Implement backward for Bhattacharyya distance

[Saeed Izadi]

Hi, 

I aim to implement my custom Bhattacharyya loss function using python layer. I've seen the following examples as guides:

https://github.com/BVLC/caffe/blob/master/examples/pycaffe/layers/pyloss.py

https://github.com/BVLC/caffe/issues/4023

Here is the cost function and its partial derivative for stochastic gradient decent:

I have implemented the forward function, but don't know how to deal with the backward function. Please help me with this implementation.

Thanks