-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathtrain_amp.py
194 lines (168 loc) · 8.54 KB
/
train_amp.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
import torch
import torch.nn as nn
from torch import optim
from torch.optim.lr_scheduler import ReduceLROnPlateau
from torch.cuda.amp import GradScaler, autocast
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.models.tr2_model import TR2
from lib.models.utils import *
def train():
# dataset
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=0,collate_fn=cuda_collate_fn, pin_memory=True)
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=0,collate_fn=cuda_collate_fn, pin_memory=True)
# detector
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
model = TR2(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, rel_classes=AG_dataset_train.relationship_classes,
enc_layer_num=conf.enc_layer, dec_layer_num=conf.dec_layer,
pre_path=conf.pre_path).to(device=gpu_device)
gradScaler = GradScaler()
# parameter
for pname, pvalue in model.named_parameters():
if pname[:4]=='clip' or pname.split('.')[1][:4]=='clip':
pvalue.requires_grad=False
params=filter(lambda p:p.requires_grad, model.parameters())
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
focaln_spa=[4643,58176,254476,46368, 69810, 12921]
focaln_con=[4008, 4076, 4377, 3214, 314, 156897, 11506, 3395, 105067,
8743, 40545, 7606, 52165, 86, 6761, 772, 1102]
ce_loss = nn.CrossEntropyLoss()
mse_loss = nn.MSELoss()
focal_loss_spa=FocalLoss(focaln_spa,conf.mode)
focal_loss_con=FocalLoss(focaln_con,conf.mode)
# optimizer
if conf.optimizer == 'adamw':
optimizer = AdamW(filter(lambda p:p.requires_grad, model.parameters()),\
lr=conf.lr, weight_decay=0.01)
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-6)
tr = []
for epoch in range(conf.nepoch):
print("*" * 50, 'train begin')
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)) # n_frame*3*h*w
im_info = copy.deepcopy(data[1].cuda(0)) # n_frame*3
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]]
with torch.no_grad(): # prevent gradients to FasterRCNN
entry = object_detector(im_data, im_info, gt_boxes, num_boxes,gt_annotation)
entry['origin_ims'] = data[6]
entry['im_info'] = im_info
entry['video_name'] = gt_annotation[0][1]['metadata']['tag'][:5]
with autocast():
pred, diff_v, diff_t = model(entry)
att_distribution = pred["attention_distribution"]
spa_distribution = pred["spatial_distribution"]
con_distribution = pred["contacting_distribution"]
# prepare labels
att_label = torch.tensor(pred["attention_gt"], dtype=torch.long,device=gpu_device).squeeze()
spa_label = torch.zeros([len(pred["spatial_gt"]), 6], dtype=torch.float32,device=gpu_device)
con_label = torch.zeros([len(pred["contacting_gt"]),17], dtype=torch.float32,device=gpu_device)
for i in range(len(pred["spatial_gt"])):
spa_label[i, pred["spatial_gt"][i]] = 1
con_label[i, pred["contacting_gt"][i]] = 1
# calculate loss
losses = {}
losses['align_a'] = mse_loss(diff_v[0], diff_t[0]) # kd loss
losses['align_s'] = mse_loss(diff_v[1], diff_t[1])
losses['align_c'] = mse_loss(diff_v[2], diff_t[2])
losses['object_loss'] = ce_loss(pred['distribution'], pred['labels'])
losses["att"] = ce_loss(att_distribution, att_label)
losses["spa_focal"] = focal_loss_spa(spa_distribution, spa_label,pred["spatial_gt"])
losses["con_focal"] = focal_loss_con(con_distribution, con_label,pred["contacting_gt"])
loss = sum(losses.values())
# optimize
optimizer.zero_grad()
gradScaler.scale(loss).backward()
# loss.backward()
gradScaler.unscale_(optimizer)
torch.nn.utils.clip_grad_norm_(model.parameters(), max_norm=5, norm_type=2)
gradScaler.step(optimizer)
gradScaler.update()
# optimizer.step()
# print
tr.append(pd.Series({x: y.item() for x, y in losses.items()}))
if b % 50 == 0 and b >= 50:
time_per_batch = (time.time() - start) / 50
print("e{: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[-50:], axis=1).mean(1)
for loss_k, loss_v in mn.items():
print('%s:%.4f'%(loss_k,loss_v),end=' ')
print("")
start = time.time()
print("*" * 50, 'test begin')
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)
entry['origin_ims'] = data[6]
entry['im_info'] = im_info
entry['video_name'] = gt_annotation[0][1]['metadata']['tag'][:5]
pred, _, _ = model(entry, ifTest=True)
evaluator.evaluate_scene_graph(gt_annotation, pred)
score = np.mean(evaluator.result_dict[conf.mode + "_recall"][20])
evaluator.print_stats()
evaluator.reset_result()
torch.save({"state_dict": model.state_dict()},conf.output_path+"model_%d_%.3f.tar"%(epoch,score))
print("save the checkpoint model_%d_%.3f.tar"%(epoch,score))
scheduler.step(score)
if __name__ == "__main__":
conf = Config()
if not os.path.exists(conf.output_path):
os.mkdir(conf.output_path)
for i in conf.args:
print(i,':', conf.args[i])
conf.mode='sgdet'
train()