Relative Mahalanobis Distance (RMD)¶

class pytorch_ood.detector.RMD(encoder: Callable[[Tensor], Tensor] | None)[source]¶

Bases: Mahalanobis

classification no segmentation requires fit

Implements the Relative Mahalanobis Distance (RMD) from the paper A Simple Fix to Mahalanobis Distance for Improving Near-OOD Detection.

This method calculates a class center \(\mu_y\) for each class, and a shared covariance matrix \(\Sigma\) from the data.

Additionally, it fits a background gaussian with mean \(\mu_0\) and covariance matrix \(\Sigma_0\) to all of the features and calculates outlier scores as

\[\min_k \lbrace d_k(f(x)) - d_0(f(x)) \rbrace\]

where \(f(x)\) are the features of \(x\), \(d_k(z) = (z - \mu_k)^\top \Sigma^{-1} (z - \mu_k)\) is the squared Mahalanobis distance to class \(k\) and \(d_0\) is the squared Mahalanobis distance under the background gaussian.

Input:: features
Paper:: A Simple Fix to Mahalanobis Distance for Improving Near-OOD Detection, arXiv 2021
Parameters:: encoder – feature extractor that maps inputs to features of shape \(B \times D\). Can be None when using fit_features(...) and predict_features(...) directly.

Attributes

cov: Tensor¶: Covariance Matrix

property device¶: The device of the detector’s owned torch state, if one can be inferred.

mu: Tensor¶: Centers

property n_classes¶: Number of classes the model is fitted for

precision: Tensor¶: Precision Matrix

requires_fit = True¶: Whether fit(...) must be called before scoring.

Methods

fit(data_loader: DataLoader) → Self[source]¶

Fit parameters of the multivariate Gaussians for the given loader. Ignores OOD Inputs.

Parameters:: data_loader – data loader with training data
Returns:: self

fit_features(z: Tensor, y: Tensor) → Self[source]¶

Fit parameters of the multivariate Gaussians. Ignores OOD inputs.

Parameters:

z – features of shape \(N \times D\)
y – class labels of shape \(N\)

Returns:

self

predict(x: Tensor) → Tensor[source]¶

Parameters:: x – input batch, will be passed through the encoder
Returns:: relative Mahalanobis scores of shape \(B\)

predict_features(z: Tensor) → Tensor[source]¶

Calculates the relative Mahalanobis distance directly on features.

Parameters:: z – features of shape \(B \times D\) (or feature maps of shape \(B \times C \times H \times W\), which are spatially averaged)
Returns:: relative Mahalanobis scores of shape \(B\)

to(device: str | device) → Self¶

Move detector-owned modules and tensor state to device.

This is a detector-level analogue of nn.Module.to(...). It moves modules, tensors, and common container-valued state stored on the detector itself.

Parameters:: device – target device, e.g. "cuda:0" or a torch.device
Returns:: self