torch¶

torch layer implementations.

Classes¶

`BatchEnsembleConv2d`	BatchEnsemble convolutional layer based on [WTB20].
`BatchEnsembleLinear`	BatchEnsemble linear layer based on [WTB20].
`BayesConv2d`	Implementation of a Bayesian convolutional layer based on [BCKW15].
`BayesLinear`	Implement a Bayesian linear layer based on [BCKW15].
`DropConnectLinear`	Custom Linear layer with DropConnect applied to weights during training.
`GaussianMixtureHead`	Per-class Gaussian density head (Gaussian Discriminant Analysis).
`HeteroscedasticLayer`	A unified PyTorch implementation of the Heteroscedastic layer.
`IRDHead`	Head that converts encoded features into Dirichlet concentration parameters (alpha).
`IntBatchNorm1d`	Interval-valued batch normalization for 1D features based on [WCM+24].
`IntBatchNorm2d`	Interval-valued batch normalization for 2D feature maps based on [WCM+24].
`IntConv2d`	Interval-arithmetic 2D convolution based on [WCM+24].
`IntLinear`	Interval-arithmetic linear layer based on [WCM+24].
`IntSoftmax`	Interval SoftMax head based on [WCM+24].
`MahalanobisHead`	Class-conditional Gaussian head with a tied covariance (Mahalanobis OOD).
`NatPNClassHead`	Dirichlet posterior head for evidential classification.
`NatPNRegHead`	Gaussian posterior head for evidential regression.
`NormalInverseGammaLinear`	Custom Linear layer for a normal-inverse-gamma-distribution based on [ASSR20].
`RadialNormalizingFlow`	Radial normalizing flow based on [RM15].
`RadialNormalizingFlowStack`	Stack of radial normalizing flows based on [RM15].
`RegressionHead`	Head that converts encoded features into evidential Normal-Gamma parameters.
`SNCoeffParametrization`	Weight parametrization that clips the spectral norm to at most `coeff`.
`SNGPLayer`	Spectral-normalized Neural Gaussian Process (SNGP) layer based on [LLP+20].
`SharedMaskDropout`	Dropout that draws a single mask per forward pass, shared across the batch.

Functions¶

`apply_spectral_norm_to_encoder`	Apply spectral normalization in-place to all Conv2d and Linear layers.
`pack_interval`	Promote `x` to a packed interval tensor with `lo == hi == x`.
`unpack_interval`	Split a packed interval tensor on `channel_dim` into its `(lo, hi)` halves.