surfaces-surrogates 0.0.1

Pre-trained ONNX surrogate models for the surfaces library

surfaces-surrogates

Pre-trained ONNX surrogate models for the surfaces library.

Surrogate models approximate computationally expensive ML hyperparameter optimization test functions. Instead of running real cross-validation (seconds to minutes per evaluation), the surrogate returns a prediction in under a millisecond via ONNX inference.

Installation

pip install surfaces-surrogates

For use with the surfaces library, install the surrogates extra which includes onnxruntime:

pip install surfaces[surrogates]

Usage

Surrogates are used transparently through the surfaces library:

from surfaces.test_functions.machine_learning import KNeighborsClassifierFunction

func = KNeighborsClassifierFunction(use_surrogate=True)
score = func({"n_neighbors": 5, "algorithm": "auto"})

The surrogate models can also be loaded directly:

from surfaces_surrogates.models import get_model_path

path = get_model_path("k_neighbors_classifier")

Available Models

Model	R2	Fidelity-Aware
decision_tree_classifier	0.995	Yes
k_neighbors_regressor	0.988	Yes
k_neighbors_classifier	0.966	Yes
gradient_boosting_regressor	n/a	No
svm_regressor	0.351	Yes

Training New Models

To retrain or extend surrogate models, install the training dependencies:

pip install surfaces-surrogates[train]

Then use the training script:

python scripts/train_surrogates.py --status    # show current state
python scripts/train_surrogates.py             # train new/stale models
python scripts/train_surrogates.py --force     # retrain everything

License

MIT