neural-network-cuda

Description

Small MATLAB library implementing neural network training exploiting CUDA, developed during the Cognitive Robotics course at Politecnico di Milano by Armando Pesenti Gritti and Oscar Tarabini.

Documentation

Here is a description of the interface offered by our library. It consists of 4 functions:

• CREATENET
• TRAINNET
• TESTNET
• APPLYNET

CREATENET Create neural network with random initialization
[ net ] = createNet( numInput, numOutput, hiddenLayers, netType, hiddenFunction )

 INPUT:
   numInput 		= number of input neurons
   numOutput 		= number of output neurons
   hiddenLayers 	= row vector containing the number of neurons for each hidden
                  		   layer from input to output (both excluded)
   netType 		= 'classification' or 'regression'
   hiddenFunction	= (optional) activation function for hidden neurons. 'sigmoid' or 'tanh'
 OUTPUT:
   net			= struct containing the neural network ready for the training

TRAINNET Train the network constructed with createNet
[ net , mse] = trainNet( net, samples, targets, gpu, batchSize, numEpochs, learningRate )

 INPUT:
   net 			= network as obtained by createNet
   samples 		= input samples, each row is a sample
   targets 		= output targets, each row is a target
   gpu			= exploit GPU if true
   batchSize 		= (optional) size of the batch, default size min(512, size(samples, 1))
   numEpochs		= (optional) maximum number of epochs, default 10
 OUTPUT:
   net 			= trained net
   mse 			= mean square error in case of regression, and cross entropy in case of
             classification. It's computed on the training samples.

TESTNET Apply the trained neural network to a test dataset, computing the error
 [ error predicted ] = testNet( net, samples, targets, gpu )

 INPUT:
   net 			 = network as obtained by trainNet
   samples 		 = input samples, each row is a sample
   targets		 = outupt targets, each row is a targert
   gpu 			 = exploit GPU if true
 OUTPUT:
   error 			 = scalar representing the mean square error in case of regression
               		    or class error in case of classification
   predicted 		 = size(targets) matrix containing the predicted output for
               		    all input samples

APPLYNET Apply the trained neural network to one or more inputs
[ predicted ] = applyNet( net, inputs, gpu )

 INPUT:
   net 			= network as obtained by trainNet
   inputs 		= input values, each row is an input
   gpu 			= exploit GPU if true
 OUTPUT:
   predicted 		= net.layers(end) matrix containing the predicted output for all input 
   values