Pre-built segmentation model training is slow

Hi,

I’m training a segmentation model, basically torchvision.models.segmentation.deeplabv3_resnet50
It takes ~45 minutes for one epoch. I feel this is too slow. I hope someone can help me find redundancy/mistakes (if any) in my codes. Or any suggestions on how to speed up each epoch are appreciated.

  • Image shape: 512x512
  • Training size: 20k
  • Batch size: 8
  • GPU: NVIDIA Tesla V100
  • Memory-Usage: 11805MiB / 16160MiB
  • GPU-Util: 85%-100%

My codes almost followed this tutorial


model = models.segmentation.deeplabv3_resnet50(pretrained=1, progress=0, num_classes=10)
model.to(device)
optimizer = optim.Adam(model.parameters(), lr=0.0001)
lr_scheduler = optim.lr_scheduler.ExponentialLR(optimizer=optimizer, gamma=0.8)

for epoch in range(num_epochs):

    for phase in ['train', 'test']:
        if phase == 'train':
            model.train()
        else:
            model.eval()

        for batch in dataloader_dict[phase]:
                
            image, label = batch['image'].to(device), batch['label'].to(device)
                
            optimizer.zero_grad()
                
            # forward pass
            with torch.set_grad_enabled(phase == 'train'): 

                output = model(image)['out']  
                loss = criterion(output, label)
                prediction = output.argmax(dim=1)  
                    
                if phase == 'train':
                    loss.backward()
                    optimizer.step()

        if phase == 'train':
            lr_scheduler.step()