Improved FireFly Algorithm

This repository contains the implementation of the improved Firefly algorithm, using a training-free evaluation of the model quality, in order to perform Neural Architecture Search on NAS-BENCH-101 and NAS-BENCH-201. This method can be used to build automatically a convolutional neural network for any image classification problem.

Our paper can be found at:

Improving Neural Architecture Search by Mixing a FireFly algorithm with a Training Free Evaluation

If you use or build on our work, please consider citing us:

@INPROCEEDINGS{9892861,
  author={Mokhtari, Nassim and Nédélec, Alexis and Gilles, Marlène and De Loor, Pierre},
  booktitle={2022 International Joint Conference on Neural Networks (IJCNN)}, 
  title={Improving Neural Architecture Search by Mixing a FireFly algorithm with a Training Free Evaluation}, 
  year={2022},
  pages={1-8},
  doi={10.1109/IJCNN55064.2022.9892861}}

Setup

In order to be able to use our implementation, please follow these instructions :

I) Nas-Bench-101 setup

1 download "nasbench_only108.tfrecord" from https://storage.googleapis.com/nasbench/nasbench_only108.tfrecord and place it into "api" folder

2 Install nasbench :

git clone https://github.com/google-research/nasbench
cd nasbench
pip install -e .

• note: If you are using tensorflow 2, you may have a problem when running nasbench, please refer to this solution google-research/nasbench#27 (comment)

3 - Install nasbench_keras :

git clone https://github.com/lienching/nasbench_keras
cd nasbench_keras
pip install -e .

4 - Install xautodl :

pip install xautodl

5 - Update numpy if necessery.

II) Nas-Bench-201 setup

1 - download "NAS-Bench-201-v1_1-096897.pth" from https://drive.google.com/open?id=16Y0UwGisiouVRxW-W5hEtbxmcHw_0hF_ and place it into "api" folder.

2 - Install nasbench201 :

pip install nas-bench-201 

Usage

You can start the Neural Architecture Search using the default parameters by running the main.py from the command line :

python ./main.py

• Samples from CIFAR-10 are provided in data/samples.pt

You can run on your own dataset (saved as pytorch tensor) by using:

python ./main.py --data_path <PATH_TO_YOUR_DATASET> --input_shape <H> <W> <C> --data_format channels_last --num_labels <#Classes>

• The input shape must be provided in the form of a set of integers separated by a blank, default is 3 32 32.
• the data format must matches the input shape, default is channels_first.

Several parameters can be used to refine the search for a neural architecture. You can find more details about these parameters using :

python ./main --help