This directory /Experiments contains codes to reproduce results from the paper. Please refer to the appendix of the paper for instructions.

To use ZORB, please clone this repository and copy /zorb to your workspace directory. You can import zorb to your script and utilize it's functionalities.

Alternatively, you can directly train and evaluate models using command line arguments. The '-n'/'--network' argument can be used to describe your architecture, without the output layer. Use the '-h'/'--help' argument for available arguments and it's usage.

Train a fully connected network with 2 sigmoid-activated hidden layers with 512 and 256 neurons on the MNIST dataset:

cd zorb

python3 zorb-cli.py -d MNIST -n Flatten Dense[512] sigmoid Dense[256] sigmoid

Train a convolutional neural network on the CIFAR10 dataset with 2 sigmoid-activated convolutional layers with 8 and 16 kernels of size 3x3 followed by 1 sigmoid-activated fully connected layer with 512 neurons:

cd zorb

python3 zorb-cli.py -d CIFAR10 -n Convolution[8,3] sigmoid Convolution[16,3] sigmoid Flatten Dense[512] sigmoid

For a fully connected layer, use "Dense[<units>]"

For a convolution layer, use "Convolution[<kernels>,<size(optional,default=3)>,<stride(optional,default=1>]"

For a flatten layer, use "Flatten"

For an activation layer, use "<activation>"