forked from deepinsight/insightface
-
Notifications
You must be signed in to change notification settings - Fork 0
/
train.py
137 lines (118 loc) · 5.65 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
import argparse
import logging
import os
import time
import torch
import torch.distributed as dist
import torch.nn.functional as F
import torch.utils.data.distributed
from torch.nn.utils import clip_grad_norm_
import losses
from backbones import get_model
from config import config as cfg
from dataset import MXFaceDataset, DataLoaderX
from partial_fc import PartialFC
from utils.utils_amp import MaxClipGradScaler
from utils.utils_callbacks import CallBackVerification, CallBackLogging, CallBackModelCheckpoint
from utils.utils_logging import AverageMeter, init_logging
def main(args):
try:
world_size = int(os.environ['WORLD_SIZE'])
rank = int(os.environ['RANK'])
dist_url = "tcp://{}:{}".format(os.environ["MASTER_ADDR"], os.environ["MASTER_PORT"])
except KeyError:
world_size = 1
rank = 0
dist_url = "tcp://127.0.0.1:12584"
dist.init_process_group(backend='nccl', init_method=dist_url, rank=rank, world_size=world_size)
local_rank = args.local_rank
torch.cuda.set_device(local_rank)
if not os.path.exists(cfg.output) and rank is 0:
os.makedirs(cfg.output)
else:
time.sleep(2)
log_root = logging.getLogger()
init_logging(log_root, rank, cfg.output)
train_set = MXFaceDataset(root_dir=cfg.rec, local_rank=local_rank)
train_sampler = torch.utils.data.distributed.DistributedSampler(
train_set, shuffle=True)
train_loader = DataLoaderX(
local_rank=local_rank, dataset=train_set, batch_size=cfg.batch_size,
sampler=train_sampler, num_workers=2, pin_memory=True, drop_last=True)
dropout = 0.4 if cfg.dataset == "webface" else 0
backbone = get_model(args.network, dropout=dropout, fp16=cfg.fp16, num_features=cfg.embedding_size).to(local_rank)
if args.resume:
try:
backbone_pth = os.path.join(cfg.output, "backbone.pth")
backbone.load_state_dict(torch.load(backbone_pth, map_location=torch.device(local_rank)))
if rank is 0:
logging.info("backbone resume successfully!")
except (FileNotFoundError, KeyError, IndexError, RuntimeError):
logging.info("resume fail, backbone init successfully!")
for ps in backbone.parameters():
dist.broadcast(ps, 0)
backbone = torch.nn.parallel.DistributedDataParallel(
module=backbone, broadcast_buffers=False, device_ids=[local_rank])
backbone.train()
margin_softmax = losses.get_loss(args.loss)
module_partial_fc = PartialFC(
rank=rank, local_rank=local_rank, world_size=world_size, resume=args.resume,
batch_size=cfg.batch_size, margin_softmax=margin_softmax, num_classes=cfg.num_classes,
sample_rate=cfg.sample_rate, embedding_size=cfg.embedding_size, prefix=cfg.output)
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_pfc = torch.optim.SGD(
params=[{'params': module_partial_fc.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_pfc = torch.optim.lr_scheduler.LambdaLR(
optimizer=opt_pfc, lr_lambda=cfg.lr_func)
start_epoch = 0
total_step = int(len(train_set) / cfg.batch_size / world_size * cfg.num_epoch)
if rank is 0: logging.info("Total Step is: %d" % total_step)
callback_verification = CallBackVerification(2000, rank, cfg.val_targets, cfg.rec)
callback_logging = CallBackLogging(50, rank, total_step, cfg.batch_size, world_size, None)
callback_checkpoint = CallBackModelCheckpoint(rank, cfg.output)
loss = AverageMeter()
global_step = 0
grad_amp = MaxClipGradScaler(cfg.batch_size, 128 * cfg.batch_size, growth_interval=100) if cfg.fp16 else None
for epoch in range(start_epoch, cfg.num_epoch):
train_sampler.set_epoch(epoch)
for step, (img, label) in enumerate(train_loader):
global_step += 1
features = F.normalize(backbone(img))
x_grad, loss_v = module_partial_fc.forward_backward(label, features, opt_pfc)
if cfg.fp16:
features.backward(grad_amp.scale(x_grad))
grad_amp.unscale_(opt_backbone)
clip_grad_norm_(backbone.parameters(), max_norm=5, norm_type=2)
grad_amp.step(opt_backbone)
grad_amp.update()
else:
features.backward(x_grad)
clip_grad_norm_(backbone.parameters(), max_norm=5, norm_type=2)
opt_backbone.step()
opt_pfc.step()
module_partial_fc.update()
opt_backbone.zero_grad()
opt_pfc.zero_grad()
loss.update(loss_v, 1)
callback_logging(global_step, loss, epoch, cfg.fp16, grad_amp)
callback_verification(global_step, backbone)
callback_checkpoint(global_step, backbone, module_partial_fc)
scheduler_backbone.step()
scheduler_pfc.step()
dist.destroy_process_group()
if __name__ == "__main__":
torch.backends.cudnn.benchmark = True
parser = argparse.ArgumentParser(description='PyTorch ArcFace Training')
parser.add_argument('--local_rank', type=int, default=0, help='local_rank')
parser.add_argument('--network', type=str, default='r50', help='backbone network')
parser.add_argument('--loss', type=str, default='arcface', help='loss function')
parser.add_argument('--resume', type=int, default=0, help='model resuming')
args_ = parser.parse_args()
main(args_)