Neural Networks Functions Convert to C++ codes

*Hi All,

I am using the Neural Networks Toolbox to design and test the Neural Networks implementations. However, the Matlab codes does not suffice the speed and memory requirement of a very large networks (with 100,000+ neurons). I am exploring ways to convert the Matlab Neural Networks implementations to C++ implementations. Can the Codes Generation Toolbox help me to do that? Has anyone tried to convert the Matlab Neural Network codes to C++ before?

I converted the Feedforward Network and the trainlm() for Back Propagation algorithm to Visual C++ implementations 9 years ago by hand. That took me 2 full days of work. And the C++ codes was about 10 to 30 times faster and used 1/5 of the memory than the Matlab implementations.

However, with much more complexed networks like the NARX with Conjugate Gradient training algorithms, I think the codes would be much much more complicated than the LM and Feedforward networks that I converted years back. So, I think it would only be practical to utilize the Coder Generation Toolbox to achieve the Matlab to C++ codes convertion.

Appreciate in advance for your inputs in this topic. Meanwhile, I am happy to share with you any progress that I might make along the way.

Thanks,

Baker Huang*

Answer 1

[Friedrich]

Hi,

the MATLAB Coder won't help you here since the Neual Network Toolbox is not supported for code generation:

http://www.mathworks.com/help/toolbox/eml/ug/bq1h2z8-12.html

Or if you have some specific function in mind:

http://www.mathworks.com/help/toolbox/eml/ug/bq1h2z7-11.html

So seems like you have to start writing from the scratch again.

Answer 2

[Yu Huang]

Hi Friedrich,

Oh, thanks for the quick answer. Gee, I think I would need a few months to do that from scratch!

There would be hundreds of functions that I need to convert and test.

Thanks, anyway.