Train Network using GPU with customlayer

Hi I wrote an custom layer which is about spherical CNN, but when I try it on GPU it returns this error messagee

 Error using trainNetwork (line 140)
Unexpected error calling cuDNN: CUDNN_STATUS_NOT_SUPPORTED.

 Caused by:
    Error using nnet.internal.cnngpu.reluBackward
    Unexpected error calling cuDNN: CUDNN_STATUS_NOT_SUPPORTED.

When I try CPU to train it, it works but very slow. How should I fix it? I can upload the code for my layer if needed.

Thanks:)

[Rui Xiang]

Is it because of the relu layer? Should I custom a relulayer according to my data structure as well?

[Joss Knight]

You would have to show us what your code is doing. Presumably you are passing invalid data out of layer and into the succeeding activation layer. You need to return a 4-D array of single precision values. Have you run checkLayer?

[Rui Xiang]

It is 4D array, and it can be runned under cpu. And my matlab is 2017b so it doesn't have the checklayer function.