v1.3.0

v1.3.0 (2019-10-25)

Diff since v1.2.7

Closed issues:

• progressbar should print on stderr by default (#509)
• RNN serialization issues (#506)
• im2col! incompatibility with Complex numbers (#500)
• v1.2.5 : Out of gpu memory (#485)
• conv4 fails for upscale > 1 (#444)
• rename upscale to dilation (#434)
• Element-wise power of KnetArray give NaN results (#108)

Merged pull requests:

• solved remaining curand out of memory problem (#514) (denizyuret)
• fixing #506: rnn serialization and gc issues (#510) (denizyuret)

v1.2.7

v1.2.7 (2019-09-29)

Diff since v1.2.6

Closed issues:

• bmm! example error (#495)
• Windows & ARM error in conv4: UndefVarError: lib not defined (#489)
• lib not defined on using Knet (#484)

Merged pull requests:

• fixes #495, adds transpose functionality to bmm, faster cpu kernel (#499) (ekinakyurek)
• Fix for Element-wise power of KnetArray give NaN results #108 (#498) (denizyuret)
• Switch to NNlib for CPU conv/pool in pure Julia. (#494) (denizyuret)
• Add CI for windows, freebsd, ARM (#490) (ianshmean)

v1.2.6

v1.2.6 (2019-09-20)

Diff since v1.2.5

Closed issues:

• ERROR: LoadError: LoadError: "loggamma" is not defined in module Knet (#483)
• RNN ERROR: AssertionError: vec(value(x)) isa WTYPE (#482)
• Feature request for KnetArray: argmax and getindex(::KnetArray{T,3}, ::Colon, ::Array{CartesianIndex}) (#368)
• Mean of a KnetArray (#110)

Merged pull requests:

• Use CuArrays.usage_limit[] to improve stability (#491) (denizyuret)
• new memory manager, more KnetArray functions (#488) (denizyuret)

v1.2.5

v1.2.5 (2019-09-04)

Diff since v1.2.4

Closed issues:

• "vcat" doesnt work for 3D KnetArrays (#470)
• Knet's vcat causing ambiguity with Base (#321)
• Feature request: more KnetArray indexing (#198)

Merged pull requests:

• dy/comp542 (#480) (denizyuret)

v1.2.4

v1.2.4 (2019-08-10)

Diff since v1.2.3

Closed issues:

• Scalar value on GPU (#476)
• Knet's permutedims is slower than Base's` (#251)
• KnetArray transpose and permutedims missing. (#39)

Merged pull requests:

• Permutedims (#474) (denizyuret)

v1.2.3

v1.2.3 (2019-07-25)

Diff since v1.2.2

Closed issues:

• mat bug for dims=3 (#472)
• Broadcast for bmm (#464)

v1.2.2

v1.2.2 (2019-05-25)

Diff since v1.2.1

Closed issues:

• rnn hidden state gradients (#463)
• erf, erfc are not defined for KnetArray (#455)
• notebook fix (#451)
• Knet.randn! gives error: ccall: could not find function cudaGetErrorString (#448)
• UndefVarError: lib not defined (#440)
• second derivative does not work for nll (#439)
• Knet cannot use GPU (#429)
• conv4 not work in CPU (#411)

Merged pull requests:

• Higher order derivatives for softmax/logp (#447) (denizyuret)
• Fix typo in LeNet Chain call (#445) (Alexander-Barth)
• gamma, lgamma, digamma work with GPU (#294) (xukai92)

Knet v1.2.1 bugfix release

• Serialization bug fix.
• Fixed eltype, size etc. for Minimize, Converge etc.
• Transpose and matmul now work with 1-D KnetArrays.
• Added intro learning notebook.
• RNN: Ignore trailing ones when comparing sizes.
• Julia 1.2 compat fixes.

Knet v1.2.0 New training interface based on iterators.

• New training interface based on iterators.
• Progressbar and converge utilities.
• RNN unboxes hidden states in backward pass making value(h) unnecessary on GPU.
• rnnparam and rnnparams no longer take a w argument.
• RNN applies dropout to input like other layers.
• mat takes a dims keyword argument that makes it useful for both RNNs and CNNs.
• Dropout automatically figures out and does nothing outside of @diff context.
• Fixed inplace assignment for Params and KnetArrays.
• Julia 1.0 fixes for goldensection.
• Improved default parameters for all optimizers tested on MLP, CNN, RNN.
• All notebooks and documentation updated.
• New iterator and quickstart notebooks.
• Updated to Documenter 0.21.

Knet v1.1.2 support for broadcasting user functions

• Support for broadcasting user defined functions.
• Added batch matrix multiplication.
• Added tests and docs for new RNN interface.
• Improved serialization and JLD file I/O.
• Added julia/base demo to tutorial/08.charlm
• Renamed broadcast.jl -> binary.jl and broadcast_ops -> binary_ops.