-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathKD_alone.py
184 lines (148 loc) · 9.36 KB
/
KD_alone.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
import argparse
import logging
import os
import time
import numpy as np
import torch
import torch.nn.functional as F
import torch.utils.data.distributed
from torch.nn.utils import clip_grad_norm_
from torch.nn import MSELoss
from utils.augmentation import get_conventional_aug_policy
from torch.utils.data import DataLoader
from utils import losses
from config.config import config as cfg
from utils.dataset import MXFaceDataset,MXFaceDataset, MXSyntheticFaceDataset, MXFaceDataset_rec
from utils.utils_callbacks import CallBackVerification, CallBackLogging, CallBackModelCheckpoint
from utils.utils_logging import AverageMeter, init_logging
from backbones.iresnet import iresnet100, iresnet50, iresnet34
torch.backends.cudnn.benchmark = True
def main(args):
local_rank = args.local_rank
torch.cuda.set_device(0)
rank = 0
world_size = 1
# directories definition
if cfg.is_teacher_baseline:
if cfg.is_dual_layer:
if cfg.has_drop_out:
cfg.output_KD="/nas-ctm01/homes/mecaldeira/output/KD/no_CEL/EArc/baseline/dropout_dual_layer_"+str(cfg.middle_layer_size)+"/"+cfg.network+"/lr_"+str(cfg.lr_fc)+"_ep_"+str(cfg.num_epoch_fc)+"/"+cfg.arc_method
cfg.path_teacher_fts="/nas-ctm01/datasets/public/BIOMETRICS/BalancedFace_embeddings/baseline_embeddings/fully_connected/EArcFace/dropout_dual_layer_"+str(cfg.middle_layer_size)+"/"+cfg.network+"/lr_"+str(cfg.lr_fc)+"_ep_"+str(cfg.num_epoch_fc)+"/"+cfg.arc_method
else:
cfg.output_KD="/nas-ctm01/homes/mecaldeira/output/KD/no_CEL/EArc/baseline/dual_layer_"+str(cfg.middle_layer_size)+"/"+cfg.network+"/lr_"+str(cfg.lr_fc)+"_ep_"+str(cfg.num_epoch_fc)+"/"+cfg.arc_method
cfg.path_teacher_fts="/nas-ctm01/datasets/public/BIOMETRICS/BalancedFace_embeddings/baseline_embeddings/fully_connected/EArcFace/dual_layer_"+str(cfg.middle_layer_size)+"/"+cfg.network+"/lr_"+str(cfg.lr_fc)+"_ep_"+str(cfg.num_epoch_fc)+"/"+cfg.arc_method
else:
cfg.output_KD="/nas-ctm01/homes/mecaldeira/output/KD/no_CEL/EArc/baseline/"+cfg.network+"/lr_"+str(cfg.lr_fc)+"_ep_"+str(cfg.num_epoch_fc)+"/"+cfg.arc_method
cfg.path_teacher_fts="/nas-ctm01/datasets/public/BIOMETRICS/BalancedFace_embeddings/baseline_embeddings/fully_connected/EArcFace/"+cfg.network+"/lr_"+str(cfg.lr_fc)+"_ep_"+str(cfg.num_epoch_fc)+"/"+cfg.arc_method
else:
if cfg.is_dual_layer:
if cfg.has_drop_out:
cfg.output_KD="/nas-ctm01/homes/mecaldeira/output/KD/no_CEL/EArc/dropout_dual_layer_"+str(cfg.middle_layer_size)+"/"+cfg.network+"/lr_"+str(cfg.lr_fc)+"_ep_"+str(cfg.num_epoch_fc)+"/"+cfg.arc_method
cfg.path_teacher_fts="/nas-ctm01/datasets/public/BIOMETRICS/BalancedFace_embeddings/fully_connected/EArcFace/dropout_dual_layer_"+str(cfg.middle_layer_size)+"/"+cfg.network+"/lr_"+str(cfg.lr_fc)+"_ep_"+str(cfg.num_epoch_fc)+"/"+cfg.arc_method
else:
cfg.output_KD="/nas-ctm01/homes/mecaldeira/output/KD/no_CEL/EArc/dual_layer_"+str(cfg.middle_layer_size)+"/"+cfg.network+"/lr_"+str(cfg.lr_fc)+"_ep_"+str(cfg.num_epoch_fc)+"/"+cfg.arc_method
cfg.path_teacher_fts="/nas-ctm01/datasets/public/BIOMETRICS/BalancedFace_embeddings/fully_connected/EArcFace/dual_layer_"+str(cfg.middle_layer_size)+"/"+cfg.network+"/lr_"+str(cfg.lr_fc)+"_ep_"+str(cfg.num_epoch_fc)+"/"+cfg.arc_method
else:
cfg.output_KD="/nas-ctm01/homes/mecaldeira/output/KD/no_CEL/EArc/"+cfg.network+"/lr_"+str(cfg.lr_fc)+"_ep_"+str(cfg.num_epoch_fc)+"/"+cfg.arc_method
cfg.path_teacher_fts="/nas-ctm01/datasets/public/BIOMETRICS/BalancedFace_embeddings/fully_connected/EArcFace/"+cfg.network+"/lr_"+str(cfg.lr_fc)+"_ep_"+str(cfg.num_epoch_fc)+"/"+cfg.arc_method
# check if the saving path exists and create it if not
if not os.path.exists(cfg.output_KD) and rank == 0:
os.makedirs(cfg.output_KD)
else:
time.sleep(2)
transform = get_conventional_aug_policy(cfg.augmentation)
log_root = logging.getLogger()
init_logging(log_root, rank, cfg.output_KD, logfile="train_KD.log")
if cfg.dataset == 'competition':
if transform is not None:
trainset = MXSyntheticFaceDataset(root_dir=cfg.rec, local_rank=local_rank,from_file=cfg.from_file,transform=transform)
else:
