neural-parser

This is a language for describing a single FFNN (feed forward neural network) or an entity of FFNN (an FFNNE) connected together, train them and test them. It does a lot to the background and made it easier for me to automate the creation of FFNN/FFNNE and assess them.

It is written in perl in around 1995 for the purposes of my PhD research. It probably needs a major rewrite.

Over the years I have added functionality for manipulating data files for training and testing the networks. Most importantly, I have added support for using Support Vector Machines (SVM) and comparing to the neural networks. SVM programs are from Libsvm, http://www.csie.ntu.edu.tw/~cjlin/libsvm, by Chih-Chung Chang and Chih-Jen Lin).

This is my work. You can use and distribute it as you wish but NOT for commercial purposes. Use it at your own risk!

build instructions:

The preferred way to install is to dowmload the tarball distribtution from this repository (https://raw.githubusercontent.com/hadjiprocopis/neural-parser/master/neuralparser-5.0.tar.gz)

tar xvzf neuralparser-5.0.tar.gz
cd neuralparser-5.0
./configure && make clean && make all

how to use this language and tools provided

Here is a first script to start with:

SIN = CreateSingle {
# this is a plain, monolithic feed forward neural network
	SingleType = FFNN;
# oh this is deeeeep neeeet!
#	Arch = 14 53 139 73 31 3;
	Arch = 14 31 3;
# weights file
	Weights = W_SINGLE;
# output sigmoid?
	Sigmoid = Yes;
# this is a classification task
       NumOutputClasses = 2;
}
$ # <<< end of file

Let us save this script to a file called create.np.

Now let's train it:

# the training file
A_FILE = OpenFileObject {
	Filename = training.txt;
}
IncludeFile {
	# read and execute the creation of the neural network
	# saved as SIN
	Filename = create.np;
}
# train the neural network
TrainSingle {
	Obj  = SIN;
	InpFileObj  = A_FILE;
	Iters = 1000;
	Beta = 0.1;
	Lamda = 0.0;
	Seed = 1234;
	# in order to observe evolution of weights we can save them occasionally
	SaveWeightsEveryNIterations = 100;
	# print progress report every so often
	ShowProgressEveryNIterations = 100;
	# to this log file
	ProgressFilename = progress.txt;

	UniqueWeightsFile = no;
	Silent = Yes;
}
# tell us that training has finished, it can also send an email...
SendInformation {
	OutFileName = finished.txt;
	Message = @ Single FFNN;
	Obj = SIN;
}
$

Fine, so let's save this file to train.np and let's proceed to testing (assess the learning performance) of the trained neural network with the following script:

INPUT = OpenFileObject{ Filename = testing.txt; }
IncludeFile {				     
	Filename = create.np;
}
TestSingle {
	Obj = SIN;
	InpFileObj  = INPUT;
	OutFileName = result.txt;
}
$

Now that we have the scripts we can call the interpreter (np) to do the hard work. Note that you should already had installed NNengine (from https://github.com/hadjiprocopis/NNengine) which provides the Neural Network executables.

And here we are doing the training

np -clog training.clog -log training.log train.np

And here is the testing

np -clog training.clog -log training.log test.np

Now, there is a lot more to this. For example input data manipulation and connecting networks together or letting np create a network according to spec.

There is a manual accompanying this software which will give you more details.

Any questions or help please drop me a line.

author: Andreas Hadjiprocopis
[email protected] (ex [email protected])
http://nfkb.scienceontheweb.net (ex http://soi.city.ac.uk/~livantes)