Adding a New Model#

DiffBench can be extended with additional machine learning or deep learning models.

This tutorial describes the main steps required to integrate a new model into the benchmark.

Before starting, it is useful to understand the standard Model Training workflow.

Model interface#

A model integrated into DiffBench must expose the interface expected by the experiment runner.

At minimum, the model must support:

fit(...)
predict(...)

The exact implementation depends on whether the model uses a scikit-learn-style estimator or a PyTorch nn.Module.

Add the model implementation to the appropriate module under the DiffBench model package.

The model should follow the conventions of the existing implementations for the corresponding backend.

Add the model to the model registry so that it can be selected through the experiment configuration.

This allows the experiment runner to instantiate it from its configured name.

Create or update the corresponding Hydra configuration with the parameters required by the new model.

Model-specific parameters should remain in the model configuration rather than being hard-coded in the training pipeline.

Once registered, the model can be selected in the same way as any existing model.

For example:

diffbenchmark-run model=my_new_model

The model will then participate in the standard cross-validation, prediction, and evaluation workflow.

See the Models Reference for examples of existing models and the Models API for details about the expected model interface.