Models API

This page describes the model interface used internally by DiffBench.

For a conceptual overview of model training, see Model Training.

Model creation

Models are instantiated through the model configuration utilities.

from diff_benchmark.models.model_configurations import get_model
from omegaconf import OmegaConf

cfg = OmegaConf.load(
    "exp_outputs/experiments/exp_myrun/config.yaml"
)

model = get_model(
    cfg.model.name,
    OmegaConf.to_container(cfg, resolve=True),
)

Training interface

Models expose a common training and prediction interface:

model.set_fold(0)

model.fit(train_loader)

predictions = model.predict(test_loader)

fit() trains the model using the data associated with the current fold.

predict() returns the model predictions in a common format so that downstream evaluation does not depend on the underlying modelling framework.

This abstraction allows DiffBench to support both scikit-learn-style estimators and PyTorch-based models through the same experiment orchestration.