Improving Transferability of Multi-Objective Evolutionary Neural Architecture Search by Utilizing Multiple Datasets in Network Evaluations

Ngoc Hoang Luong, Tu Do

In NICS'21.

Installation

Clone this repo:

git clone https://github.com/MinhTuDo/MD-MOENAS.git
cd MD-MOENAS

Install dependencies:

pip install -r requirements.txt

Usage

0. Prepare the NAS Benchmarks

Follow the instructions here to install benchmark files for NATS-Bench.
Remember to properly set the benchmark paths in config files, default data path is ~/.torch.

1. Search

Topology Search Space

# Baseline MOENAS evaluating only on CIFAR-10 dataset
python search.py --console_log -sw --use_archive --search_space tss -dts cifar10 --efficiency flops --eval_dts ImageNet16-120

# MD-MOENAS evaluating on CIFAR-10 & CIFAR-100 datasets
python search.py--console_log -sw --use_archive --search_space tss -dts cifar10 -dts cifar100 --efficiency flops --eval_dts ImageNet16-120

Size Search Space

# Baseline MOENAS evaluating only on CIFAR-10 dataset
python search.py --console_log -sw --use_archive --search_space sss -dts cifar10 --efficiency params --eval_dts ImageNet16-120

# MD-MOENAS evaluating on CIFAR-10 & CIFAR-100 datasets
python search.py--console_log -sw --use_archive --search_space sss -dts cifar10 -dts cifar100 --efficiency params --eval_dts ImageNet16-120

To evaluate IGD score on pre-computed optimal front during the search, simply provide --eval_igd flag.

You change dataset for IGD evaluation by providing value for --eval_dts. Note that this will work only if --eval_igd flag is used.

To change efficiency objective, simply change --efficiency parameters. Available efficiency objectives are params, flops and latency

For customized search, additional configurations can be modified through yaml config files in config folder.

Acknowledgement

Code inspired from:

TuDo1403/MD-MOENAS