def main(args):
init_distributed_mode(args)
print(args)
device = torch.device(args.device)
# 用来保存coco_info的文件
results_file = "results{}.txt".format(
datetime.datetime.now().strftime("%Y%m%d-%H%M%S"))
# Data loading code
print("Loading data")
data_transform = {
"train":
transforms.Compose(
[transforms.ToTensor(),
transforms.RandomHorizontalFlip(0.5)]),
"val":
transforms.Compose([transforms.ToTensor()])
}
COCO_root = args.data_path
# load train data set
# coco2017 -> annotations -> instances_train2017.json
train_data_set = CocoDetection(COCO_root, "train", data_transform["train"])
# load validation data set
# coco2017 -> annotations -> instances_val2017.json
val_data_set = CocoDetection(COCO_root, "val", data_transform["val"])
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")
# create model num_classes equal background + 80 classes
model = create_model(num_classes=args.num_classes + 1, device=device)
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
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,
warmup=True)
# update learning rate
lr_scheduler.step()
# evaluate after every epoch
coco_info = utils.evaluate(model, data_loader_test, device=device)
# 只在主进程上进行写操作
if args.rank in [-1, 0]:
train_loss.append(mean_loss.item())
learning_rate.append(lr)
val_map.append(coco_info[1]) # pascal mAP
# write into txt
with open(results_file, "a") as f:
# 写入的数据包括coco指标还有loss和learning rate
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")
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(parser_data):
device = torch.device(
parser_data.device if torch.cuda.is_available() else "cpu")
print("Using {} device training.".format(device.type))
# 用来保存coco_info的文件
results_file = "results{}.txt".format(
datetime.datetime.now().strftime("%Y%m%d-%H%M%S"))
data_transform = {
"train":
transforms.Compose(
[transforms.ToTensor(),
transforms.RandomHorizontalFlip(0.5)]),
"val":
transforms.Compose([transforms.ToTensor()])
}
VOC_root = parser_data.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_dataset = VOCDataSet(VOC_root, "2012", data_transform["train"],
"train.txt")
train_sampler = None
# 是否按图片相似高宽比采样图片组成batch
# 使用的话能够减小训练时所需GPU显存，默认使用
if args.aspect_ratio_group_factor >= 0:
train_sampler = torch.utils.data.RandomSampler(train_dataset)
# 统计所有图像高宽比例在bins区间中的位置索引
group_ids = create_aspect_ratio_groups(
train_dataset, k=args.aspect_ratio_group_factor)
# 每个batch图片从同一高宽比例区间中取
train_batch_sampler = GroupedBatchSampler(train_sampler, group_ids,
args.batch_size)
# 注意这里的collate_fn是自定义的，因为读取的数据包括image和targets，不能直接使用默认的方法合成batch
batch_size = parser_data.batch_size
nw = min([os.cpu_count(), batch_size if batch_size > 1 else 0,
8]) # number of workers
print('Using %g dataloader workers' % nw)
if train_sampler:
# 如果按照图片高宽比采样图片，dataloader中需要使用batch_sampler
train_data_loader = torch.utils.data.DataLoader(
train_dataset,
batch_sampler=train_batch_sampler,
pin_memory=True,
num_workers=nw,
collate_fn=train_dataset.collate_fn)
else:
train_data_loader = torch.utils.data.DataLoader(
train_dataset,
batch_size=batch_size,
shuffle=True,
pin_memory=True,
num_workers=nw,
collate_fn=train_dataset.collate_fn)
# load validation data set
# VOCdevkit -> VOC2012 -> ImageSets -> Main -> val.txt
val_dataset = VOCDataSet(VOC_root, "2012", data_transform["val"],
"val.txt")
val_data_set_loader = torch.utils.data.DataLoader(
val_dataset,
batch_size=1,
shuffle=False,
pin_memory=True,
num_workers=nw,
collate_fn=val_dataset.collate_fn)
# create model num_classes equal background + 20 classes
model = create_model(num_classes=parser_data.num_classes + 1)
