Interface

The interface of all of the models implemented in Bensemble is (almost) the same to provide interchangeability. In this section we will explore the common methods used in all of Bensemble's classes.

Initialization (__init__)

All of the methods' constructors take different input arguments, however, one thing that they have in common is that they all require a nn.Module model as their first argument. This is because Bensemble models are in fact wrappers around PyTorch models allowing for posterior approximation rather than independent models themselves.

fit

All methods have a fit method that trains the model as well as calculates the parameters needed for model sampling and uncertainty estimation. The fit method always requires the train_loader parameter that is a torch.utils.data.DataLoader instance. You can additionally optionally provide a val_loader of the same type for validation.

Keep in mind that if you use LaplaceApproximation, which is a post hoc method primarily used to approximate posterior weight distributions for pretrained models, the fit method will by default just call the compute_posterior method without model training (more on that in the section about Laplace approximation).

predict

The predict method is use to make predictions for test objects with uncertainty estimation by sampling a given amount of model weights and making the corresponding amount of predctions. The parameters of predict in all models are:

• X: torch.Tensor: the object for which you are making a prediction. Keep in mind that it must be a torch.Tensor instance of the same shape as training objects.

• n_samples: int = 100: the number of weight samples used to make the prediction. The default value of this parameter may vary across different methods depending on sampling speeds.

The predict method outputs a tuple of torch.Tensor object containing the mean predictions and their uncertainties.

sample_models

This is the go-to method for model ensemble sampling. It takes the n_models parameter as input, defining the number of models to be sampled, and returns a list of sampled models as nn.Module instances. The sampled models are in fact just copies of the original model with perturbed weights.

save

This method takes a path file name as input and saves the ensemble state to the specified file.

load

This method takes a path file name as input and loads the ensemble state for the specified file.