v0.7.13

NNlib v0.7.13

Diff since v0.7.12

Closed issues:

• batched_gemm! emits "cannot take the CPU address of a CuArray" (#256)
• Batched_mul adjoint fails (#268)

Merged pull requests:

• improve bilinear upsampling (#266) (@maxfreu)
• Fix stride & size inference of 'T'/'N' in batched_mul (#271) (@mcabbott)
• Gradients for batched_adjoint etc. (#272) (@mcabbott)

v0.7.12

NNlib v0.7.12

Diff since v0.7.11

Merged pull requests:

• Add upsample_nearest (#269) (@mcabbott)
• Partially revert #259 (#270) (@mcabbott)

v0.7.11

NNlib v0.7.11

Diff since v0.7.10

Closed issues:

• ∇softmax modifies input argument (#264)

Merged pull requests:

• add bilinear upsampling (#262) (@CarloLucibello)

v0.7.10

NNlib v0.7.10

Diff since v0.7.9

Closed issues:

• Integration with Batched.jl (#76)
• no method matching conv(::Array{Float32,4}, ::Array{Float32,4}; stride=(1, 1), pad=(0, 0), dilation=(1, 1)) (#106)
• Convolution for mixed-precision inputs (#107)
• Asymmetric padding fails on gpu models (#117)
• ambiguity in calling maxpool (#128)
• move adjoint definitions from Zygote (#219)
• softmax gradients inefficient, API change necessary (#248)
• @eval (#263)

Merged pull requests:

• port rule definitions to ChainRulesCore (#242) (@simeonschaub)
• new API grad softmax (#250) (@CarloLucibello)
• remove ZygoteRules (#257) (@CarloLucibello)
• Remove TODO about not defining broadcasted for ADs other than Zygote (#258) (@oxinabox)
• more extensive tests and some reorg (#259) (@CarloLucibello)
• add pixel_shuffle (#260) (@CarloLucibello)
• fix hardsigmoid and use float(x) instead of x/1 (#261) (@CarloLucibello)

v0.7.9

NNlib v0.7.9

Diff since v0.7.8

Closed issues:

• softmax! and logsoftmax! implementations incomplete (#249)

Merged pull requests:

• added softmax!. reduced softmax's memory usage. (#247) (@norci)

v0.7.8

NNlib v0.7.8

Diff since v0.7.7

Closed issues:

• conv_bias_act! does not modify its output argument correctly (#243)

Merged pull requests:

• add logsumexp (#244) (@CarloLucibello)
• fix conv_act_bias (#245) (@CarloLucibello)
• github CI (#246) (@CarloLucibello)

v0.7.7

NNlib v0.7.7

Diff since v0.7.6

Closed issues:

• pool padding is buggy (#229)
• CUDNNError: CUDNN_STATUS_BAD_PARAM (code 3) while training lstm neural network (#237)

Merged pull requests:

• Allow batched_mul to work through PermutedDimsArray, II (#191) (@mcabbott)
• Add docs link (#240) (@mcabbott)
• Improvements to batched_mul (#241) (@mcabbott)

v0.7.6

NNlib v0.7.6

Diff since v0.7.5

Merged pull requests:

• fix meanpool bug #229 (#238) (@jeremiedb)

v0.7.5

NNlib v0.7.5

Diff since v0.7.4

Closed issues:

• Conv Fallback Warning Hindering Intentional Work (With FixedPointNumbers) (#232)

Merged pull requests:

• Added wrapper for convolution, bias, and activation (#228) (@gartangh)
• Update links and use main page of papers instead of their PDFs (#231) (@hieronimo)
• Fix multithreading bug in ∇conv_filter_im2col (#235) (@pazner)
• Limit logging of fallback @warning 'maxlog=1' (#236) (@RocketRoss)

v0.7.4

NNlib v0.7.4

Diff since v0.7.3

Closed issues:

• Hardware is unsupported by NNPACK so falling back to default NNlib (#225)

Merged pull requests:

• Revert LoopVectorization (#226) (@DhairyaLGandhi)
• requires for nnpack (#227) (@CarloLucibello)
• Bump version (#230) (@DhairyaLGandhi)