# print(model)
model.to(device)
# define optimizer
params = [p for p in model.parameters() if p.requires_grad]
optimizer = torch.optim.SGD(params,
lr=0.005,
momentum=0.9,
weight_decay=0.0005)
# learning rate scheduler
lr_scheduler = torch.optim.lr_scheduler.StepLR(optimizer,
step_size=3,
gamma=0.33)
# 如果指定了上次训练保存的权重文件地址，则接着上次结果接着训练
if parser_data.resume != "":
checkpoint = torch.load(parser_data.resume, map_location='cpu')
model.load_state_dict(checkpoint['model'])
optimizer.load_state_dict(checkpoint['optimizer'])
lr_scheduler.load_state_dict(checkpoint['lr_scheduler'])
parser_data.start_epoch = checkpoint['epoch'] + 1
print("the training process from epoch{}...".format(
parser_data.start_epoch))
train_loss = []
learning_rate = []
val_map = []
for epoch in range(parser_data.start_epoch, parser_data.epochs):
# train for one epoch, printing every 10 iterations
mean_loss, lr = utils.train_one_epoch(model,
optimizer,
train_data_loader,
device=device,
epoch=epoch,
print_freq=50,
warmup=True)
train_loss.append(mean_loss.item())
learning_rate.append(lr)
# update the learning rate
lr_scheduler.step()
# evaluate on the test dataset
coco_info = utils.evaluate(model, val_data_set_loader, device=device)
# write into txt
with open(results_file, "a") as f:
# 写入的数据包括coco指标还有loss和learning rate
result_info = [
str(round(i, 4)) for i in coco_info + [mean_loss.item()]
] + [str(round(lr, 6))]
txt = "epoch:{} {}".format(epoch, ' '.join(result_info))
f.write(txt + "\n")
val_map.append(coco_info[1]) # pascal mAP
# save weights
save_files = {
'model': model.state_dict(),
'optimizer': optimizer.state_dict(),
'lr_scheduler': lr_scheduler.state_dict(),
'epoch': epoch
}
torch.save(save_files,
"./save_weights/resNetFpn-model-{}.pth".format(epoch))
# 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():
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
print("Using {} device training.".format(device.type))
# 用来保存coco_info的文件
results_file = "results{}.txt".format(
datetime.datetime.now().strftime("%Y%m%d-%H%M%S"))
# 检查保存权重文件夹是否存在，不存在则创建
if not os.path.exists("save_weights"):
os.makedirs("save_weights")
data_transform = {
"train":
transforms.Compose(
[transforms.ToTensor(),
transforms.RandomHorizontalFlip(0.5)]),
"val":
transforms.Compose([transforms.ToTensor()])
}
VOC_root = "./" # VOCdevkit
aspect_ratio_group_factor = 3
batch_size = 8
# 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_dataset = VOCDataSet(VOC_root, "2012", data_transform["train"],
"train.txt")
train_sampler = None
# 是否按图片相似高宽比采样图片组成batch
# 使用的话能够减小训练时所需GPU显存，默认使用
if aspect_ratio_group_factor >= 0:
train_sampler = torch.utils.data.RandomSampler(train_dataset)
# 统计所有图像高宽比例在bins区间中的位置索引
group_ids = create_aspect_ratio_groups(train_dataset,
k=aspect_ratio_group_factor)
# 每个batch图片从同一高宽比例区间中取
train_batch_sampler = GroupedBatchSampler(train_sampler, group_ids,
batch_size)
nw = min([os.cpu_count(), batch_size if batch_size > 1 else 0,
8]) # number of workers
print('Using %g dataloader workers' % nw)
# 注意这里的collate_fn是自定义的，因为读取的数据包括image和targets，不能直接使用默认的方法合成batch
if train_sampler:
# 如果按照图片高宽比采样图片，dataloader中需要使用batch_sampler
train_data_loader = torch.utils.data.DataLoader(
train_dataset,
batch_sampler=train_batch_sampler,
pin_memory=True,
num_workers=nw,
collate_fn=train_dataset.collate_fn)
else:
train_data_loader = torch.utils.data.DataLoader(
train_dataset,
batch_size=batch_size,
shuffle=True,
pin_memory=True,
num_workers=nw,
collate_fn=train_dataset.collate_fn)
