Inference code shape mismatches

[sabetAI]

Getting the error below when I attempt to run python inference.py --input_dir examples/ --output_dir output/

/mnt/asabet/IntrinsicCompositing/intrinsic_compositing/shading/pipeline.py:108: UserWarning: Using a target size (torch.Size([169247, 352])) that is different to the input size (torch.Size([169247])). This will likely lead to incorrect results due to broadcasting. Please ensure they have the same size. loss = torch.nn.functional.mse_loss(pred_shd.reshape(-1), b) Traceback (most recent call last): File "/mnt/asabet/IntrinsicCompositing/inference/inference.py", line 201, in <module> coeffs, lgt_vis = get_light_coeffs( File "/mnt/asabet/IntrinsicCompositing/intrinsic_compositing/shading/pipeline.py", line 138, in get_light_coeffs init_loss = test_init(params, A, b) File "/mnt/asabet/IntrinsicCompositing/intrinsic_compositing/shading/pipeline.py", line 108, in test_init loss = torch.nn.functional.mse_loss(pred_shd.reshape(-1), b) File "/mnt/asabet/IntrinsicCompositing/.venv/lib/python3.10/site-packages/torch/nn/functional.py", line 3365, in mse_loss expanded_input, expanded_target = torch.broadcast_tensors(input, target) File "/mnt/asabet/IntrinsicCompositing/.venv/lib/python3.10/site-packages/torch/functional.py", line 76, in broadcast_tensors return _VF.broadcast_tensors(tensors) # type: ignore[attr-defined] RuntimeError: The size of tensor a (169247) must match the size of tensor b (352) at non-singleton dimension 1

If I fix this bug by reshaping the pred_shd tensor, I run into errors elsewhere, ie

File "/mnt/asabet/IntrinsicCompositing/intrinsic_compositing/shading/pipeline.py", line 153, in get_light_coeffs return coeffs, out_shd.reshape(shd.shape) ValueError: cannot reshape array of size 174592 into shape (512,341,352)

Code would be easier to execute if exact dependencies were pinned. ie there are breaking changes made to https://github.com/compphoto/intrinsic, which have to be manually set to a specific commit. Also unclear which version of torch to use, etc.

[wjd-322418]

hello i also meet this prblem,did you solve it?

[CCareaga]

Hi, sorry for the delay. I've been meaning to fix this, so thank you for bringing it up.

For now, I've pinned the version of the intrinsic decomposition pipeline to a specific commit before we updated the method. It should work if you run the pip installation command after pulling this change. As for the version s of libraries I used, it's not super strict and should work with the latest versions of pytorch, numpy, etc. I can add information in the repo for which versions are reasonable though.

Let me know if the issue still persists.