What's Changed
- Fix mix docs build failure by @aphillipo in #632
- Missed the serialisation guide being added to the build of the docs. by @aphillipo in #633
- Correct implementation of weight tying by @seanmor5 in #625
- Add scale layer by @seanmor5 in #635
- Add a residual layer combinator by @seanmor5 in #637
- fix: fix Axon.Display for Tensor parameter templates by @karlosmid in #638
- chore: update to nx 0.13 by @polvalente in #641
- Annotate activation functions with Nx blocks by @Chapaman in #639
- Add SwiGLU activation by @Chapaman in #642
- Default Axon.scale/2 to a ones initializer by @seanmor5 in #649
- Support grouped transposed convolutions by @seanmor5 in #648
- Support container outputs in eval_step and Axon.Display by @seanmor5 in #647
- Add Axon.deferred/2 by @seanmor5 in #651
- Add Axon.elem/3 and Axon.fetch/3 container accessors by @seanmor5 in #653
- Donate step state buffers during training by @seanmor5 in #645
- Define Axon.ModelState functions as transforms by @seanmor5 in #656
- Fix RNN unrolling gradients and bidirectional LSTMs by @seanmor5 in #646
- Remove loop output transforms by @seanmor5 in #657
- Add Axon.namespace/2 and /3 by @seanmor5 in #652
- Log iteration instead of batch in Axon.Loop progress output by @seanmor5 in #659
- Improve shape mismatch errors in training loops by @seanmor5 in #662
- chore: Update to Nx, EXLA and Torchx 1.0 by @georgeguimaraes in #671
New Contributors
- @karlosmid made their first contribution in #638
- @georgeguimaraes made their first contribution in #671
Full Changelog: v0.8.1...v0.9.0
Contributors
georgeguimaraes, aphillipo, and 4 other contributors