v1.4.0

Knet v1.4.0

Diff since v1.3.9

Closed issues:

• Use GPUArray / KnetArray (#150)
• Illegal memory access with CuArrays (#251)
• cat 4D arrays along channel dimension (#319)
• Concatenation in 3 dimension missing (#400)
• Feature request: group convolutions (#407)
• StackOverflowError when broadcast between number and KnetArray{Bool} (#502)
• Tutorial 15. quickstart fail (Knet 1.2.7, Julia 1.1.1, Windows 10) (#508)
• [bug when broadcasting KnetArray] UndefVarError: lib not defined (#534)
• Use CuArrays, eliminate dependency on nvcc by implementing gpu kernels in julia (#550)
• Is it possible to provide Knet GPU support in BinaryBuilder (#566)
• cudnn 8.0 compatibility (#569)
• Get rid of @cu calls, replace them with CUDA calls (#580)
• Create Knet.Ops20 and add benchmarking for operator set (#583)
• Knet GPU selection and auto device selection (#586)
• Create a profiling suite (#588)
• Missing website (#592)
• CUDA.pool not defined (#595)

Merged pull requests:

• Ops20 (#596) (@denizyuret)

v1.3.9

Knet v1.3.9

Diff since v1.3.8

Closed issues:

• CUDAapi doesn't seem to have find_host_compiler (#567)
• Warning: Hardware is unsupported by NNPACK so falling back to default NNlib (#576)
• addtoindex! ambiguity (#578)
• no method matching length error at model optimization (#579)

v1.3.8

Knet v1.3.8

Diff since v1.3.7

v1.3.7

Knet v1.3.7

Diff since v1.3.6

Closed issues:

• problems on test Knet and find_host_compiler (#530)
• CuArrays.jl: DEPRECATED, use CUDA.jl instead! (#570)
• travis does not push new docs (#571)
• Feature request: Depthwise and groupwise convolution (#574)

v1.3.6

Knet v1.3.6

Diff since v1.3.5

Closed issues:

• SpecialFunctions loggamma update (#486)
• test with CUDAapi v4.0 (#543)
• Memory allocation and deallocation of temporary variables (#548)
• Error when testing Knet (#553)
• Broadcasted copyto! .(+/*)= operator is undefined for AutoGrad (#554)
• Bug: Loading a model made using the Chain struct fails (#557)
• collect now needs size, collect(progress()) fails (#558)
• How to cite Knet in a research? (#559)
• Loading Knet breaks Julia's copyto! on 1.5 (#561)
• Performance degradation (#562)
• NLL loss (#563)
• Installing and using Knet doesn't work (#565)
• GPU tests suggested in the documentation failing (#568)

Merged pull requests:

• add gpu install docs for azure/ubuntu18.04 (#546) (@jw3126)
• GPU discoverability with CUDAapi v4.0 (#549) (@iuliancioarca)
• Further notes on GPU tools, especially for windows (#560) (@RocketRoss)

v1.3.5

Knet v1.3.5

Diff since v1.3.4

Closed issues:

• KnetArray error : expected UnionAll, got typeof(Knet.CuArray) (#540)
• test with CUDAapi v4.0 (#543)
• Reading from a KNetArray within a GPU kernel function (#544)
• Knet.save error in julia 1.4 (#545)

Merged pull requests:

• respect to CUDA_VISIBLE_DEVICES (#541) (@ekinakyurek)

v1.3.4

Knet v1.3.4

Diff since v1.3.3

Closed issues:

• AssertionError: vec(value(x)) isa WTYPE (#537)
• Cannot convert an object of type AutoGrad.Result{Float32} to an object of type Float64 (#538)

Merged pull requests:

• Install TagBot as a GitHub Action (#539) (@JuliaTagBot)

v1.3.3

v1.3.3 (2020-02-01)

Diff since v1.3.2

Closed issues:

• Convert KnetArray to CuArray and back (#535)
• aws images do not work (#533)
• Feature Request: Compatibility with CuArrays 1.6.0 (#532)
• julia1.3.0/knet132 build conf.jl gradcheck error (#528)
• Test error on test/serialize.jl and broken tests on test/conv.jl (#527)
• the error of test for Knet in dropout.jl (#524)
• memory usage printings (#523)

Merged pull requests:

• fixed primitive for bmm, added more gradient tests for bmm (#526) (ekinakyurek)
• A numerical example for Knet.nll for the documentation (#525) (Alexander-Barth)

v1.3.2

v1.3.2 (2019-11-29)

Diff since v1.3.1

Closed issues:

• conv4 allocates outside of the memory pool (#518)
• scaling xavier intializer comparared to other frameworks (#515)

Merged pull requests:

• Alexander barth master (#522) (denizyuret)
• fix issue #518: conv4 allocates outside of the memory pool #518 (#521) (denizyuret)
• xavier intializer using the same convention as TensorFlow and PyTorch (#520) (Alexander-Barth)

v1.3.1

v1.3.1 (2019-11-07)

Diff since v1.3.0

Merged pull requests:

• Dy/cuarrays13 (#516) (denizyuret)