Computed gradients for a keras model in MATLAB

MATLAB via NN Toolbox now supports directly importing and working with keras models. Is it possible to compute gradients w.r.to input/output of any layer, similar to keras backend function K.gradients() ?

Answer 1

[Don Mathis]

If you're referring to the Tensorflow-Keras Importer support package, that imports a Keras model into MATLAB and returns a MATLAB (Neural Network Toolbox) model. The Keras model is not retained; it is translated into a MATLAB model, so you can only perform operations defined on MATLAB models from that point onward.

[James Walt]

@Don Mathis Thanks for a response, and clarification. Sorry but I still don’t have the full answer since I am new to MATLAB. I saw documentation which mentions defaultderiv() can be used to compute gradients but I am not sure how to use it to compute say dy_i/dx_ik where y_i is the ith target class/node and x_ik is the ith input/feature map at layer k.

Thanks