Re: Working on nnet package

[Francesco Faccio]

Frankie Robertson wrote:

 

Hi,

I saw your post on the mailing list about resurrecting the nnet package. I have just submitted a few patches against it to the patch tracker on GNU Savannah. I have recently started using this package to try and replace Matlab for a Machine Learning course I'm taking and I'm also interested in improving it. I wonder what is your approach with regards to implementing external libraries. Is the plan to integrate them in addition to Octave only implementations or not? I don't think that, for example, Tensorflow supports Levenberg-Marquardt optimisation so it might not be possible to do everything in terms of 3rd party libraries.

My initial plan before learning of your own efforts was to improve the Octave only implementations to support Matlab 2010 versions of functions and add extra training strategies as well as to make the training/test set/validation stuff more consistent with Matlab. Would any of this work still be useful or would it be made redundant by your own efforts? For that matter would you be interested in a little extra help with part of what you're planning on working on?

Regards,

Frankie Robertson

--

http://frankie.robertson.name/

Hi!

Thank you for showing your interest on this. The deep learning part of the package will rely on 3rd party libraries, but, as you said, it's not possible to do everything, so most of the functionalities will be implemented as m-files. I would say that every train function that is not supported by TensorFlow should be implemented as an m-file (or in C++). This means that any contribution in terms of m-implementations will be highly appreciated.

I checked your patch on newff and the code looks better now. In the next few days I will test it and I will give you a feedback. Although newff has been declared obsolete [1] and has been replaced with feedforwardnet [2], I see that it can still be used in Matlab, so it is nice to keep improving it.

Do you have a specific function you would like o start improving or you would like to implement from scratch?
You can find the complete list of functions already implemented in Matlab here [3].

Thank you again,

Francesco Faccio

[1]https://it.mathworks.com/help/nnet/release-notes.html?searchHighlight=newff&s_tid=doc_srchtitle
[2]https://it.mathworks.com/help/nnet/ref/feedforwardnet.html
[3]https://it.mathworks.com/help/nnet/functionlist-alpha.html