def main(args):
print(args)
# mp.spawn(main_worker, args=(args,), nprocs=args.world_size, join=True)
init_distributed_mode(args)
device = torch.device(args.device)
results_file = "results{}.txt".format(
datetime.datetime.now().strftime("%Y%m%d-%H%M%S"))
# Data loading code
print("Loading data")
data_transform = {
"train":
transform.Compose([
transform.SSDCropping(),
transform.Resize(),
transform.ColorJitter(),
transform.ToTensor(),
transform.RandomHorizontalFlip(),
transform.Normalization(),
transform.AssignGTtoDefaultBox()
]),
"val":
transform.Compose([
transform.Resize(),
transform.ToTensor(),
transform.Normalization()
])
}
VOC_root = args.data_path
# check voc root
if os.path.exists(os.path.join(VOC_root, "VOCdevkit")) is False:
raise FileNotFoundError(
"VOCdevkit dose not in path:'{}'.".format(VOC_root))
# load train data set
train_data_set = VOC2012DataSet(VOC_root,
data_transform["train"],
train_set='train.txt')
# load validation data set
val_data_set = VOC2012DataSet(VOC_root,
data_transform["val"],
train_set='val.txt')
print("Creating data loaders")
if args.distributed:
train_sampler = torch.utils.data.distributed.DistributedSampler(
train_data_set)
test_sampler = torch.utils.data.distributed.DistributedSampler(
val_data_set)
else:
train_sampler = torch.utils.data.RandomSampler(train_data_set)
test_sampler = torch.utils.data.SequentialSampler(val_data_set)
if args.aspect_ratio_group_factor >= 0:
# 统计所有图像比例在bins区间中的位置索引
group_ids = create_aspect_ratio_groups(
train_data_set, k=args.aspect_ratio_group_factor)
train_batch_sampler = GroupedBatchSampler(train_sampler, group_ids,
args.batch_size)
else:
train_batch_sampler = torch.utils.data.BatchSampler(train_sampler,
args.batch_size,
drop_last=True)
data_loader = torch.utils.data.DataLoader(
train_data_set,
batch_sampler=train_batch_sampler,
num_workers=args.workers,
collate_fn=train_data_set.collate_fn)
data_loader_test = torch.utils.data.DataLoader(
val_data_set,
batch_size=1,
sampler=test_sampler,
num_workers=args.workers,
collate_fn=train_data_set.collate_fn)
print("Creating model")
model = create_model(num_classes=args.num_classes + 1, device=device)
model_without_ddp = model
if args.distributed:
model = torch.nn.parallel.DistributedDataParallel(
model, device_ids=[args.gpu])
model_without_ddp = model.module
params = [p for p in model.parameters() if p.requires_grad]
optimizer = torch.optim.SGD(params,
lr=args.lr,
momentum=args.momentum,
weight_decay=args.weight_decay)
lr_scheduler = torch.optim.lr_scheduler.StepLR(optimizer,
step_size=args.lr_step_size,
gamma=args.lr_gamma)
# lr_scheduler = torch.optim.lr_scheduler.MultiStepLR(optimizer, milestones=args.lr_steps, gamma=args.lr_gamma)
# 如果传入resume参数，即上次训练的权重地址，则接着上次的参数训练
if args.resume:
# If map_location is missing, torch.load will first load the module to CPU
# and then copy each parameter to where it was saved,
# which would result in all processes on the same machine using the same set of devices.
checkpoint = torch.load(
args.resume, map_location='cpu') # 读取之前保存的权重文件(包括优化器以及学习率策略)
model_without_ddp.load_state_dict(checkpoint['model'])
optimizer.load_state_dict(checkpoint['optimizer'])
lr_scheduler.load_state_dict(checkpoint['lr_scheduler'])
args.start_epoch = checkpoint['epoch'] + 1
if args.test_only:
utils.evaluate(model, data_loader_test, device=device)
return
train_loss = []
learning_rate = []
val_map = []
print("Start training")
start_time = time.time()
for epoch in range(args.start_epoch, args.epochs):
if args.distributed:
train_sampler.set_epoch(epoch)
mean_loss, lr = utils.train_one_epoch(model, optimizer, data_loader,
device, epoch, args.print_freq)
# only first process to save training info
if args.rank in [-1, 0]:
train_loss.append(mean_loss.item())
learning_rate.append(lr)
# update learning rate
lr_scheduler.step()
# evaluate after every epoch
coco_info = utils.evaluate(model, data_loader_test, device=device)
if args.rank in [-1, 0]:
# write into txt
with open(results_file, "a") as f:
result_info = [
str(round(i, 4))
for i in coco_info + [mean_loss.item(), lr]
]
txt = "epoch:{} {}".format(epoch, ' '.join(result_info))
f.write(txt + "\n")
val_map.append(coco_info[1]) # pascal mAP
if args.output_dir:
# 只在主节点上执行保存权重操作
save_on_master(
{
'model': model_without_ddp.state_dict(),
'optimizer': optimizer.state_dict(),
'lr_scheduler': lr_scheduler.state_dict(),
'args': args,
'epoch': epoch
