A PyTorch class that implements an approximate Gaussian process as the last layer of a neural network - compatible with any architecture and with regression, binary logistic classification and classification. This provides a simple way to obtain uncertainty calibration.
We recommend using this in combination with spectral normalization which is approximately distance-preserving (see Liu et al for details). This ensures that datapoints far from the training set in the input space are appropriately associated with high uncertainty. We may add standard spectral-normalized layers to a future release to make this easier to implement.
For details on installation and usage, see the docs