v0.2.0

Tsunami v0.2.0

Diff since v0.1.7

Breaking changes:

• Devices indexing now starts from 1, as in MLDataDevices.gpu_device.

Highlights:

• Updated to Flux v0.16.
• Models (i.e. subtypes of FluxModule) are now not required to be mutable.
• Tsunami.fit is deprecated in favor of Tsunami.fit!.
• Added Trainer option to use Enzyme for automatic differentiation
• Improved test infrastructure.
• Improved documentation.

Merged pull requests:

• follow @functor -> @layer changes in Flux (#62) (@CarloLucibello)
• add VAE example (#64) (@CarloLucibello)
• changes for Flux v0.16 (#73) (@CarloLucibello)
• remove fit in favor of fit! (#76) (@CarloLucibello)
• set default idx = nothing (#77) (@CarloLucibello)
• use TestItems (#78) (@CarloLucibello)
• deprecate bson for jld2 (#79) (@CarloLucibello)
• make tests GPU backend agnostic (#80) (@CarloLucibello)
• implement Enzyme support (#81) (@CarloLucibello)
• fix enzyme pullback (#84) (@CarloLucibello)

Closed issues:

• cannot checkpoint models containing closures (#1)
• replace bson checkpointing with jld2? (#12)
• relax requirement that FluxModule is mutable (#65)
• add Enzyme support (#67)
• remove :metal,:amdgpu, :cuda accelerator options in favor of :gpu (#68)
• use JLD2 for checkpointing (#70)
• make tests GPU backend independent (#71)
• move tests to TestItems.jl (#74)