End-to-end Framework for Metasurface-based Optical Deep Learning

This repository contains the files that are necessary to run the pipeline of out hybrid CNNs on different datasets using different architectures. Each folder corresponds to a different (architecture,dataset) pair. In order to run the pipeline in the ideal case (ignoring the effects of cross-talk between channels), run the hybrid_onn_pipeline_ideal.py file of the corresponding folder on a terminal. In order to run the pipeline in the non-ideal case (incorporating the effects of cross-talk between channels), run the hybrid_onn_pipeline_nonideal.py file of the corresponding folder on a terminal.

For these pipelines to work, some files (which could not be uploaded to Github because of their size) need to be added:

• The CIFAR-10 python files downloaded from https://www.cs.toronto.edu/~kriz/cifar.html need to be added to HybridCNN_CodeRepo/AlexNet_CIFAR10/cifar-10-batches-py/
• HybridCNN_CodeRepo/VGG_CIFAR100/weights/ needs to have a file called cifar100vgg_pretrained.h5, which can be downloaded from
• HybridCNN_CodeRepo/ResNet50_CIFAR100/weights/Final/ needs to have a file called model_finetuned_2020-09-01-07-46.h5, which can be downloaded from

After having added these files to the indicated directories, install Anaconda (if it is not installed already) following the instructions in: https://docs.anaconda.com/anaconda/install/

After that, move to the cloned repository's directory and run this command to install the required packages:

conda create --name tensorflowenv --file spec-file.txt

Then move to the folder that corresponds to the (architecture,dataset) pair you wish to run, and run these commands if you want to run the ideal case (no cross-talk):

conda activate tensorflowenv
python hybrid_onn_pipeline_ideal.py

If instead you want to run the non-ideal case (with cross-talk), run these commands from the folder that corresponds to the the (architecture,dataset) pair you wish to run:

conda activate tensorflowenv
python hybrid_onn_pipeline_nonideal.py

Publication

This project contains the artifact for our paper End-to-End Framework for Efficient Deep Learning Using Metasurfaces Optics.