}, os.path.join(args.output_dir, 'model_{}.pth'.format(epoch)))
total_time = time.time() - start_time
total_time_str = str(datetime.timedelta(seconds=int(total_time)))
print('Training time {}'.format(total_time_str))
if args.rank in [-1, 0]:
# plot loss and lr curve
if len(train_loss) != 0 and len(learning_rate) != 0:
from plot_curve import plot_loss_and_lr
plot_loss_and_lr(train_loss, learning_rate)
# plot mAP curve
if len(val_map) != 0:
from plot_curve import plot_map
plot_map(val_map)
def main(args):
print(args)
# mp.spawn(main_worker, args=(args,), nprocs=args.world_size, join=True)
init_distributed_mode(args)
device = torch.device(args.device)
# Data loading code
print("Loading data")
data_transform = {
"train": transforms.Compose([transforms.ToTensor(),
transforms.RandomHorizontalFlip(0.5)]),
"val": transforms.Compose([transforms.ToTensor()])
}
VOC_root = args.data_path
# check voc root
if os.path.exists(os.path.join(VOC_root, "VOCdevkit")) is False:
raise FileNotFoundError("VOCdevkit dose not in path:'{}'.".format(VOC_root))
# load train data set
# VOCdevkit -> VOC2012 -> ImageSets -> Main -> train.txt
train_data_set = VOC2012DataSet(VOC_root, data_transform["train"], "train.txt")
# load validation data set
# VOCdevkit -> VOC2012 -> ImageSets -> Main -> val.txt
val_data_set = VOC2012DataSet(VOC_root, data_transform["val"], "val.txt")
print("Creating data loaders")
if args.distributed:
train_sampler = torch.utils.data.distributed.DistributedSampler(train_data_set)
test_sampler = torch.utils.data.distributed.DistributedSampler(val_data_set)
else:
train_sampler = torch.utils.data.RandomSampler(train_data_set)
test_sampler = torch.utils.data.SequentialSampler(val_data_set)
if args.aspect_ratio_group_factor >= 0:
# 统计所有图像比例在bins区间中的位置索引
group_ids = create_aspect_ratio_groups(train_data_set, k=args.aspect_ratio_group_factor)
train_batch_sampler = GroupedBatchSampler(train_sampler, group_ids, args.batch_size)
else:
train_batch_sampler = torch.utils.data.BatchSampler(
train_sampler, args.batch_size, drop_last=True)
data_loader = torch.utils.data.DataLoader(
train_data_set, batch_sampler=train_batch_sampler, num_workers=args.workers,
collate_fn=train_data_set.collate_fn)
data_loader_test = torch.utils.data.DataLoader(
val_data_set, batch_size=1,
sampler=test_sampler, num_workers=args.workers,
collate_fn=train_data_set.collate_fn)
print("Creating model")
model = create_model(num_classes=21)
model.to(device)
model_without_ddp = model
if args.distributed:
model = torch.nn.parallel.DistributedDataParallel(model, device_ids=[args.gpu])
model_without_ddp = model.module
params = [p for p in model.parameters() if p.requires_grad]
optimizer = torch.optim.SGD(
params, lr=args.lr, momentum=args.momentum, weight_decay=args.weight_decay)
# lr_scheduler = torch.optim.lr_scheduler.StepLR(optimizer, step_size=args.lr_step_size, gamma=args.lr_gamma)
lr_scheduler = torch.optim.lr_scheduler.MultiStepLR(optimizer, milestones=args.lr_steps, gamma=args.lr_gamma)
# 如果传入resume参数，即上次训练的权重地址，则接着上次的参数训练
if args.resume:
# If map_location is missing, torch.load will first load the module to CPU
# and then copy each parameter to where it was saved,
# which would result in all processes on the same machine using the same set of devices.
checkpoint = torch.load(args.resume, map_location='cpu') # 读取之前保存的权重文件(包括优化器以及学习率策略)
model_without_ddp.load_state_dict(checkpoint['model'])
optimizer.load_state_dict(checkpoint['optimizer'])
lr_scheduler.load_state_dict(checkpoint['lr_scheduler'])
args.start_epoch = checkpoint['epoch'] + 1
if args.test_only:
utils.evaluate(model, data_loader_test, device=device)
return
print("Start training")
start_time = time.time()
for epoch in range(args.start_epoch, args.epochs):
if args.distributed:
train_sampler.set_epoch(epoch)
utils.train_one_epoch(model, optimizer, data_loader, device, epoch, args.print_freq)
lr_scheduler.step()
if args.output_dir:
# 只在主节点上执行保存权重操作
save_on_master({
'model': model_without_ddp.state_dict(),
'optimizer': optimizer.state_dict(),
'lr_scheduler': lr_scheduler.state_dict(),
'args': args,
'epoch': epoch},
os.path.join(args.output_dir, 'model_{}.pth'.format(epoch)))
# evaluate after every epoch
utils.evaluate(model, data_loader_test, device=device)
total_time = time.time() - start_time
total_time_str = str(datetime.timedelta(seconds=int(total_time)))
print('Training time {}'.format(total_time_str))