Does SlidingWindowInferer support test time augmentation?

[lizhuoq]

[lizhuoq]

The following code solves my problem

def predict(data, model, model_name, batch_size=32, n_fold=0, device='cuda', tta=False):
    model     = model.to(device) # load the model into the GPU
    model.load_state_dict(torch.load(os.path.join(model_name + str(n_fold), 'checkpoint.pth')))
    
    model.eval()
    with torch.no_grad():
        inferer = SlidingWindowInferer(roi_size=(128, 128), sw_batch_size=batch_size, overlap=0.5, mode="gaussian", progress=True, sw_device=device, device=torch.device('cpu'))
        outputs = inferer(data, model)
        outputs = torch.softmax(outputs, dim=1)
    if tta:
        tta_list = [Flip(spatial_axis=0), Flip(spatial_axis=1), Compose([Flip(spatial_axis=0), Flip(spatial_axis=1)])]
        tta_res = [outputs]
        for aug in tta_list:
            with torch.no_grad():
                inferer = SlidingWindowInferer(roi_size=(128, 128), sw_batch_size=batch_size, overlap=0.5, mode="gaussian", progress=True, sw_device=device, device=torch.device('cpu'))
                transformed_data = aug(data[0]).unsqueeze(0)
                outputs = inferer(transformed_data, model)
                outputs = aug.inverse(outputs[0]).unsqueeze(0)
                outputs = torch.softmax(outputs, dim=1)
                tta_res.append(outputs)
            gc.collect()
        outputs = torch.stack(tta_res, dim=0).mean(dim=0)

    return outputs