-
Notifications
You must be signed in to change notification settings - Fork 9
/
train_cr_fiqa.py
181 lines (142 loc) · 6.78 KB
/
train_cr_fiqa.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
import argparse
import logging
import os
from threading import local
import time
import torch
import torch.distributed as dist
import torch.nn.functional as F
from torch.nn.parallel.distributed import DistributedDataParallel
import torch.utils.data.distributed
from torch.nn.utils import clip_grad_norm_
from torch.nn import CrossEntropyLoss
import losses
from config import config as cfg
from dataset import MXFaceDataset, DataLoaderX
from utils.utils_callbacks import CallBackVerification, CallBackLogging, CallBackModelCheckpoint
from utils.utils_logging import AverageMeter, init_logging
from backbones.iresnet import iresnet100, iresnet50
torch.backends.cudnn.benchmark = True
def main(args):
dist.init_process_group(backend='nccl', init_method='env://')
local_rank = args.local_rank
torch.cuda.set_device(local_rank)
rank = dist.get_rank()
world_size = dist.get_world_size()
if not os.path.exists(cfg.output) and rank == 0:
os.makedirs(cfg.output)
else:
time.sleep(2)
log_root = logging.getLogger()
init_logging(log_root, rank, cfg.output)
trainset = MXFaceDataset(root_dir=cfg.rec, local_rank=local_rank)
train_sampler = torch.utils.data.distributed.DistributedSampler(
trainset, shuffle=True)
train_loader = DataLoaderX(
local_rank=local_rank, dataset=trainset, batch_size=cfg.batch_size,
sampler=train_sampler, num_workers=16, pin_memory=True, drop_last=True)
# load evaluation
if cfg.network == "iresnet100":
backbone = iresnet100(dropout=0.4,num_features=cfg.embedding_size).to(local_rank)
elif cfg.network == "iresnet50":
backbone = iresnet50(dropout=0.4,num_features=cfg.embedding_size, use_se=False, qs=1).to(local_rank)
else:
backbone = None
logging.info("load backbone failed!")
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 student loaded successfully!")
if rank == 0:
logging.info("backbone resume loaded successfully!")
except (FileNotFoundError, KeyError, IndexError, RuntimeError):
logging.info("load backbone resume init, failed!")
for ps in backbone.parameters():
dist.broadcast(ps, 0)
backbone = DistributedDataParallel(
module=backbone, broadcast_buffers=False, device_ids=[local_rank])
backbone.train()
# get header
if args.loss == "CR_FIQA_LOSS":
header = losses.CR_FIQA_LOSS(in_features=cfg.embedding_size, out_features=cfg.num_classes, s=cfg.s, m=cfg.m).to(local_rank)
else:
print("Header not implemented")
if args.resume:
try:
header_pth = os.path.join(cfg.identity_model, str(cfg.identity_step) + "header.pth")
header.load_state_dict(torch.load(header_pth, map_location=torch.device(local_rank)))
if rank == 0:
logging.info("header resume loaded successfully!")
except (FileNotFoundError, KeyError, IndexError, RuntimeError):
logging.info("header resume init, failed!")
header = DistributedDataParallel(
module=header, broadcast_buffers=False, device_ids=[local_rank])
header.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)
opt_header = torch.optim.SGD(
params=[{'params': header.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)
scheduler_header = torch.optim.lr_scheduler.LambdaLR(
optimizer=opt_header, lr_lambda=cfg.lr_func)
criterion = CrossEntropyLoss()
criterion_qs= torch.nn.SmoothL1Loss(beta=0.5)
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
scheduler_header.last_epoch = cur_epoch
# --------- this could be solved more elegant ----------------
opt_backbone.param_groups[0]['lr'] = scheduler_backbone.get_lr()[0]
opt_header.param_groups[0]['lr'] = scheduler_header.get_lr()[0]
print("last learning rate: {}".format(scheduler_header.get_lr()))
# ------------------------------------------------------------
callback_verification = CallBackVerification(cfg.eval_step, rank, cfg.val_targets, cfg.rec)
callback_logging = CallBackLogging(50, rank, total_step, cfg.batch_size, world_size, writer=None)
callback_checkpoint = CallBackModelCheckpoint(rank, cfg.output)
alpha=10.0 #10.0
loss = AverageMeter()
global_step = cfg.global_step
for epoch in range(start_epoch, cfg.num_epoch):
train_sampler.set_epoch(epoch)
for _, (img, label) in enumerate(train_loader):
global_step += 1
img = img.cuda(local_rank, non_blocking=True)
label = label.cuda(local_rank, non_blocking=True)
features, qs = backbone(img)
thetas, std, ccs,nnccs = header(features, label)
loss_qs=criterion_qs(ccs/ nnccs,qs)
loss_v = criterion(thetas, label) + alpha* loss_qs
loss_v.backward()
clip_grad_norm_(backbone.parameters(), max_norm=5, norm_type=2)
opt_backbone.step()
opt_header.step()
opt_backbone.zero_grad()
opt_header.zero_grad()
loss.update(loss_v.item(), 1)
callback_logging(global_step, loss, epoch, 0,loss_qs)
callback_verification(global_step, backbone)
scheduler_backbone.step()
scheduler_header.step()
callback_checkpoint(global_step, backbone, header)
dist.destroy_process_group()
if __name__ == "__main__":
parser = argparse.ArgumentParser(description='PyTorch CR-FIQA Training')
parser.add_argument('--local_rank', type=int, default=0, help='local_rank')
parser.add_argument('--loss', type=str, default="CR_FIQA_LOSS", help="loss function")
parser.add_argument('--resume', type=int, default=0, help="resume training")
args_ = parser.parse_args()
main(args_)