vbll.torch¶

Collection of torch VBLL training losses.

These functions implement the negative ELBO objectives used to fit the variational Bayesian last layers of [HWS24]. Each takes the layer together with the backbone features feeding it (captured with a forward pre-hook), the integer targets, and the weight on the KL/regularization terms.

Functions¶

`disc_vbll_loss`	Negative discriminative ELBO of a `VBLLLayer` using the double-Jensen bound.
`g_vbll_loss`	Negative generative ELBO (the Jensen bound) of a `GVBLLLayer`.
`het_vbll_loss`	Negative ELBO of a `HetVBLLLayer` using the reduced Knowles-Minka bound.
`t_vbll_loss`	Negative ELBO of a `TVBLLLayer` using the reduced Knowles-Minka bound.