Models Reference

DiffBench supports both classical machine learning models and deep learning models. The available model is selected through the experiment configuration.

For an overview of how models are trained within the benchmark, see Model Training.

Classical machine learning models

Model	Description
linear	Linear or logistic regression, depending on the prediction task
pca_linear	PCA followed by a linear model
lasso	Lasso regression
forest	Random forest
pca_forest	PCA followed by a random forest
svm	Support vector machine
pca_svm	PCA followed by a support vector machine
dummy_classifier	Majority-class baseline for classification
dummy_regressor	Mean-prediction baseline for regression

Deep learning models

Model	Description
dinov2	DINOv2 Vision Transformer backbone
curia	CURIA backbone
medicalnet	MedicalNet 3D backbone

Model selection

The model is selected through the corresponding Hydra configuration.

For example, an experiment may select a model using:

model=linear

or:

model=medicalnet

The exact configuration structure depends on the model implementation.

For training-related parameters, see the Configuration Reference.

For details about extending DiffBench with a new model, see Adding a New Model.