v0.7.24

NNlib v0.7.24

Diff since v0.7.23

Merged pull requests:

• gather shouldn't fallback to getindex (#332) (@CarloLucibello)

v0.7.23

NNlib v0.7.23

Diff since v0.7.22

Merged pull requests:

• Support reverse_indices for integer, tuple, CartesianIndex index (#327) (@yuehhua)
• Fix the output type of reverse_indices (#328) (@yuehhua)
• Add links to readme (#330) (@mcabbott)

v0.7.22

NNlib v0.7.22

Diff since v0.7.21

Closed issues:

• Conv missing negative (#208)
• ERROR: LoadError: could not open file /home/sarthak/.juliapro/JuliaPro_v1.4.2-1/packages/NNlib/FAI3o/deps/deps.jl (#209)
• Conv with batch dim adds channels (#217)
• add adjoint for softplus (#254)
• ∇softmax modifies input argument (#264)
• An error occurred while using the softmax,CUDNNError: CUDNN_STATUS_BAD_PARAM (code 3) (#285)

Merged pull requests:

• revisit softplus (#324) (@CarloLucibello)
• cleanup gradtest (#325) (@CarloLucibello)
• Make reverse_indices gpu compatible (#326) (@yuehhua)

v0.7.21

NNlib v0.7.21

Diff since v0.7.20

Closed issues:

• batched_mul docstring example fails (#322)

Merged pull requests:

• Update ChainRulesCore (#323) (@willtebbutt)

v0.7.20

NNlib v0.7.20

Diff since v0.7.19

Closed issues:

• add rrules for max/min scatter (#317)

Merged pull requests:

• Add rrule for scatter (#297) (@yuehhua)
• introduce trilinear upsampling (#315) (@maxfreu)
• Add NNlibCUDA to test deps (#316) (@DhairyaLGandhi)
• Add gradient for gather (#318) (@yuehhua)
• scatter cleanup and fix min/max (#319) (@CarloLucibello)

v0.7.19

NNlib v0.7.19

Diff since v0.7.18

Closed issues:

• register NNlibCUDA (#300)

Merged pull requests:

• scatter and gather support element type of idx to be CartesianIndex (#308) (@yuehhua)
• [nnlibcuda] updates for CUDA 3.0 (#311) (@CarloLucibello)
• NNlibCUDA cleanup (#312) (@CarloLucibello)
• Avoiding scalar indexing in cuda (#313) (@yuehhua)
• Move NNlibCUDA (#314) (@DhairyaLGandhi)

v0.7.18

NNlib v0.7.18

Diff since v0.7.17

Closed issues:

• Split off the CPU implementation (#224)
• possible huge performance regression in batched_mul (#282)
• pad_constant fails for vectors (#294)
• scatter!() on CuArrays now triggers scalar operations (#298)
• ConvTranspose with negative padding fails on GPU (#307)

Merged pull requests:

• Fix #282 (#299) (@mcabbott)
• Add a LICENSE.md file for the NNlibCUDA.jl package (#301) (@DilumAluthge)
• Adapt rules for batched_* wrappers (#302) (@aussetg)
• CompatHelper: add new compat entry for "Adapt" at version "3.2" (#305) (@github-actions[bot])
• [nnlibcuda] introduce a manifest and remove cufunc (#306) (@CarloLucibello)
• unexport gather/scatter (#309) (@CarloLucibello)

v0.7.17

NNlib v0.7.17

Diff since v0.7.16

Closed issues:

• batched_mul(A, B) not differentiable when B is an Array{T, 2} (#290)
• rrule for pad_constant fails when using (l1, r1, ..., ln, rn) padding (#293)

Merged pull requests:

• add gather (#280) (@CarloLucibello)
• create NNlibCUDA sub-package (#286) (@CarloLucibello)
• Fix batched_mul gradient, for cases with one matrix (#291) (@mcabbott)
• Fixes #293 and #294 (#295) (@torfjelde)

v0.7.16

NNlib v0.7.16

Diff since v0.7.15

Closed issues:

• issues with pad_constant (#287)

Merged pull requests:

• Add some user forwards (#288) (@DhairyaLGandhi)

v0.7.15

NNlib v0.7.15

Diff since v0.7.14

Merged pull requests:

• Add scatter operations (#255) (@yuehhua)
• Upsample docstring improvements (#278) (@mcabbott)
• add pad_constant, pad_repeat, pad_reflect, pad_zeros (#279) (@CarloLucibello)
• cleanup AD (#281) (@CarloLucibello)
• Faster constant padding (#283) (@DhairyaLGandhi)
• update ci.yml (#284) (@CarloLucibello)