Ngoc Hoang Luong, Tu Do

In ICONIP 2021.

• Clone this repo:

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

• Install dependencies:

pip install -r requirements.txt

• Follow the instructions here to install benchmark files for NAS-Bench-101.
• Follow the instructions here to install benchmark files for NAS-Bench-201.
• Optional: To evaluate IGD on the optimal front during a NAS run, for NAS-Bench-101, you need to download the pre-computed benchmark query data here and put it in the data folder.
• Remember to properly set the benchmark paths in config files, default data path is ~/.torch.

# Baseline MOENAS
python search.py -cfg config/baseline_moenas-101.yml --n_evals 5000 --pop_size 50 --loops_if_rand 30 -sw --use_archive

# Training-free MOENAS
python search.py -cfg config/tf_moenas-101.yml --n_evals 5000 --pop_size 50 --loops_if_rand 30 -sw --use_archive

# Baseline MOENAS
python search.py -cfg config/baseline_moenas-201.yml --n_evals 3000 --pop_size 50 --loops_if_rand 30 -sw --use_archive

# Training-free MOENAS
python search.py -cfg config/tf_moenas-201.yml --n_evals 3000 --pop_size 50 --loops_if_rand 30 -sw --use_archive

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

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

Code inspired from:

• NASBench: A Neural Architecture Search Dataset and Benchmark,
• NASBench-Pytorch,
• NATS-Bench: Benchmarking NAS Algorithms for Architecture Topology and Size,
• Pymoo: Multi-Objective Optimization in Python,
• Neural Architecture Search on ImageNet in Four GPU Hours: A Theoretically Inspired Perspective,
• Automated Deep Learning Projects (AutoDL-Projects)