00:44 < zoq> luffy1996: I'm not sure I get the jacobian matrix part.

05:25 < luffy1996> template<typename InputDataType, typename OutputDataType>

05:25 < luffy1996> template<typename InputType, typename OutputType>

05:25 < luffy1996> void Softmax<InputDataType, OutputDataType>::Forward(

05:25 < luffy1996> const InputType&& input, OutputType&& output)

05:25 < luffy1996> {

05:25 < luffy1996> // This is how I am computing forward pass for softmax layer on mlpack

05:25 < luffy1996> arma::mat maxInput = arma::repmat(arma::max(input), input.n_rows, 1);

05:25 < luffy1996> arma::mat expInput = arma::exp(maxInput - input);

05:25 < luffy1996> // We will normalize the values to get probability.

05:25 < luffy1996> double sumExpInput = arma::sum(expInput);

05:25 < luffy1996> output = expInput/sumExpInput;

05:25 < luffy1996> }

05:25 < luffy1996> https://www.irccloud.com/pastebin/Ttb9MeK5/

05:26 < luffy1996> zoq : Please go throught the snippet. The about code wasn't sent successfully because of flooding

05:27 < luffy1996> @zoq, I want a softmax layer to be implemented which should work with categorical crossentropy. I want to send a one hot bit vector at the time of back propagation. Would you please help me in implementing this

15:04 < dmatt> .

16:20 < zoq> luffy1996: For the backward path compute the derivative, see https://github.com/mlpack/mlpack/blob/master/src/mlpack/methods/ann/activation_functions/logistic_function.hpp for an example. Once we have the derivate we can implement it as a sperate layer or use the base layer.

19:19 < Atharva> zoq: Please look at the input/output PR when you get time. If the approach is okay then I will proceed and the required constructors for the other layers.