06 June 2011
A neural network implementation made for the Machine Learning course at the University of Pisa.
Allows you to create and test a neural network for classification with the following characteristics:
- The outputs are in the range (0,1)
- Each layers is fully connected with the follow layer and only to that
- The activation function of each unit is the sigmoid function f(x) = 1/(1+e^(-x))
- The weights of each unit are initialized randomly in the range [-0.7,+0.7] except the 0.
- Allows an arbitrary number of inputs, outputs and hidden layers with an arbitrary number of units
Depending on the mode selected (with parameter --mode) you can:
- Create a custom neural network and train it with the back-propagation algorithm on a dataset passed in csv format (mode training)
- Test a neural network, previously created, on a dataset in csv format (mode test)
In general, the format of the file containing the dataset must be:
id, input[1], ..., input[n], output[1], ..., output[m]
where
- each row is an instance with a unique identifier id
- one instance has n inputs and m outputs
- each instance has the same number of inputs and outputs to the others
In test mode you can have a dataset without outputs to get the answers by the neural network.
See help.txt for more.