Branch yang

[jinglunYang]

I have pulled request again, please to check it. And i have checked the emd.cuh is in pkg/include/cuda/

[louis-cl]

The code outputs a value (hopefully correct ?) for big tensors but the gradient fails, even for small inputs.
In that case my error is:
THCudaCheck FAIL file=.../THCReduceAll.cuh line=327 error=4 : unspecified launch failure

[jinglunYang]

Maybe you should adjust the version of your CUDA & Cudnn & Pytorch. They should be satisfied with a certainly relationship. Can you tell me about your environment?

[louis-cl]

I am using Python 3.6.8, Pytorch 1.1.0, CUDA 9.0.176 and Cudnn 7501.

[meder411]

Sorry I've been backed up with work. I will do my best to get to this ASAP.

[meder411]

This doesn't repair the bug necessarily. It just fixes the number of dimensions to be 3, as opposed to some arbitrary dimension D.

[rjtshrm]

Hi @jinglunYang , I picked you commit with changes. I can run test_emd_loss.py file. But while using the loss in my training set it gives me error RuntimeError: cuDNN error: CUDNN_STATUS_NOT_INITIALIZED. I already tested my training data with Chamfer Loss and it was working fine.