forked from yrcong/STTran
-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathtrain.py
179 lines (148 loc) · 8.13 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
import torch
import torch.nn as nn
from torch import optim
from torch.optim.lr_scheduler import ReduceLROnPlateau
import numpy as np
np.set_printoptions(precision=3)
import time
import os
import pandas as pd
import copy
from dataloader.action_genome import AG, cuda_collate_fn
from lib.object_detector import detector
from lib.config import Config
from lib.evaluation_recall import BasicSceneGraphEvaluator
from lib.AdamW import AdamW
from lib.sttran import STTran
"""------------------------------------some settings----------------------------------------"""
conf = Config()
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))
for i in conf.args:
print(i,':', conf.args[i])
"""-----------------------------------------------------------------------------------------"""
AG_dataset_train = AG(mode="train", datasize=conf.datasize, data_path=conf.data_path, 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=4,
collate_fn=cuda_collate_fn, pin_memory=False)
AG_dataset_test = AG(mode="test", datasize=conf.datasize, data_path=conf.data_path, 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=4,
collate_fn=cuda_collate_fn, pin_memory=False)
gpu_device = torch.device("cuda:0")
# freeze the detection backbone
object_detector = detector(train=True, object_classes=AG_dataset_train.object_classes, use_SUPPLY=True, mode=conf.mode).to(device=gpu_device)
object_detector.eval()
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).to(device=gpu_device)
evaluator =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')
# loss function, default Multi-label margin loss
if conf.bce_loss:
ce_loss = nn.CrossEntropyLoss()
bce_loss = nn.BCELoss()
else:
ce_loss = nn.CrossEntropyLoss()
mlm_loss = nn.MultiLabelMarginLoss()
# optimizer
if conf.optimizer == 'adamw':
optimizer = AdamW(model.parameters(), 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)
# some parameters
tr = []
for epoch in range(conf.nepoch):
model.train()
object_detector.is_train = True
start = time.time()
train_iter = iter(dataloader_train)
test_iter = iter(dataloader_test)
for b in range(len(dataloader_train)):
data = next(train_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_train.gt_annotations[data[4]]
# prevent gradients to FasterRCNN
with torch.no_grad():
entry = object_detector(im_data, im_info, gt_boxes, num_boxes, gt_annotation ,im_all=None)
pred = model(entry)
attention_distribution = pred["attention_distribution"]
spatial_distribution = pred["spatial_distribution"]
contact_distribution = pred["contacting_distribution"]
attention_label = torch.tensor(pred["attention_gt"], dtype=torch.long).to(device=attention_distribution.device).squeeze()
if not conf.bce_loss:
# multi-label margin loss or adaptive loss
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 = 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
losses = {}
if conf.mode == 'sgcls' or conf.mode == 'sgdet':
losses['object_loss'] = ce_loss(pred['distribution'], pred['labels'])
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:
losses["spatial_relation_loss"] = bce_loss(spatial_distribution, spatial_label)
losses["contact_relation_loss"] = 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()
tr.append(pd.Series({x: y.item() for x, y in losses.items()}))
if b % 1000 == 0 and b >= 1000:
time_per_batch = (time.time() - start) / 1000
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[-1000:], axis=1).mean(1)
print(mn)
start = time.time()
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():
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]]
entry = object_detector(im_data, im_info, gt_boxes, num_boxes, gt_annotation, im_all=None)
pred = model(entry)
evaluator.evaluate_scene_graph(gt_annotation, pred)
print('-----------', flush=True)
score = np.mean(evaluator.result_dict[conf.mode + "_recall"][20])
evaluator.print_stats()
evaluator.reset_result()
scheduler.step(score)