Experimental

This module provides experimental features. Use cautiously.

class experimental.TorchPredictor(model, is_trained=False, optimizer=None, criterion=None, **compile_kwargs)

Bases: object

Wrapper of a torch prediction model \(\hat{f}\). Enables to standardize the interface of torch predictors and to expose generic fit(), predict() and copy() methods.

Parameters:

• model (Any) – torch prediction model \(\hat{f}\)

• is_trained (bool) – boolean flag that informs if the model is pre-trained. If True, the call to fit() will be skipped

• optimizer (torch.optim.Optimizer) – torch optimizer. Defaults to torch.optim.Adam.

• criterion (torch.nn.modules.loss) – criterion that measures the distance between predictions and outputs. Default to torch.nn.MSELoss.

• compile_kwargs – keyword arguments passed to the optimizer constructor when creating the optimizer in fit().

copy()

Returns a copy of the predictor.

Returns:

copy of the predictor.

Return type:

TorchPredictor

fit(X, y, **kwargs)

Fit model to the training data.

Parameters:

• X (torch.Tensor) – train features.

• y (torch.Tensor) – train labels.

• kwargs – keyword arguments to configure the training.

predict(X)

Compute predictions on new examples.

Parameters:

X (torch.Tensor) – new examples’ features.

Returns:

predictions \(\hat{f}(X)\) associated to the new examples X.

Return type:

torch.Tensor