Implementation of a straight-through gradient wrapper to allow for discrete latent representations. Provides binary discretizer which maps hidden representations to {0, 1} and a learnable multi-value discretizer, which maps hidden activations to their closest value in a set of given size.
Implementation of a straight-through gradient wrapper to allow for discrete latent representations. Provides binary discretizer which maps hidden representations to {0, 1} and a learnable multi-value discretizer, which maps hidden activations to their closest value in a set of given size.
Implementation of a straight-through gradient wrapper to allow for discrete latent representations. Provides binary discretizer which maps hidden representations to {0, 1} and a learnable multi-value discretizer, which maps hidden activations to their closest value in a set of given size.