L_Softmax_TensorFlow

TensorFlow version of L_SoftMax.

Results:

I found prelu is quite stable than relu, so I used prelu as paper said.

Reference:

• mx-lsoftmax
• Large-Margin Softmax Loss for Convolutional Neural Networks

Contribution

This is mainly implemented by py_func, which is quite slow. If anyone have implemented a tf_op in C++ or cuda, pull request is warmly welcome.