v0.7.2

NNlib v0.7.2

Diff since v0.7.1

Merged pull requests:

• use LoopVectorization to vectorize activation functions and softmax (#199) (@AStupidBear)

v0.7.1

NNlib v0.7.1

Diff since v0.7.0

Merged pull requests:

• Remove redundant loops (#213) (@pxl-th)
• dev comments for conv (#214) (@CarloLucibello)
• remove manifest (#215) (@CarloLucibello)
• nnpack no more default for conv (#216) (@CarloLucibello)

v0.7.0

NNlib v0.7.0

Diff since v0.6.6

Closed issues:

• Overflow of Float64 logistic sigmoid derivative (#78)
• gradient causes a 'MethodError: no method matching ∇maxpool' error for NNlib.maxpool (#188)
• Provide artifact for BB download (#192)
• NNPACK_CPU_THREADS and thread pool issue (ERROR: KeyError: key 3 not found) (#200)
• Wrong MaxPool gradient (#205)

Merged pull requests:

• update travis (#182) (@CarloLucibello)
• Add comment for wrapper functions in CuArrays.jl (#183) (@matsueushi)
• Improve docs (#190) (@matsueushi)
• Mutating softmax tests (#194) (@sudo-rushil)
• BatchedAdjOrTrans interface tests (#196) (@sudo-rushil)
• fix sigmoid (#198) (@cossio)
• fix allocate_threadpool when NNPACK_CPU_THREADS is set (#201) (@Alexander-Barth)
• Update package to use jll (#204) (@aviks)
• Fix maxpool gradient (#207) (@yiyuezhuo)
• updates for NNpack (#212) (@CarloLucibello)

v0.6.6

NNlib v0.6.6

Diff since v0.6.5

Closed issues:

• tag new release (#178)
• add docstrings for batch multiplications (#179)

Merged pull requests:

• Added some Activation functions (#175) (@AdarshKumar712)
• Add fallback batched_mul! (#181) (@mcabbott)
• update travis (#182) (@CarloLucibello)

v0.6.5

NNlib v0.6.5

Diff since v0.6.4

Closed issues:

• generalize softmax (#77)
• Add maxpool and meanpool methods (#120)
• Feature Request: Add Mish activation (#144)
• ∇conv_data_direct produces incorrectly shaped output for 5D arrays (#156)
• conv fails on GPU for first run (#161)
• Potential overflow of tanh (#164)

Merged pull requests:

• implementation for batch-wise matrix multiplication (#100) (@chengchingwen)
• Added Mish activation (#145) (@digantamisra98)
• Preserve the type in differentiation (#149) (@matsueushi)
• conv.jl: fix @debug message (#155) (@bcdarwin)
• Update test cases of activation functions (#162) (@matsueushi)
• Update deps for CI (#166) (@matsueushi)
• Install TagBot as a GitHub Action (#167) (@JuliaTagBot)
• Activation functions added (#168) (@sriyash421)
• Fix debug string (#169) (@mcabbott)
• docstring, type annotation (#170) (@matsueushi)
• Fix output shape of ∇conv_data_direct! (#171) (@matsueushi)
• add some docstrings (#172) (@CarloLucibello)
• Update Codecov badge (#173) (@matsueushi)
• Typo (#174) (@dhairyagandhi96)
• tag new release (#177) (@CarloLucibello)

v0.6.4

v0.6.4 (2020-01-16)

Diff since v0.6.3

Closed issues:

• Conv gives incorrect results when multithreading (#158)

Merged pull requests:

• Use a thread-sharded col in im2col-related functions (#160) (staticfloat)

v0.6.3

v0.6.3 (2020-01-14)

Diff since v0.6.2

Closed issues:

• Fails on loading NNlib/udnNA/deps/deps.jl (#157)

Merged pull requests:

• handle dims properly in gradient of logsoftmax (#148) (tanhevg)

v0.6.2

v0.6.2 (2019-12-24)

Diff since v0.6.1

Merged pull requests:

• Refactor tests and make NNPACK optional (#154) (dhairyagandhi96)
• Fix asymmetric padding arguments to NNPACK's conv\(\) (#153) (staticfloat)

v0.6.1

v0.6.1 (2019-12-20)

Diff since v0.6.0

Closed issues:

• UndefVarError: project not defined (#141)
• conv with No Dilation (#138)
• Bad gradient for pool with padding (#133)
• Convolution is not inferable (#125)
• Error building NNlib (#116)
• NNPACK convolutions behaving weirdly! (#114)
• Register v0.6.0 (#111)

Merged pull requests:

• Update Requires compat and add CompatHelper (#150) (ChrisRackauckas)
• Add compat section to Project.toml and rm REQUIRE (#143) (dhairyagandhi96)
• conv_direct!(): The performance fix (#142) (staticfloat)
• Fix typo: actvation -> activation (#136) (mprat)
• logsoftmax with dims (#135) (mcabbott)
• Fix typo in maxpool backprop (#134) (staticfloat)
• Make softmax! dimension-agnostic (#130) (merckxiaan)
• Remove timer outputs as they mess with type inference (#129) (avik-pal)
• Added wrappers for conv and depthwiseconv (#127) (thebhatman)
• Improve numerical stability of logsoftmax gradient (#126) (ornithos)
• Avoid error while trying to build NNPACK on unsupported platforms (#122) (avik-pal)
• maxpool and meanpool added (#121) (thebhatman)
• NNPACK Convolution Patch (#115) (avik-pal)
• Add NNPACK support (#67) (avik-pal)

v0.6.0

<description of version>