trainset = MXSyntheticFaceDataset(root_dir=cfg.rec, local_rank=local_rank,from_file=cfg.from_file)
elif cfg.dataset == 'competition_baseline':
trainset = MXFaceDataset_rec(root_dir=cfg.rec, local_rank=local_rank, transform=transform)
else:
trainset = MXFaceDataset(root_dir=cfg.rec, ethnicity=cfg.ethnicity, local_rank=local_rank, is_train=True, to_sample=cfg.to_sample)
train_sampler = torch.utils.data.RandomSampler(trainset)
train_loader = DataLoader(dataset=trainset, batch_size=cfg.batch_size,
sampler = train_sampler ,num_workers=4, pin_memory=True, drop_last=True,prefetch_factor=2)
# load model
if cfg.student_network == "iresnet100":
backbone = iresnet100(num_features=cfg.embedding_size, use_se=cfg.SE).to(local_rank)
# FIXME Verify dropout rate
elif cfg.student_network == "iresnet50":
backbone = iresnet50(dropout=0.4,num_features=cfg.embedding_size, use_se=cfg.SE).to(local_rank)
elif cfg.student_network == "iresnet34":
backbone = iresnet34(num_features=cfg.embedding_size, use_se=cfg.SE).to(local_rank)
else:
backbone = None
logging.info("load backbone failed!")
exit()
if args.resume:
try:
backbone_pth = os.path.join(cfg.output, str(cfg.global_step) + "backbone.pth")
backbone.load_state_dict(torch.load(backbone_pth, map_location=torch.device(local_rank)))
if rank == 0:
logging.info("backbone resume loaded successfully!")
except (FileNotFoundError, KeyError, IndexError, RuntimeError):
logging.info("load backbone resume init, failed!")
backbone.train()
opt_backbone = torch.optim.SGD(
params=[{'params': backbone.parameters()}],
lr=cfg.lr / 512 * cfg.batch_size * world_size,
momentum=0.9, weight_decay=cfg.weight_decay)
scheduler_backbone = torch.optim.lr_scheduler.LambdaLR(
optimizer=opt_backbone, lr_lambda=cfg.lr_func)
header=None
start_epoch = 0
total_step = int(len(trainset) / cfg.batch_size / world_size * cfg.num_epoch)
if rank == 0: logging.info("Total Step is: %d" % total_step)
if args.resume:
rem_steps = (total_step - cfg.global_step)
cur_epoch = cfg.num_epoch - int(cfg.num_epoch / total_step * rem_steps)
logging.info("resume from estimated epoch {}".format(cur_epoch))
logging.info("remaining steps {}".format(rem_steps))
start_epoch = cur_epoch
scheduler_backbone.last_epoch = cur_epoch
# --------- this could be solved more elegant ----------------
opt_backbone.param_groups[0]['lr'] = scheduler_backbone.get_lr()[0]
# ------------------------------------------------------------
callback_verification = CallBackVerification(cfg.eval_step, rank, cfg.val_targets, cfg.rec)
callback_logging = CallBackLogging(500, rank, total_step, cfg.batch_size, world_size, writer=None)
callback_checkpoint = CallBackModelCheckpoint(rank, cfg.output_KD)
loss = AverageMeter()
global_step = cfg.global_step
criterion_KD = MSELoss()
logging.info("Train starting for %s samples.", cfg.ethnicity)
for epoch in range(start_epoch, cfg.num_epoch):
running_loss = 0
steps = 0
for _, (img, label, extra_img_path) in enumerate(train_loader):
global_step += 1
img = img.cuda(local_rank, non_blocking=True)
label = label.cuda(local_rank, non_blocking=True)
features = F.normalize(backbone(img))
teacher_fts=np.zeros((len(label), cfg.embedding_size))
for sample in range(len(label)):
teacher_fts[sample,:]=np.load(cfg.path_teacher_fts+extra_img_path[sample]+".npy").reshape(-1, cfg.embedding_size)
teacher_fts=torch.from_numpy(teacher_fts).float()
teacher_fts=teacher_fts.cuda(local_rank, non_blocking=True)
loss_v = cfg.loss_lambda*criterion_KD(teacher_fts, features)
loss_v.backward()
clip_grad_norm_(backbone.parameters(), max_norm=5, norm_type=2)
opt_backbone.step()
opt_backbone.zero_grad()
loss.update(loss_v.item(), 1)
running_loss += loss_v.item()
steps+=1
callback_logging(global_step, loss, epoch)
logging.info("Loss check: total -> %f.", running_loss)
callback_verification(global_step, backbone)
scheduler_backbone.step()
callback_checkpoint(global_step, backbone, header)
logging.info("KD train complete for %d identities.", cfg.to_sample)
if __name__ == "__main__":
parser = argparse.ArgumentParser(description='PyTorch margin penalty loss training')
parser.add_argument('--local_rank', type=int, default=0, help='local_rank')
parser.add_argument('--resume', type=int, default=0, help="resume training")
args_ = parser.parse_args()
main(args_)