/jahs_bench_201_experiments

Experiments for the JAHS-Bench-201 paper.

Primary LanguagePythonOtherNOASSERTION

JAHS-Bench-201 Experiments

Installation

Clone this repository:

git clone https://github.com/automl/jahs_bench_201_experiments.git
cd jahs_bench_201_experiments

Install JAHS-Bench-201 using pip:

pip install git+https://github.com/automl/jahs_bench_201.git

Install this repository using pip:

pip install .

Download the surrogates:

python -m jahs_bench.download --target surrogates

Experiments

First research question

To reproduce results for JAHS, run:

python jahs_bench_201_experiments/tasks/run_model_free.py --dataset DATASET --seed SEED --fidelity None

where DATASET may be one of cifar10, colorectal_histology, fashion_mnist.

Append --use_default_hps or --use_default_arch for NAS-only or HPO-only, respectively.

Second research question

To reproduce results, run:

python jahs_bench_201_experiments/tasks/run_model_free.py --dataset DATASET --seed SEED --fidelity FIDELITY --min_budget MIN_BUDGET --max_budget MAX_BUDGET

where FIDELITY may be one of Epochs, N, W, Resolution and DATASET may be one of cifar10, colorectal_histology, fashion_mnist.

Leaderboard

We provide a script for SLURM-job submission for all leaderboard entries, before running:

bash jahs_bench_201_experiments/tasks/run_all.sh

NOTE: please adjust the user-specific paths in startup.sh and run_all.sh

Analysis

To reproduce plots from Section 4:

python jahs_bench_201_experiments/analysis/analysis.py

To reproduce Table 4 and 5:

python jahs_bench_201_experiments/analysis/compare_runtime.py

NOTE: this command requires all data.

To generate leaderboard:

python jahs_bench_201_experiments/analysis/leaderboard.py

NOTE: this command requires all data.

Repository structure:

jahs_bench_201_experiments          # This repository
├── jahs_bench_201_data             # To be downloaded
├── jahs_bench_201_plots            # Created upon regenerating plots
├── jahs_bench_201_results          # Created upon running any task
├── jahs_bench_201_experiments
│   ├── analysis                    # Scipts for results analysis
│   ├── tasks                       # Scipts for results geneartion
│   │   │── ...
│   │   │── run_all.sh              # Script submitting SLURM array job
│   │   │── run_model_free.py       # Main task script (model-free)
│   │   │── run_model_based.py      # Main task script (model-based)
│   ├── utils