v0.4.17

Lux v0.4.17

Diff since v0.4.16

Closed issues:

• Inconsistent descripition of PairwiseFusion (#130)
• No method matching with argument IRTools.Inner.Undefined in gradient computation. (#134)

Merged pull requests:

• LSTM docs: don't go over first element in sequence twice (#132) (@visr)
• fix PairwiseFusion docs (#133) (@MilkshakeForReal)
• Generic recurrent cells (#136) (@jumerckx)
• relu tests with finite diff is too unreliable (#137) (@avik-pal)

v0.4.16

Lux v0.4.16

Diff since v0.4.15

Merged pull requests:

• Trainable hidden states (#117) (@lungd)

v0.4.15

Lux v0.4.15

Diff since v0.4.14

Merged pull requests:

• Relax ComponentArrays compat (#126) (@avik-pal)

v0.4.14

Lux v0.4.14

Diff since v0.4.13

Merged pull requests:

• Rnn bias deprecation (#120) (@lungd)
• Add use_bias kwarg to LSTMCell and GRUCell (#121) (@lungd)

v0.4.13

Lux v0.4.13

Diff since v0.4.12

Closed issues:

• Make it easier to pass empty state st = (;) (#118)
• is there transposed convolution (#122)
• Support for multidimensional data? (#123)

Merged pull requests:

• Fix Scale's paramlength (#116) (@lungd)
• Update docs for dense layer (#124) (@avik-pal)
• Upper bound ComponentArrays (#125) (@avik-pal)
• Layer Normalization Implementation (#127) (@avik-pal)

v0.4.12

Lux v0.4.12

Diff since v0.4.11

Closed issues:

• optimising parameters with Optimization.jl (#108)
• add OrdinaryDiffEq downstream test (#110)

Merged pull requests:

• fix state update in NeuralODE example (#107) (@gabrevaya)
• Deprecate elementwise_* and applyactivation (#113) (@avik-pal)
• Go through the dense bias deprecation (#114) (@avik-pal)

v0.4.11

Lux v0.4.11

Diff since v0.4.10

Closed issues:

• WeightNorm causes NaN for Conv layer gradients (#95)

Merged pull requests:

• [LuxTraining] Wrappers for less clunky training loops (#104) (@avik-pal)
• Fixes WeightNorm with zero Parameter bug (#106) (@avik-pal)

v0.4.10

Lux v0.4.10

Diff since v0.4.9

Closed issues:

• Lighter syntax for stateless networks? (#83)
• Scalar indexing problem for the NeuralODE example (#92)
• Basic example from Migrating from Flux to Lux is broken || normalization issue (#94)
• [Feature request] Another type of Chain that sequentially passing x and st (#96)
• RNN and LSTM break when using GPU (#100)
• Can one compose lux layers with graph neural network (#102)

Merged pull requests:

• CompatHelper: bump compat for FluxMPI to 0.6 for package examples, (keep existing compat) (#86) (@github-actions[bot])
• Update comparison section in overview.md (#88) (@ToucheSir)
• Fix typos (#89) (@claforte)
• Fix minor typos in the docs (#93) (@gabrevaya)
• making x Float32 in migrate from Flux example (#97) (@gabrevaya)
• add init_hidden_state function (#101) (@gabrevaya)
• JLArray is now registered (#103) (@MilkshakeForReal)
• Use OneHotArrays (#105) (@MilkshakeForReal)

v0.4.9

Lux v0.4.9

Diff since v0.4.8

Merged pull requests:

• Update rrules so that we can support Yota (#85) (@avik-pal)

v0.4.8

Lux v0.4.8

Diff since v0.4.7

Merged pull requests:

• More Testing + Deprecate Nonsensical Functions + Better Naming for Kwargs (#80) (@avik-pal)
• CompatHelper: add new compat entry for Optimisers at version 0.2, (keep existing compat) (#82) (@github-actions[bot])