Skip to content

Latest commit

 

History

History
62 lines (45 loc) · 3.66 KB

README.md

File metadata and controls

62 lines (45 loc) · 3.66 KB

EmProx

Implementation of paper EmProx: Neural Network Performance Estimation For Neural Architecture Search

Adaption of NASLib.

Setup

  1. It is recommended to create a new conda environment.
conda create -n mvenv python=3.8
  1. Install the packages from requirements.txt.
pip install -r requirements.txt
  1. Download the CIFAR-10 dataset here and the file nb201_cifar10_full_training.pickle here and place them in the directory EmProx/naslib/data (optional: CIFAR100 data for NAS-Bench-201, DARTS architectures and CIFAR10 data for DARTS).

Usage

To reproduce the findings in the paper, simply run the following command in the directory EmProx/experiments:

python run_experiments.py

This evaluates the predictors listed below on architectures from NAS-Bench-201 on the CIFAR-10 dataset. Evaluation metrics are MAE and RMSE between the predicted and actual validation accuracy of the architectures. Additionally, since in many NAS algorithms the exact predicted performance is not of importance, rather than the rank of a certain architecture among other candidates, correlation-based performance measures Pearson, Spearman and Kendall’s Tau on the predicted and validation accuracies are included. Lastly, fit (training) times and query times are incorporated.

Results are averaged over 20 trials and outputted in seperate Excel files per predictor.

To reproduce the additional experiments, the search space and dataset can be changed in the experiment_config.yaml file.

Results

Based on our own experiments, results are as follows:

Predictor MAE RMSE Pearson Spearman Kendall Fit time Query time
EmProx (k = 10) 4.4027 10.7163 0.4771 0.7304 0.5453 6.4498 0.0009
EmProx (k = 60) 4.5264 10.7953 0.5044 0.7332 0.5468 7.2310 0.0032
NAO 4.7336 10.9394 0.4512 0.6433 0.4835 54.1674 0.0026
SemiNAS 4.0283 10.1222 0.5307 0.7677 0.5822 152.2268 0.0012
XGB 5.3989 12.2955 0.4008 0.6466 0.4719 31.6526 0.0004
BANANAS 7.3007 12.2760 0.3793 0.4169 0.2910 507.3936 0.0002
MLP 6.8584 11.3592 0.4417 0.5298 0.3759 471.7508 <0.0001

The upper two rows are the results of the method proposed in this study, below are several methods evaluated in White et al. (2021). In bold are the best results of EmProx and the best results among the other predictors for ease of comparison. As can be seen, EmProx scores competitively regarding regression and correlation coefficients and scores much better regarding fit time.