Neural Architecture Search Without Training

This repository contains code for replicating our paper on NAS without training.

Setup

1. Download the datasets.
2. Download NAS-Bench-201.
3. Install the requirements in a conda environment with conda env create -f environment.yml.

We also refer the reader to instructions in the official NASBench-201 README.

Reproducing our results

To reproduce our results:

conda activate nas-wot
./reproduce.sh 3 # average accuracy over 3 runs
./reproduce.sh 500 # average accuracy over 500 runs (this will take longer)

Each command will finish by calling process_results.py, which will print a table. ./reproduce.sh 3 should print the following table:

Method	Search time (s)	CIFAR-10 (val)	CIFAR-10 (test)	CIFAR-100 (val)	CIFAR-100 (test)	ImageNet16-120 (val)	ImageNet16-120 (test)
Ours (N=10)	1.73435	89.25 +- 0.08	92.21 +- 0.11	68.53 +- 0.17	68.40 +- 0.14	40.42 +- 1.15	40.66 +- 0.97
Ours (N=100)	17.4139	89.18 +- 0.29	91.76 +- 1.28	67.17 +- 2.79	67.27 +- 2.68	40.84 +- 5.36	41.33 +- 5.74

./reproduce 500 will produce the following table (which is the same as what we report in the paper):

Method	Search time (s)	CIFAR-10 (val)	CIFAR-10 (test)	CIFAR-100 (val)	CIFAR-100 (test)	ImageNet16-120 (val)	ImageNet16-120 (test)
Ours (N=10)	1.73435	88.47 +- 1.33	91.53 +- 1.62	66.49 +- 3.08	66.63 +- 3.14	38.33 +- 4.98	38.33 +- 5.22
Ours (N=100)	17.4139	88.45 +- 1.46	91.61 +- 1.71	66.42 +- 3.27	66.56 +- 3.28	36.56 +- 6.70	36.37 +- 6.97

To try different sample sizes, simply change the --n_samples argument in the call to search.py, and update the list of sample sizes on line 51 of process_results.py.

Note that search times may vary from the reported result owing to hardware setup.

The code is licensed under the MIT licence.

Acknowledgements

This repository makes liberal use of code from the AutoDL library. We also rely on NAS-Bench-201.