-
Notifications
You must be signed in to change notification settings - Fork 0
/
train_multi_gpu.py
149 lines (124 loc) · 6.6 KB
/
train_multi_gpu.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
import os
import torch
import tempfile
from config import Config
import torch.optim as optim
from models.SGANet import SGANet
from torch.utils.data import DataLoader
from utils.tools import safe_load_weights
from torch.utils.tensorboard import SummaryWriter
from utils.train_eval_utils import train_one_epoch, evaluate
from dataset.dataset import CorrespondencesDataset, collate_fn
from utils.distributed_utils import init_distributed_mode, dist, cleanup
if __name__ == '__main__':
if torch.cuda.is_available() is False:
raise EnvironmentError("not find GPU device for training.")
conf = Config()
init_distributed_mode(args=conf)
rank = conf.rank
device = torch.device(conf.device)
world_size = 4 # Set up world size what you use
batch_size = 32 # Set up batch size that suits your device
true_batch_size = world_size * batch_size
_scaling = true_batch_size / conf.canonical_bs
true_lr = conf.canonical_lr * _scaling
conf.loss_essential_init_iter = int(conf.canonical_bs * 20000 // true_batch_size)
checkpoint_path = ""
if rank == 0:
print('Start Tensorboard with "tensorboard --logdir=runs", view at http://localhost:6006/')
tb_writer = SummaryWriter(log_dir=conf.writer_dir)
# Create directory
if not os.path.isdir(conf.checkpoint_path):
os.makedirs(conf.checkpoint_path)
if not os.path.isdir(conf.best_model_path):
os.makedirs(conf.best_model_path)
# Load data
train_dataset = CorrespondencesDataset(conf.data_tr, conf)
valid_dataset = CorrespondencesDataset(conf.data_va, conf)
train_sampler = torch.utils.data.distributed.DistributedSampler(train_dataset)
train_batch_sampler = torch.utils.data.BatchSampler(train_sampler, batch_size, drop_last=True)
if rank == 0:
print('Using {} dataloader workers every process'.format(conf.num_workers))
train_loader = DataLoader(train_dataset,
batch_sampler=train_batch_sampler,
pin_memory=True,
num_workers=conf.num_workers,
collate_fn=collate_fn)
valid_loader = DataLoader(valid_dataset,
batch_size=1,
shuffle=False,
pin_memory=True,
num_workers=8,
collate_fn=collate_fn)
# Create model
model = SGANet(conf).to(device)
# Set optimizer
pg = [p for p in model.parameters() if p.requires_grad]
optimizer = optim.Adam(pg, lr=true_lr, weight_decay=conf.weight_decay)
best_auc = -1
start_epoch = -1
if os.path.exists(conf.resume):
weights_dict = torch.load(conf.resume, map_location=device)
best_auc = weights_dict['best_auc']
start_epoch = weights_dict['epoch']
safe_load_weights(model, weights_dict['state_dict'])
else:
checkpoint_path = os.path.join(tempfile.gettempdir(), "initial_weights.pt")
if rank == 0:
torch.save(model.state_dict(), checkpoint_path)
dist.barrier()
model.load_state_dict(torch.load(checkpoint_path, map_location=device))
model = torch.nn.parallel.DistributedDataParallel(model, device_ids=[conf.gpu], find_unused_parameters=True)
for epoch in range(start_epoch + 1, conf.epochs):
train_sampler.set_epoch(epoch)
cur_global_step = (epoch - 1) * train_dataset.__len__()
if rank == 0:
mean_loss = train_one_epoch(model=model,
optimizer=optimizer,
data_loader=train_loader,
device="cuda",
epoch=epoch,
conf=conf,
cur_global_step=cur_global_step,
tb_writer=tb_writer)
else:
mean_loss = train_one_epoch(model=model,
optimizer=optimizer,
data_loader=train_loader,
device="cuda",
epoch=epoch,
conf=conf,
cur_global_step=cur_global_step)
if rank == 0:
aucs5, aucs10, aucs20, va_res, precisions, recalls, f_scores = evaluate(model, valid_loader, conf, epoch=epoch)
print("[AUC result epoch {}] AUC@5: {} AUC@10: {} AUC@20: {}".format(epoch, round(aucs5, 3), round(aucs10, 3), round(aucs20, 3)))
print("[Pose metric epoch {}] mAP5: {} mAP10: {} mAP20: {}".format(epoch, round(va_res[0] * 100, 3), round(va_res[1] * 100, 3), round(va_res[3] * 100, 3)))
print("[Outlier metric epoch {}] Precisions: {} Recalls: {} F_scores: {}\n".format(epoch, round(precisions * 100, 3), round(recalls * 100, 3), round(f_scores * 100, 3)))
tags = ["train_loss", "AUC@5", "AUC@10", "AUC@20", "mAP5", "mAP10", "mAP20", "Precisions", "Recalls", "F_scores", "learning_rate"]
tb_writer.add_scalar(tags[1], aucs5, epoch)
tb_writer.add_scalar(tags[2], aucs10, epoch)
tb_writer.add_scalar(tags[3], aucs20, epoch)
tb_writer.add_scalar(tags[4], va_res[0] * 100, epoch)
tb_writer.add_scalar(tags[5], va_res[1] * 100, epoch)
tb_writer.add_scalar(tags[6], va_res[3] * 100, epoch)
tb_writer.add_scalar(tags[7], precisions * 100, epoch)
tb_writer.add_scalar(tags[8], recalls * 100, epoch)
tb_writer.add_scalar(tags[9], f_scores * 100, epoch)
tb_writer.add_scalar(tags[10], optimizer.param_groups[0]["lr"], epoch)
if aucs5 > best_auc:
print("Saving best model with auc5 = {}\n".format(aucs5))
best_auc = aucs5
torch.save({
'epoch': epoch,
'state_dict': model.module.state_dict(),
'best_auc': best_auc,
}, os.path.join(conf.best_model_path, 'model_best.pth'))
torch.save({
'epoch': epoch,
'state_dict': model.module.state_dict(),
'best_auc': best_auc,
}, os.path.join(conf.checkpoint_path, 'checkpoint{}.pth'.format(epoch)))
if rank == 0:
if os.path.exists(checkpoint_path) is True:
os.remove(checkpoint_path)
cleanup()