# Generate list of filenames from your dataset
image_list = pytorch_superpixels.list_loader.ImageList(
    'pascal-seg', './VOCdevkit/VOC2012', 'trainval')
# Use this list to create and save 100 superpixel dataset
pytorch_superpixels.preprocess.create_masks(image_list, 100)

# -----------------------------------------------
# code that sets up model, optimizer, dataloader, metrics, etc.
# -----------------------------------------------

# Iterate through images, labels and masks
# (Note that masks and labels are already in superpixel form)
for (images, labels, masks) in trainloader:
        # Set model to training mode
        model.train()
        # Send images, labels and masks to the GPU
        images = images.to('cuda:0')
        labels = labels.to('cuda:0')
        masks = masks.to('cuda:0')
        # Pass images through the model
        optimizer.zero_grad()
        outputs = model(images)
        # Convert outputs to superpixel form
        outputs_sp, sizes = pixels_to_superpixels(
            outputs, masks)
        # Calculate loss using size weighted superpixels
        loss = loss_fn(scores=outputs_sp, target=labels, size=sizes)
        # Convert back to pixels for metric evaluation
        outputs = superpixels_to_pixels(outputs_sp, masks)
        # Accumulate train metrics during train
        pred = outputs.data.max(1)[1].cpu().numpy()
        gt = labels.data.cpu().numpy()
        running_metrics_train.update(gt, pred)
        train_loss_meter.update(loss.item())
        # Backprop and optimizer step
        loss.backward()
        optimizer.step()