Two Hidden-Layer Perceptron in Java

This repository contains a Java implementation of a two hidden-layer perceptron that can conduct some image analysis (e.g. counting how many fingers are in a picture of a hand). All code was implemented from scratch, with the only external library used being a JSON parsing library.

Usage Instructions

External Libraries

This project is dependent on the json-simple Java library, which can be downloaded here.

Configuration

All configuration options can be set in a JSON file passed as a command line option when running the program. These can be placed in the src/config/ directory. The program comes with a default configuration file and presets for basic boolean operators as well as image recognition. The default configuration options, located in the src/defaults/ directory, should only be edited with caution. A list of configuration options can be found below.

Testing/Training Sets

The first line of any training/testing sets file should contain the number of sets in the file. Each successive line should contain a single set, formatted as a list of inputs separated by spaces followed by a list of outputs separated by spaces.

Configuration Options

General Network Configuration

• train (boolean): true if training the network upon the set, false if testing the network upon the set
• nodes (int[]): the number of nodes in each layer of the network (nodes[0] corresponds to the input layer, nodes[1-2] correspond to the hidden layer, and nodes[3] corresponds to the output layer)
• printDetailed (boolean): true if the network should have more detailed output while training/testing, false otherwise
• weightsPath (String): a path to the text file containing the weights that the network should start with, if any
• setsPath (String): a path to the text file containing the training/testing sets for the network

Training Configuration

• useTrainingWeights (boolean): true if the network should begin training with predetermined weights, false if beginning with random weights
• lambda (int): the learning rate for the network
• maxIterations (int): the maximum amount of training cycles the network should go through
• errorThreshold (double): the goal threshold error to be met through training
• minRandom (double): the lower bound of random generation for weights
• maxRandom (double): the upper bound of random generation for weights
• saveWeights (boolean): true if the weights should be saved after training, false otherwise
• savedWeightsPath (String): a path to the text file where the weights should be saved after training
• autosaveInterval (int): the number of iterations that must occur before the weights are automatically saved