-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathtrain_kd.py
174 lines (132 loc) · 6.33 KB
/
train_kd.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
import os
import sys
import time
import json
import torch
from torch import nn
from tqdm import tqdm
import numpy as np
import random
from arguments.argument_kd import get_args
from models.model_kd import PoseModuleKD as PoseModule
from libs.distributed import (
get_rank,
synchronize,
)
from tensorboardX import SummaryWriter
from libs.eval_libs import valid
from libs.train_libs import close_shared_memory, build_model, build_model_teacher
from libs.train_libs import build_dataset
# close shared memory of pytorch
if True:
close_shared_memory()
if __name__ == '__main__':
torch.cuda.empty_cache()
# reproducibility: https://pytorch.org/docs/stable/notes/randomness.html
torch.manual_seed(0)
np.random.seed(0)
random.seed(0)
cfg, cfg_t = get_args()
n_gpu = int(os.environ['WORLD_SIZE']) if 'WORLD_SIZE' in os.environ else 1
cfg['RUNTIME']['N_GPU'] = n_gpu
cfg['RUNTIME']['DISTRIBUTED'] = n_gpu > 1
cfg_t['RUNTIME']['DISTRIBUTED'] = n_gpu > 1
if cfg['RUNTIME']['DISTRIBUTED']:
torch.cuda.set_device(cfg['RUNTIME']['LOCAL_RANK'])
torch.distributed.init_process_group(backend='gloo', init_method='env://')
synchronize()
# device = 'cuda'
device = cfg['RUNTIME']['RUNNING_DEVICE']
# build dataset
train_loader, valid_loader = build_dataset(cfg)
cfg['KD']['vis_dir'] = cfg['RUNTIME']['WORKING_DIR']
print('Building teacher ......')
model_t= build_model_teacher(cfg_t, PoseModule, device)
print('Building student ......')
model, optimizer, scheduler, total_steps = build_model(cfg, PoseModule, device)
VAL_FREQ = cfg['SOLVER']['VAL_FREQ']
#
print("working directory: " + cfg['RUNTIME']['WORKING_DIR'])
if get_rank() == 0:
os.makedirs(cfg['RUNTIME']['WORKING_DIR'], exist_ok=True)
logger = SummaryWriter(cfg['RUNTIME']['WORKING_DIR'])
# compute model size
total_params_count = sum(p.numel() for p in model.parameters())
total_params_count_t = sum(p.numel() for p in model_t.parameters())
print(f"Model size: Student VS Teacher: {total_params_count:d} vs {total_params_count_t:d}")
# write cfg to working_dir
with open(cfg['RUNTIME']['WORKING_DIR'] + 'cfg.json', 'w') as f:
json.dump(cfg, f, indent=4, sort_keys=True)
print("--- evaluate teacher ---")
valid(cfg, total_steps, valid_loader, model_t, device, logger=None)
model.train()
model_t.eval()
cfg_kd = cfg['KD']
should_keep_training = True
while should_keep_training:
pbar = tqdm(enumerate(train_loader), total=len(train_loader), dynamic_ncols=True)
for idx, (images, targets, _) in pbar:
if total_steps >= cfg['SOLVER']['MAX_ITER']:
should_keep_training = False
valid(cfg, total_steps, valid_loader, model, device, logger=logger)
torch.save(model.state_dict(), cfg['RUNTIME']['WORKING_DIR'] + 'final.pth')
print('Training finished')
break
total_steps += 1
model.zero_grad()
images = images.to(device)
targets = [target.to(device) for target in targets]
with torch.no_grad():
pred_t = model_t(images, targets=targets, is_teacher=True, cfg_kd=cfg_kd)
_, loss_dict = model(images, targets=targets, pred_t=pred_t, cfg_kd=cfg_kd)
# add pure loss value to tensorboard
if get_rank() == 0:
current_lr = optimizer.param_groups[0]['lr']
# writing log to tensorboard
if logger and idx % 10 == 0:
logger.add_scalar('training/learning_rate', current_lr, total_steps)
logger.add_scalar('training/loss_cls', loss_dict['loss_cls'], total_steps)
logger.add_scalar('training/loss_reg', loss_dict['loss_reg'], total_steps)
logger.add_scalar('training/loss_cls_reg', (loss_dict['loss_cls'] + loss_dict['loss_reg']), total_steps)
logger.add_scalar('training/loss_kd', loss_dict['loss_kd'], total_steps)
loss_dict['loss_cls'] = loss_dict['loss_cls'] * cfg['SOLVER']['LOSS_WEIGHT_CLS']
loss_dict['loss_reg'] = loss_dict['loss_reg'] * cfg['SOLVER']['LOSS_WEIGHT_REG']
loss_cls = loss_dict['loss_cls'].mean()
loss_reg = loss_dict['loss_reg'].mean()
loss = loss_cls + loss_reg
if "loss_kd" in loss_dict.keys():
loss_dict['loss_kd']= loss_dict['loss_kd'] * cfg['KD']['LOSS_WEIGHT_KD']
loss_kd = loss_dict['loss_kd'].mean()
if cfg['KD']['LOSS_WEIGHT_KD'] > 0.0:
loss = loss + loss_kd
loss.backward()
nn.utils.clip_grad_norm_(model.parameters(), cfg['SOLVER']['GRAD_CLIP'])
optimizer.step()
scheduler.step()
if get_rank() == 0:
if "loss_kd" in loss_dict.keys():
pbar_str = (("steps: %d/%d, lr:%.6f, cls:%.4f, reg:%.4f, kd:%.4f") % (total_steps, cfg['SOLVER']['MAX_ITER'], current_lr, loss_cls, loss_reg, loss_kd))
else:
pbar_str = (("steps: %d/%d, lr:%.6f, cls:%.4f, reg:%.4f") % (total_steps, cfg['SOLVER']['MAX_ITER'], current_lr, loss_cls, loss_reg))
pbar.set_description(pbar_str)
if total_steps % VAL_FREQ == 0:
valid(cfg, total_steps, valid_loader, model, device, logger=logger)
model.train()
torch.save({
'steps': total_steps,
'model': model.state_dict(),
'optim': optimizer.state_dict(),
'sched': scheduler.state_dict(),
},
cfg['RUNTIME']['WORKING_DIR'] + 'latest.pth',
)
# output final info
if get_rank() == 0:
timestr = time.strftime('%Y%m%d_%H%M%S',time.localtime(time.time()))
commandstr = ' '.join([str(elem) for elem in sys.argv])
final_msg = ("finished at: %s\nworking_dir: %s\ncommands:%s" % (timestr, cfg['RUNTIME']['WORKING_DIR'], commandstr))
with open(cfg['RUNTIME']['WORKING_DIR'] + 'info.txt', 'w') as f:
f.write(final_msg)
print(final_msg)
torch.cuda.empty_cache()