dataloader_train) + b)
tr.append(pd.Series({x: y.item() for x, y in losses.items()}))
loss_cnt_tr.append(pd.Series({x: y.item() for x, y in losses.items()}))
if b % save_epoch == 0 and b >= save_epoch:
time_per_batch = (time.time() - start) / save_epoch
print("\ne{:2d} b{:5d}/{:5d} {:.3f}s/batch, {:.1f}m/epoch".format(epoch, b, len(dataloader_train),
time_per_batch, len(dataloader_train) * time_per_batch / 60))
mn = pd.concat(tr[-save_epoch:], axis=1).mean(1)
print(mn)
start = time.time()
t.set_description(desc="Epoch {} ".format(epoch))
# t.set_postfix(steps=step, loss=loss.data.item())
t.update(1)
# 一个epoch结束,存储这个epoch的loss
save_loss[str(epoch)] = {}
loss_all = pd.concat(loss_cnt_tr, axis=1).mean(1)
for loss_name in loss_all.keys():
save_loss[str(epoch)][loss_name] = loss_all[loss_name]
save_loss[str(epoch)]['alpha1'] = alpha1_cnt / b
save_loss[str(epoch)]['alpha2'] = alpha2_cnt / b
with open(os.path.join(conf.save_path, "save_loss_{}.json".format(epoch)), 'w') as f:
json.dump(save_loss, f)
if trans_module is not None:
torch.save({"state_dict": model.state_dict(), "state_dict_3": trans_module.state_dict()}, os.path.join(conf.save_path, "model_{}.tar".format(epoch)))
else:
torch.save({"state_dict": model.state_dict()}, os.path.join(conf.save_path, "model_{}.tar".format(epoch)))
print("*" * 40)
print("save the checkpoint after {} epochs".format(epoch))
model.eval()
object_detector.is_train = False
with torch.no_grad():
with tqdm(total=len(dataloader_test)) as t:
for b in range(len(dataloader_test)):
data = next(test_iter)
im_data = copy.deepcopy(data[0].cuda(0))
im_info = copy.deepcopy(data[1].cuda(0))
gt_boxes = copy.deepcopy(data[2].cuda(0))
num_boxes = copy.deepcopy(data[3].cuda(0))
gt_annotation = AG_dataset_test.gt_annotations[data[4]]
frame_names = AG_dataset_test.video_list[data[4]]
entry = object_detector(im_data, im_info, gt_boxes, num_boxes, gt_annotation, frame_names, faset_rcnn_model, transforms)
if entry != None:
pred = model(entry)
else:
pred = {}
# evaluator.evaluate_scene_graph(gt_annotation, pred)
evaluator1.evaluate_scene_graph(gt_annotation, dict(pred))
evaluator2.evaluate_scene_graph(gt_annotation, dict(pred))
evaluator3.evaluate_scene_graph(gt_annotation, dict(pred))
t.update(1)
print('-----------', flush=True)
score = np.mean(evaluator1.result_dict[conf.mode + "_recall"][20])
evaluator1.print_stats()
# save res
with_res = evaluator1.save_stats()
semi_res = evaluator2.save_stats()
no_res = evaluator3.save_stats()
res = {'with': with_res, 'semi': semi_res, 'no': no_res}
test_res['epoch' + str(epoch)] = res
for k in with_res.keys():
writer.add_scalar('with' + k, with_res[k], epoch)
for k in semi_res.keys():
writer.add_scalar('semi' + k, semi_res[k], epoch)
for k in no_res.keys():
writer.add_scalar('no' + k, no_res[k], epoch)
with open(os.path.join(conf.save_path, "save_res_{}.json".format(epoch)), 'w') as f:
json.dump(test_res, f)
evaluator1.reset_result()
evaluator2.reset_result()
evaluator3.reset_result()
scheduler.step(score)