# Blackbox Repository

This dataset contains hyperparameter optimization (HPO) evaluations from several paper:

- fcnet: Tabular benchmarks for joint architecture and hyperparameter optimization. Klein, A. and Hutter, F. 2019.
- icml-deepar, icml-xgboost: A quantile-based approach for hyperparameter transfer learning. Salinas, D., Shen, H., and Perrone, V. 2021.
- lcbench: Auto-PyTorch: Multi-Fidelity MetaLearning for Efficient and Robust AutoDL. Lucas Zimmer, Marius Lindauer, Frank Hutter. 2020.
- nasbench201: NAS-Bench-201: Extending the scope of reproducible neural architecture search. Dong, X. and Yang, Y. 2020.
- pd1: Pre-trained Gaussian processes for Bayesian optimization. Wang, Z. and Dahl G. and Swersky K. and Lee C. and Mariet Z. and Nado Z. and Gilmer J. and Snoek J. and Ghahramani Z. 2021.
- yahpo: YAHPO Gym - An Efficient Multi-Objective Multi-Fidelity Benchmark for Hyperparameter Optimization. Pfisterer F., Schneider S., Moosbauer J., Binder M., Bischl B., 2022

The evaluations can be accessed through [Syne Tune](https://github.com/syne-tune/syne-tune) HPO library by calling the following:

```python
from syne_tune.blackbox_repository import load_blackbox

blackbox = load_blackbox("nasbench201")["cifar10"]
blackbox_hyperparameter = next(iter(blackbox.hyperparameters.to_dict(orient="records")))
print(f"First hyperparameter: {blackbox_hyperparameter}")
print(
    f"Objectives for first hyperparameters: {blackbox(configuration=blackbox_hyperparameter, fidelity=100)}"
)

# > First hyperparameter: {'hp_x0': 'avg_pool_3x3', 'hp_x1': 'nor_conv_1x1', 'hp_x2': 'skip_connect', 'hp_x3': 'nor_conv_1x1', 'hp_x4': 'skip_connect', 'hp_x5': 'skip_connect'}
# > Objective for first hyperparameters: {'metric_valid_error': 0.4177, 'metric_train_error': 0.2246, 'metric_runtime': 15.461778, 'metric_elapsed_time': 1546.179, 'metric_latency': 0.013935976, 'metric_flops': 15.64737, 'metric_params': 0.129306}
```

In addition, the blackboxes can be used to simulate HPO methods such as ASHA or Bayesian Optimization very fast while keeping identical results with non-simulated tuning.

The files can also be accessed directly from [here](https://huggingface.co/datasets/synetune/blackbox-repository/tree/main).

If you are interested in having other blackboxes feel free to create an issue on Syne Tune project, we aim to grow the set over time.