pytorch 如何使用amp进行混合精度训练

def train():
    batch_size = 8
    epochs = 10
    lr = 1e-3
    size = 256
    num_class = 35
    use_amp = True 
    device = 'cuda' if torch.cuda.is_available() else 'cpu' 
    print('torch version: {}'.fORMat(torch.__version__))
    print('amp:           {}'.format(use_amp))
    print('device:        {}'.format(device))
    print('epochs:        {}'.format(epochs))
    print('learn rate:    {}'.format(lr))
    print('batch size:    {}'.format(batch_size))
 
    net = ERFNet(num_classes=num_class).to(device) 
    train_data = CityScapesDataset('D:\\dataset\\cityscapes',
                                   'D:\\dataset\\cityscapes\\trainImages.txt',
                                   'D:\\dataset\\cityscapes\\trainLabels.txt',
                                   size, num_class)
    val_data = CityScapesDataset('D:\\dataset\\cityscapes',
                                 'D:\\dataset\\cityscapes\\valImages.txt',
                                 'D:\\dataset\\cityscapes\\valLabels.txt',
                                 size, num_class)
 
    train_dataloader = DataLoader(train_data, batch_size=batch_size, shuffle=False, num_workers=8)
    val_dataloader = DataLoader(val_data, batch_size=batch_size, shuffle=False, num_workers=4)
 
    opt = torch.optim.Adam(net.parameters(), lr=lr)
    criterion = torch.nn.CrossEntropyLoss(ignore_index=255)
 
    if use_amp:
        scaler = torch.cuda.amp.GradScaler() 
    writer = SummaryWriter("summary") 
    train_loss = AverageMeter()
    val_acc = AverageMeter()
    val_miou = AverageMeter()
 
    for epoch in range(0, epochs):
        train_loss.reset()
        val_acc.reset()
        val_miou.reset()
 
        with tqdm(total=train_data.__len__(), unit='img', desc="Epoch {}/{}".format(epoch + 1, epochs)) as pbar:
            # train
            net.train()
            for img, mask in train_dataloader:
                img = img.to(device)
                mask = mask.to(device)
                n = img.size()[0] 
                opt.zero_grad()
 
                if use_amp:
                    with torch.cuda.amp.autocast():
                        output = net(img)
                        loss = criterion(output, mask)
 
                    scaler.scale(loss).backward()
                    scaler.step(opt)
                    scaler.update()
                else:
                    output = net(img)
                    loss = criterion(output, mask)
                    loss.backward()
                    opt.step() 
                train_loss.update(loss.item(), n)
 
                pbar.set_postfix(**{"loss": train_loss.avg})
                pbar.update(img.size()[0])
 
            writer.add_Scalar('train_loss', train_loss.avg, epoch)
            # eval
            net.eval()
            for img, mask in val_dataloader:
                img = img.to(device)
                mask = mask
                n = img.size()[0] 
                output = net(img) 
                pred_mask = torch.softmax(output, dim=1)
                pred_mask = pred_mask.detach().cpu().numpy()
                pred_mask = np.argmax(pred_mask, axis=1)
                true_mask = mask.numpy()
                acc, acc_cls, mean_iu, fwavacc = evaluate(pred_mask, true_mask, num_class)
 
                val_acc.update(acc)
                val_miou.update(mean_iu)
 
            writer.add_scalar('val_acc', val_acc.avg, epoch)
            writer.add_scalar('val_miou', val_miou.avg, epoch) 
            pbar.set_postfix(**{"loss": train_loss.avg, "val_acc": val_acc.avg, "val_miou": val_miou.avg})