incerto.ood.Mahalanobis

class incerto.ood.Mahalanobis(model, layer_name='penultimate')

Bases: OODDetector

Feature-space Mahalanobis (Lee et al., NeurIPS 2018).

__init__(model, layer_name='penultimate')

Initialize the OOD detector with a trained model.

The model is automatically: 1. Set to eval mode 2. Has gradients disabled (requires_grad=False)

Parameters:

model – A trained PyTorch model (nn.Module)

Methods

__init__(model[, layer_name])	Initialize the OOD detector with a trained model.
fit(loader)
load_state_dict(state)	Load fitted Mahalanobis parameters.
predict(x, threshold)	Predict whether inputs are OOD using a threshold.
save(path)	Save detector state to a file (excluding the model).
score(x)	Compute OOD scores for input samples.
state_dict()	Save fitted Mahalanobis parameters.

__init__(model, layer_name='penultimate')

Initialize the OOD detector with a trained model.

The model is automatically: 1. Set to eval mode 2. Has gradients disabled (requires_grad=False)

Parameters:

model – A trained PyTorch model (nn.Module)

fit(loader)

score(x)

Compute OOD scores for input samples.

Higher scores indicate the input is more likely to be out-of-distribution.

Parameters:

x – Input tensor of shape (batch_size, *input_dims)

Returns:

OOD scores of shape (batch_size,) where higher values indicate more OOD-like samples.

Note

The scale of scores depends on the detection method. Use the predict() method with a threshold for binary OOD decisions.

state_dict()

Save fitted Mahalanobis parameters.

Return type:

dict

load_state_dict(state)

Load fitted Mahalanobis parameters.

Parameters:

state (dict)

Return type:

None