How Can I use sliding_window_inference for a network that has two branches outputting different things?

[OeslleLucena]

Hi,

I am currently having the following error and I would appreciate how I could sort that out.

  File "/home/ol18/miniconda3/envs/tract_uncertainty/lib/python3.6/site-packages/pytorch_lightning/accelerators/gpu_accelerator.py", line 66, in _step
    output = model_step(*args)
  File "/home/ol18/Codes/TractUncertainty/segmentation_sh_torchIO_PIL_MONAI.py", line 281, in validation_step
    outputs = sliding_window_inference(inputs, roi_size, sw_batch_size, self.forward)
  File "/home/ol18/miniconda3/envs/tract_uncertainty/lib/python3.6/site-packages/monai/inferers/utils.py", line 129, in sliding_window_inference
    seg_prob = predictor(window_data, *args, **kwargs).to(device)  # batched patch segmentation
AttributeError: 'tuple' object has no attribute 'to'

Many thanks!

[rijobro]

The output of predictor should be of type torch.Tensor so that it can be copied to the device (e.g., GPU) with .to(device). For whatever reason, in your case predictor is returning a tuple (perhaps multiple torch.Tensor?).

I can't say much else without seeing more code, but I think you need to put in some print statements and check that the return types are as you expect throughout the script.

[OeslleLucena]

@rijobro outputs from predictor was indeed a tuple. I have adapted the script of sliding_window_inference to handle two outputs instead of one suiting my needs. Thanks for the help.

[CataRay]

@OeslleLucena could you share your changes to sliding_window_inference? I am running into a similar problem when trying to get the attention maps out of my network. my question