What's new - Support configurable number of stacked gated residuals in `SkipConnectionHead`, useful for training deeper `EncoderHead`s.
0.1.18
What's new - Fix a bug that causes loss of custom `encoder_selection` implementation when restoring `SwitchEncoder` 45
0.1.17
What's new? - Introduce `Encoder.extract_meta()` hook to be able to pass meta data from `Encoder` to `EncoderHead`. - Introduce `SwitchHead` similar to `SwitchEncoder`, which can be used in two-tower models like CLIP.