I have an implementation I wrote in C++ with Python bindings.  You can specify an arbitrary structure for the neural network using a dot schema.  It also supports recursion. Its part of a more general open source project of mine called "AILib". You can download it here:

http://kovachtech.com/downloads.html

DISCLAIMER: its still in beta mode, so forgive the occasional compiler error, especially if you're using Windows (its been tested in Ubuntu 12.04 and Mac OS X).

If you are interested in something a solution a little more 'out of the box', R has a package called 'nnet' which has a parameter called 'hidden' in which you can set the number of hidden layers.  There's a pretty good article about it here:

https://beckmw.wordpress.com/tag/nnet/

Dear Sir,

This tool may be useful for you: http://www.heatonresearch.com/encog

Kind regards,

Olivier PARISOT

You may want to check out packages like Theano, Pylearn2 and PyBrain (discontinued).

http://deeplearning.net/software/theano/

http://deeplearning.net/software/pylearn2/

http://pybrain.org/

Thank you all for your contributions. I am most grateful.

With Matlab  software and Artificial neural network Toolbox , you can write your own program to training the ANN

Alfonso

I found tensor flow is there and it can be implemented with python https://www.tensorflow.org/versions/r0.10/get_started/index.html 

Thank you Sandeep.