/SpeechRecognitionUsingNeuralNetwork

feedforward Neural Network with back propagation. Bipolar logistic function as activation function

Primary LanguageC++

Neural Network

EXAMPLE TO TRAIN AND TEST NEURAL NETWORK
  • created by Mitesh Patel on <March 2007>
  • amended by Mitesh Patel on <Feb 2014>
PRE-REQUISITE
  • boost/ublas libraries 1.5 or greater (older libraries should work but haven’t tested)
  • gcc4.3 compiler
CURRENT NEURAL NETWORK (NN) ARCHITECTURE

  • Current code can only handle one hidden layer

  • Current setup required the data be separated training-validation-testing before hand

  • The NN architecture is limited to a feed forward neural network with back propagation.

  • The NN uses bipolar logistic function as the activation function.

    INPUT NODES:

    • The number of input nodes of the NN is calculated from the input of the training data file.
    • It used the length of the rows in the data file to determine the number of input nodes HIDDEN NODES:
    • The default number of hidden nodes in the single layer are calculated using the formula as under:
    • (hiddenNodes_ = ceil((pow(outputNodes_,2.0) + outputNodes_+ 2)/2)+1.
    • The number of hidden nodes can also be specified through the command line. OUTPUT NODES:
    • The number of output nodes correspond to the number of classes to be predicted.

    DATA FORMAT: TRAINING NEURAL NETWORK

    • When training the NN, the program expects five types of files.

    • Training data file

    • training label file

    • validation data file

    • validation label file

    • the name of the model file in which the trained parameters of the NN will be saved.

      TRAINING DATA FILE:

      • Each row of the training data file is a sample for training and the columns are the features. TRAINING LABEL FILE:
      • Each row in the label file represents the label of the corresponding sample in the training data file. VALIDATION DATA FILE:
      • Each row in the label file represents the label of the corresponding sample in the data file. VALIDATION LABEL FILE:
      • Each row in the label file represents the label of the corresponding sample in the validation data file. MODEL FILE:
      • The model file specified is used to store the trained parameters of the Neural Network.

    TESTING NEURAL NETWORK

    • When testing the NN, the program expects three types of files.
    • Testing data file
    • Testing label file
    • the name of the model file from where the NN parameters are loaded. TESTING DATA FILE:
      • Each row in the label file represents the label of the corresponding sample in the data file. TESTING LABEL FILE:
      • Each row in the label file represents the label of the corresponding sample in the testing data file. MODEL FILE:
      • The trained parameters of the NN are loaded from this file.
COMMAND LINE ARGUMENTS
FOR TRAINING
* ./NeuralNetwork -t train [options] training_data.txt training_label.txt validation_data.txt validation_label.txt trained_model.txt
* [options]
* "-h number of hidden_nodes : (default calculated using (hiddenNodes_ = ceil((pow(outputNodes_,2.0) + outputNodes_+ 2)/2)+1 \n"
* "-c training cycles : iteration for optimising the weights of NN (default 300)\n"
* “"-v displays NN parameters : displays the trained parameters of the model (default will not display)\n"
FOR TESTING
* ./NeuralNetwork -t test testing_data.txt testing_label.txt trained_model.txt