# load validation data set
# VOCdevkit -> VOC2012 -> ImageSets -> Main -> val.txt
val_dataset = VOCDataSet(VOC_root, "2012", data_transform["val"],
"val.txt")
val_data_loader = torch.utils.data.DataLoader(
val_dataset,
batch_size=1,
shuffle=False,
pin_memory=True,
num_workers=nw,
collate_fn=val_dataset.collate_fn)
# create model num_classes equal background + 20 classes
model = create_model(num_classes=21)
# print(model)
model.to(device)
train_loss = []
learning_rate = []
val_map = []
# # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # #
# first frozen backbone and train 5 epochs #
# 首先冻结前置特征提取网络权重（backbone），训练rpn以及最终预测网络部分 #
# # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # #
for param in model.backbone.parameters():
param.requires_grad = False
# define optimizer
params = [p for p in model.parameters() if p.requires_grad]
optimizer = torch.optim.SGD(params,
lr=0.005,
momentum=0.9,
weight_decay=0.0005)
init_epochs = 5
for epoch in range(init_epochs):
# train for one epoch, printing every 10 iterations
mean_loss, lr = utils.train_one_epoch(model,
optimizer,
train_data_loader,
device,
epoch,
print_freq=50,
warmup=True)
train_loss.append(mean_loss.item())
learning_rate.append(lr)
# evaluate on the test dataset
coco_info = utils.evaluate(model, val_data_loader, device=device)
# write into txt
with open(results_file, "a") as f:
# 写入的数据包括coco指标还有loss和learning rate
result_info = [
str(round(i, 4)) for i in coco_info + [mean_loss.item()]
] + [str(round(lr, 6))]
txt = "epoch:{} {}".format(epoch, ' '.join(result_info))
f.write(txt + "\n")
val_map.append(coco_info[1]) # pascal mAP
torch.save(model.state_dict(), "./save_weights/pretrain.pth")
# # # # # # # # # # # # # # # # # # # # # # # # # # # #
# second unfrozen backbone and train all network #
# 解冻前置特征提取网络权重（backbone），接着训练整个网络权重 #
# # # # # # # # # # # # # # # # # # # # # # # # # # # #
# 冻结backbone部分底层权重
for name, parameter in model.backbone.named_parameters():
split_name = name.split(".")[0]
if split_name in ["0", "1", "2", "3"]:
parameter.requires_grad = False
else:
parameter.requires_grad = True
# define optimizer
params = [p for p in model.parameters() if p.requires_grad]
optimizer = torch.optim.SGD(params,
lr=0.005,
momentum=0.9,
weight_decay=0.0005)
# learning rate scheduler
lr_scheduler = torch.optim.lr_scheduler.StepLR(optimizer,
step_size=3,
gamma=0.33)
num_epochs = 20
for epoch in range(init_epochs, num_epochs + init_epochs, 1):
# train for one epoch, printing every 50 iterations
mean_loss, lr = utils.train_one_epoch(model,
optimizer,
train_data_loader,
device,
epoch,
print_freq=50,
warmup=True)
train_loss.append(mean_loss.item())
learning_rate.append(lr)
# update the learning rate
lr_scheduler.step()
# evaluate on the test dataset
coco_info = utils.evaluate(model, val_data_loader, device=device)
# write into txt
with open(results_file, "a") as f:
# 写入的数据包括coco指标还有loss和learning rate
result_info = [
str(round(i, 4)) for i in coco_info + [mean_loss.item()]
] + [str(round(lr, 6))]
txt = "epoch:{} {}".format(epoch, ' '.join(result_info))
f.write(txt + "\n")
val_map.append(coco_info[1]) # pascal mAP
# save weights
# 仅保存最后5个epoch的权重
if epoch in range(num_epochs + init_epochs)[-5:]:
save_files = {
'model': model.state_dict(),
'optimizer': optimizer.state_dict(),
'lr_scheduler': lr_scheduler.state_dict(),
'epoch': epoch
}
torch.save(save_files,
"./save_weights/mobile-model-{}.pth".format(epoch))
# 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)
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, device=device)
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))