Changelog

Changelog#

All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

Breaking: StackedLinearLayer.weight shape changed from (n_stacks, in_features, out_features) to (n_stacks, out_features, in_features), aligning with nn.Linear’s (out_features, in_features) convention and with the bias shape (n_stacks, out_features). The forward pass and output_subset / stack_subset behaviour are unchanged; only the stored weight layout differs.

StackedLinearLayer.forward now accepts an arbitrary number of leading (batch) dimensions, like nn.Linear’s (*, in_features). Input of shape (..., n_stacks, in_features) produces output of shape (..., n_stacks, out_features); the transformation is applied only to the last two dimensions. Previously only a fixed 2D batch (batch, n_stacks, in_features) was supported. output_subset and stack_subset work with the extra leading dimensions as well.