v0.8.7

NNlib v0.8.7

Diff since v0.8.6

Closed issues:

• v0.8.6 contains breaking changes (#412)

Merged pull requests:

• partly revert changes to stay compatible w NNlibCUDA (#414) (@maxfreu)
• Update CompatHelper.yml (#419) (@CarloLucibello)
• CompatHelper: bump compat for Compat to 4, (keep existing compat) (#420) (@github-actions[bot])

v0.8.6

NNlib v0.8.6

Diff since v0.8.5

Merged pull requests:

• boil down upsampling code via dynamic dispatch (#410) (@maxfreu)

v0.8.5

NNlib v0.8.5

Diff since v0.8.4

Closed issues:

• Installing on Mac M1 (#403)
• NaN caused by tanh_fast (#407)

Merged pull requests:

• Fix tanh_fast for large Float64 (#408) (@mcabbott)

v0.8.4

NNlib v0.8.4

Diff since v0.8.3

Closed issues:

• Bug report: leakyrelu'.(CuArray(rand(Float32, 10))) fails [email protected] (@0.8.2 works) (#398)
• NNlib 0.8.3 broke NNlibCUDA (#400)

Merged pull requests:

• Avoid Rational in activation function gradients (#399) (@mcabbott)

v0.8.3

NNlib v0.8.3

Diff since v0.8.2

Closed issues:

• Nested AD not defined for conv; any plan? (#119)
• Latest release (0.7.9) broke softmax gradients with Tracker (#251)
• ForwardDiff and Conv: Slow fallback implementation invoked for conv! (#349)
• rrule(::typeof(∇conv_filter) (#362)
• move the exports to the main source file (#394)

Merged pull requests:

• Silence some warnings (#383) (@mcabbott)
• Fix gradient of convolution for complex values (#389) (@zsoerenm)
• Conjugate weight before multiplying it with input (#390) (@zsoerenm)
• Improve some activation function gradients (#392) (@mcabbott)
• Simplify softmax, test second derivatives (#393) (@mcabbott)
• Change libblas to libblastrampoline (#396) (@theabhirath)
• Move exports to main source file (#397) (@vincentmolin)

v0.8.2

NNlib v0.8.2

Diff since v0.8.1

Merged pull requests:

• add rrule(::typeof(∇conv_filter) (#387) (@vincentmolin)

v0.8.1

NNlib v0.8.1

Diff since v0.8.0

Merged pull requests:

• Fix CI badge (#382) (@darsnack)
• trigger op specialization in scatter (#384) (@CarloLucibello)

v0.8.0

NNlib v0.8.0

Diff since v0.7.34

Closed issues:

• Add a friendly warning to softmax (#373)
• Reverse CI (#374)
• Fix non test related CI error (#377)

Merged pull requests:

• Fix convolution & pooling type-stability (#370) (@pxl-th)
• Informative error message for softmax variants (#378) (@theabhirath)
• Make ACTIVATIONS vector constant. (#379) (@eliascarv)
• Add downstream reverse CI (#380) (@ToucheSir)
• Add codecov token secret (#381) (@ToucheSir)

v0.7.34

NNlib v0.7.34

Diff since v0.7.33

Merged pull requests:

• Added hard swish activation function (#376) (@theabhirath)

v0.7.33

NNlib v0.7.33

Diff since v0.7.32

Merged pull requests:

• Widen tanh_fast signature (#375) (@mcabbott)