- New ConvTranspose layer @tejank10
- Datasets are now hash verified on download to avoid corruption. @KristofferC
- NamedTuples now work like Tuples when doing mapleaves @Keno
- Added “Performance Tips” section to the documentation thanks to @oxinabox
- The training loop has been revamped and better shows the lower level APIs @oxinabox @MikeInnes
- Support for CuPtr for updated APIs for use in the updated GPU stack @maleadt
- The Iris dataset can now be accessed as Data.Iris thanks to @JoshuaWhittemore
- The major ongoing project is to incorporate Zygote.jl as the AD. The work can also be tracked on Flux’s zygote branch.
Tracker has now been moved into its own package as a step towards making Flux compatible with a larger number of AD libraries.
Please try out this release and file issues and regressions as necessary.
Please reach out if you wish to participate in the Google Summer of Code. The proposed projects list can be found here. But do feel free to come up with your own!
A huge thanks to everyone’s contributions with issues and PRs.