def train(args, train_dataloader, valid_dataloader):
if str(args.model).lower() == 'fcn32s':
model = VGG16_FCN32s(n_classes=7)
elif str(args.model).lower() == 'fcn8s':
model = VGG16_FCN8s(n_classes=7)
else:
model = UNet(n_channels=3, n_classes=7)
#model = nn.DataParallel(model, device_ids=['cuda:0','cuda:1'])
model.to(args.device)
# loss
# 0.79, 0.14, 1.0, 0.73, 2.74, 1.04, 132, 0
weight = torch.tensor([0.79, 0.14, 1.0, 0.73, 2.74, 1.04, 1.0])
criterion = nn.CrossEntropyLoss(weight).to(args.device)
# optim
optimizer = optim.SGD(model.parameters(),
lr=1e-3,
momentum=0.9,
weight_decay=5e-4)
if str(args.model) == 'fcn32s':
milestones = [1, 10, 20, 50]
elif str(args.model) == 'fcn8s':
milestones = [1, 10, 20, 60]
else:
milestones = [25, 50, 80]
train_scheduler = optim.lr_scheduler.MultiStepLR(
optimizer, milestones=milestones, gamma=0.2) #learning rate decay
best_iou = 0
for epoch in range(args.epochs):
print(f"\tEpoch {epoch}")
loss, acc, iou = _run_train(args, train_dataloader, model, criterion,
optimizer)
print("\t train loss:{:.5f}, acc:{:.3f}, iou:{:.2f}".format(
loss, acc, iou))
loss, acc, iou = _run_eval(args, valid_dataloader, model, criterion)
print("\t valid loss:{:.5f}, acc:{:.3f}, iou:{:.2f}".format(
loss, acc, iou))
if epoch in milestones:
torch.save(model.state_dict(),
f"./result/{epoch}_{args.model}.pth")
print('\t [Info] save weights')
if epoch > milestones[1] and iou > best_iou:
best_iou = iou
torch.save(model.state_dict(), f"./result/best_{args.model}.pth")
print('\t [Info] save weights')
def get_segment_model_and_criterion(device):
"""
Create U-NET and changes fully connected layer.
Parameters
----------
device: torch.device
Define CPU or GPU will be used for training
Returns
-------
model architecture and criterion in tuple
"""
model = UNet(n_channels=3, n_classes=1)
model = model.to(device)
criterion = dice
return model, criterion
pin_memory=True,
drop_last=True)
# Create validation dataset
valid_data = dset.ImageFolder(root=args.valid_dir,
transform=default_transform)
valid_loader = torch.utils.data.DataLoader(valid_data,
batch_size=64,
shuffle=True,
num_workers=args.workers,
pin_memory=True,
drop_last=True)
model = UNet(n_channels=3, n_classes=3, bilinear=True)
model.to(args.device)
optimizer = torch.optim.Adam(params=model.parameters(),
lr=args.lr,
weight_decay=0)
state_dict = {'itr': 0}
if args.resume:
print('Loading weights & resuming from iteration {}'.format(
args.checkpoint))
model, optimizer, logger = load_UNET_checkpoint(
model, optimizer, '256', args)
state_dict['itr'] = args.checkpoint
for epoch in range(args.num_epochs):
train_256(epoch, state_dict, model, optimizer, train_loader,
default_transform = transforms.Compose([
transforms.CenterCrop(args.image_size),
transforms.Resize(args.image_size),
transforms.ToTensor()
])
# Create train dataset
train_dataset = dset.ImageFolder(root=args.train_dir, transform=default_transform)
train_loader = torch.utils.data.DataLoader(train_dataset, batch_size=args.batch_size,
shuffle=False, num_workers=args.workers,
pin_memory=True, drop_last=True)
# Create validation dataset
valid_dataset = dset.ImageFolder(root=args.valid_dir, transform=default_transform)
valid_loader = torch.utils.data.DataLoader(valid_dataset, batch_size=args.batch_size,
shuffle=False, num_workers=args.workers,
pin_memory=True, drop_last=True)
model_128 = UNet(n_channels=3, n_classes=3, bilinear=True)
model_128.to(args.device)
model_128 = load_UNET_weights(model_128, '128', args)
model_256 = UNet(n_channels=3, n_classes=3, bilinear=True)
model_256.to(args.device)
model_256 = load_UNET_weights(model_256, '256', args)
eval_unet_128_256(model_128, model_256, train_loader, 'train', args)
eval_unet_128_256(model_128, model_256, valid_loader, 'valid', args)
net.train()
# 按照batch_size开始训练
for image, label in train_loader:
optimizer.zero_grad()
# 将数据拷贝到device中
image = image.to(device=device, dtype=torch.float32)
label = label.to(device=device, dtype=torch.float32)
# 使用网络参数，输出预测结果
pred = net(image)
# 计算loss
loss = criterion(pred, label)
print('Loss/train', loss.item())
# 保存loss值最小的网络参数
if loss < best_loss:
best_loss = loss
torch.save(net.state_dict(), 'best_model.pth')
# 更新参数
loss.backward()
optimizer.step()
if __name__ == "__main__":
# 选择设备，有cuda用cuda，没有就用cpu
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
# 加载网络，图片单通道1，分类为1。
net = UNet(n_channels=1, n_classes=1)
# 将网络拷贝到deivce中
net.to(device=device)
# 指定训练集地址，开始训练
data_path = "E:/AI_data/ISBI/data/train/"
train_net(net, device, data_path)