datasuperman/convolution-neural-network

CNN in Java

Convolution-Neural-Network

CNN in Java

#Modifying the network To modify the architecture just open the MyNeuralNetwork class under test folder. Just to clarify softmax doesn't seem to work properly.

Training

MyNeuralNetworkTrain class under the test folder, all the hyperparameters can be defined there. Just a thing this class saves all the weights under weights folder. Typically if a folder is named as "weights-16.20000" it means the network is stored for which validation accuracy was 16.20000% .Current weights are saved under weights folder. Everytime one stops the network, weights are saved inside the folder and if you restart the MyNeuralNetworkTrain class it will start from those weights. If you want to start from random initialization just delete all 4 files under weights folder.

Test

MyNeuralNetworkTest : This class does the forward pass only and predicts the labels, output values are saved in in the "Yte.csv" file under test folder. One can specify which accuracy weight folder he/she wants to use simply by changing the name of the folder inside.

This project is best viewed with IntelliJ IDE.

Helper Files

All the files under the Mosig folder are the helper files for the project to run.