Coherence-Awareness-in-Diffractive-Neural-Networks

Official pytorch implementation of the paper: "Coherence Awareness in Diffractive Neural Networks"

System Requirements

Hardware Requirements

This code has been tested on a Linux machine (Ubuntu 22.04.4 LTS) with NVIDIA GeForce GTX 1080 Ti GPU.

GPU is not mandatory, however it expadite training.

Software Requirements

This code has the following dependencies:

  python >= 3.8.12 
  torch >= 1.12.1
  torchvision >= 0.13.1
  numpy >= 1.23.4
  tqdm >= 4.64.0

Setting an environment

Create a python virtual environment, install all dependecies using the requirements.txt file and then run the code on your computer.

cd DIR_NAME
python3 -m venv VENV_NAME
source VENV_NAME/bin/activate
pip install -r requirements.txt 

Installation time should take around 10 minutes.

Usage Instructions

After installation one can run our code.

Data

The data used in our work is the MNIST and FashionMNIST datasets. Both datasets are available via torchvision. See get_data.py.

Hyperparameters

config.py include all the hyperparameters used for each trial.

Usage

A usage example can be found in run_trials.py.

Training of a coherence aware diffractive network with two layers and the hyperparameters in the usage exmaple requires approximately 4 GB of memory and takes approximately 25 hours to complete on the mentioned machine.

The different hyperparameters used for running different experiemnts are detailed in the paper.

Licence

Our code is under the MIT License.

Citation

If you use this code for your research, please cite our paper:

@article{kleiner2024coherence,
    title={Coherence Awareness in Diffractive Neural Networks},
    author={Kleiner, Matan and Michaeli, Lior and Michaeli, Tomer},
    journal={arXiv preprint arXiv:2408.06681},
    year={2024}
}