Optical Neural Network simulator and FatNet

This repository contains the library, which serves as the simulator of the 4f free-space system. It uses the custom convolutional layer in PyTorch, which uses the simulation of the light propagation in the 4F device.

4f system, is an optical system containing the laser, light modulators, a pair of lenses and the camera. This system can accelerate the speed of inference of the convolution operation compared to the traditional on GPU convolution.

FatNet is the type of the neural network transformation, which can potentially make the classifier convolutional neural network more compatible with the free-space optics. The description of FatNet can be found in the paper described in the "Paper" section.

Install using pip

pip install optnn

Link to PyPI

How to Run

Simulation of 4f system

If using PyTorch, the layer can be used in the similar way as the standrd Conv2D. It has to be imported as:

from optnn import OpticalConv2d

Optical convolutional layer structure is:

OpticalConv2d(input_channels, output_channels, kernel_size, is_bias=True, pseudo_negativity=False, input_size=28, noise=True)

PS: if input_size of the layer has to be passed, the default value for input is 28x28.

Examples and FatNet

To run the example models and FatNet described in the paper use the command below by running the MAIN.py file:

python3 MAIN.py --batchsize=(your batch size here) --model (model name) --epochs=(number of epochs)

The models you can use are:

• RESNET
• FATNET
• OPTICALFATNET

PS: Optical Fatnet is the only model that will use the simulator

Paper

Title: FatNet: High Resolution Kernels for Classification Using Fully Convolutional Optical Neural Networks

Authors: Riad Ibadulla, Thomas M. Chen and Constantino Carlos Reyes-Aldasoro

Correspondence: riad.ibadulla@city.ac.uk