LOMONAS: Local-search algorithm for Multi-Objective Neural Architecture Search

Quan Minh Phan, Ngoc Hoang Luong.

In GECCO 2023.

Setup

• Clone repo.
• Install necessary packages.

$ pip install -r requirements.txt

• Download databases in this link, unzip and put all folders into data folder for building APIs (benchmarks).

In our experiments, we do not implement directly the API benchmarks published in their repos (e.g., NAS-Bench-101, NAS-Bench-201, etc). Instead, we create smaller-size databases by accessing their databases and only logging necessary content.

You can compare our databases and the original databases in check_log_database.ipynb

Reproducing the results

You can reproduce our results by running the below script:

$ python main.py --optimizer [MOEA_NSGAII, MOEA_MOEAD, RR_LS, LOMONAS] --problem [NAS101, MacroNAS-C10, MacroNAS-C100, NAS201-C10, NAS201-C100, NAS201-IN16, NAS-ASR]

Search with different hyperparameters

Moreover, you can search with different hyperparameter settings

For environment

• n_run: the number times of running algorithms (default: 31)
• max_eval: the maximum number of evaluation each run (default: 3000)
• init_seed: the initial random seed (default: 0)
• api_benchmark_path: the path that contains the api databases (default: ./data)
• res_path: the path for logging results (default: ./exp_res)
• debug: print the search performance at each generation if debug is True (default: False)

For evolution-based MONAS (NSGA-II & MOEA/D)

• pop_size: the population size

For Random Restart - Local Search

• loop: implement the loop RR-LS variant if loop is True (default: False)

For LOMONAS

• NF: the number of fronts kept for performing neighborhood checks (default: 3)
• check_all_neighbors: evaluate all solutions in the neighborhood if set True (default: False)
• neighborhood_check_on_all_sols: perform neighborhood check on all solutions instead of knee and extremes ones if set True (default: False)

The results of LOMONAS in the paper are run with the default hyperparameters.

Transferability Evaluation

In our study, we evaluate the transferability of algorithms by evaluating the found architectures (search on CIFAR-10) on CIFAR-100 (for MacroNAS and NAS-201) and ImageNet16-120 (for NAS-201 only).

$ python transferability_evaluation.py
--problem [MacroNAS-C100, NAS201-C100, NAS201-IN16]
--api_benchmark_path <path_of_benchmark_databases, default: ./data>
--cifar10_res_path

Note: --cifar10_res_path must contain the searching result at MacroNAS-C10 or NAS201-C10 problems.

For example:

$ python transferability_evaluation.py --problem NAS201-C100 --cifar10_res_path ./exp_res/NAS201-C10 --api_benchmark_path ./data

Visualization and T-test

You can create all figures presented in our paper by run the visualization_and_statistical_test.ipynb file.

Acknowledgement

Our source code is inspired by:

• pymoo: Multi-objective Optimization in Python
• NSGA-Net: Neural Architecture Search using Multi-Objective Genetic Algorithm
• NAS-Bench-101: Towards Reproducible Neural Architecture Search
• NAS-Bench-201: Extending the Scope of Reproducible Neural Architecture Search
• NAS-Bench-ASR: Reproducible Neural Architecture Search for Speech Recognition
• Local Search is a Remarkably Strong Baseline for Neural Architecture Search