Generic Neural Architecture Search via Regression

This repository is the official implementation of Generic Neural Architecture Search via Regression.

Requirements

PyTorch >= 1.1.0 (We suggest to use version 1.5.0. It may work for version < 1.1.0 but we didn't test.)

Minimum required datasets:

Download the data for NDS dataset

mv <path_to_data> <path_to_GenNAS>/data/

Download the data for NASBench-NLP

git clone https://github.com/fmsnew/nas-bench-nlp-release.git
mv ./nas-bench-nlp-release/train_logs_single_run <path_to_GenNAS>/data/
mv ./nas-bench-nlp-release/train_logs_wikitext-2 <path_to_GenNAS>/data/

Download the data for ImageNet16

mv <path_to_data>/* <path_to_GenNAS>/data/ImageNet16

Suggest datasets & API:

NASBench-101 NASBench-201

Proxy Task Search

To search for a proxy task, run the following examples:

python do_search.py --search_space=nasbench101 --json_loc=data/nasbench1_search_20samples.json --json_description=nasbench1_search_20samples #NASBench-101
python do_search.py --search_space=nlp #NASBench-NLP
python do_search.py --search_space=DARTS #NDS

Sampling Experiments

To do the sampling experiments, run the following examples:

python do_sample.py --search_space=nasbench101 --config=CONF_NB101 --json_loc=data/nasbench1_500_fb.json --json_description=nasbench1_500_fb #NASBench-101
python do_sample.py --search_space=nasbench201 --config=CONF_NB101 --json_loc=data/nasbench2_1000_0.json --json_description=nasbench2_1000_0 --dataset=cifar10 #NASBench-201
python do_sample.py --search_space=nlp --config=CONF_NLP #NASBench-NLP
python do_sample.py --search_space=DARTS --config=CONF_DARTS #NDS

Exploring Experiments

To do the exploring experiments, run the following examples:

python do_explore.py --search_space=nasbench101 --config=CONF_NB101 #NASBench-101
python do_explore.py --search_space=nasbench201 --config=CONF_NB101 #NASBench-201
python do_explore.py --search_space=nlp --config=CONF_NLP #NASBench-NLP
python do_explore.py --search_space=DARTS --config=CONF_DARTS #NDS

lilujunai / GenNAS