Re: Working on nnet package

[Mike Miller]

On Thu, Apr 11, 2019 at 19:05:09 -0400, Mae wrote:
> 
> > On  11, 2019, at 18:42, Alois Schlögl <address@hidden> wrote:
> > 
> > A much more promising approach seems incorporating some open sourced
> > Deep Learning framework into Octave. E.g. setting up an interface to
> > Tensorflow or ONNX through mex/oct C/C++ interface would be an
> > interesting task. Looking at the c_api of TF 1.13, it seems that one
> > need to get familiar with dataflow graphs, and protobuf, etc. Matlab
> > uses ONNX, so for compatibility one might use ONNX; otoh, Google's
> > Tensorflow seems to be technologically more advanced, therefore, I'd go
> > with a mex interface to tensorflow. It's certainly not a trivial
> > project, but it should be possible. There are already TF interfaces for
> > a number of other languages, so way not Octave as well.
> 
> As a person who uses pytorch for experimental deep learning experiments
> it also is an attractive target which seems to be the most popular (at
> least in research) in recent months.

It should be possible to use either TensorFlow or PyTorch directly today
with Octave and the Python interface. I have not worked with either, but
just glancing at a tutorial, the following works in Octave for me

    >> x1 = py.tensorflow.constant(py.list([1, 2, 3, 4]));
    >> x2 = py.tensorflow.constant(py.list([5, 6, 7, 8]));
    >> result = py.tensorflow.multiply(x1, x2);
    >> session = py.tensorflow.Session();
    >> session.run(result)
    ans = [Python object of type numpy.ndarray]
    
      [ 5. 12. 21. 32.]
    
    >> session.close()

I would be happy to work with anyone interested to develop better
motivating examples that can demonstrate how to leverage these popular
Python libraries in Octave.

-- 
mike