v0.6.0

MLJFlux v0.6.0

Diff since v0.5.1

All models, except ImageClassifier, now support categorical features (presented as table columns with a CategoricalVector type). Rather than one-hot encoding, embeddings into a continuous space are learned (i.e, by adding an embedding layer) and the dimension of theses spaces can be specified by the user, using a new dictionary-valued hyperparameter, embedding_dims. The learned embeddings are exposed by a new implementation of transform, which means they can be used with other models (transfer learning) as described in Cheng Guo and Felix Berkhahn (2016): Entity Embeddings of Categorical Variables.

Also, all continuous input presented to these models is now forced to be Float32, but this is the only breaking change.

Merged pull requests:

• Update docs (#265) (@ablaom)
• Introduce EntityEmbeddings (#267) (@EssamWisam)
• Fix l2 loss in MultitargetNeuralNetworkRegressor doctring (#270) (@ablaom)
• automatically convert input matrix to Float32 (#272) (@tiemvanderdeure)
• Force Float32 as type presented to Flux chains (#276) (@ablaom)
• For a 0.6.0 release (#277) (@ablaom)

Closed issues:

• deprecated warning in documentation (#236)
• Some minor doc issues (#258)
• Fix code snippet in Readme (#266)
• MultitargetNeuralNetworkRegressor doc example doesn't work as intended (#268)

v0.5.1

MLJFlux v0.5.1

Diff since v0.5.0

Merged pull requests:

• Revamp of docs (#263) (@ablaom)
• For a 0.5.1 release (#264) (@ablaom)

v0.5.0

MLJFlux v0.5.0

Diff since v0.4.0

• (new model) Add NeuralNetworkBinaryClasssifier, an optimised form of NeuralNetworkClassifier for the special case of two target classes. Use Flux.σ instead of softmax for the default finaliser (#248)
• (internals) Switch from implicit to explicit differentiation (#251)
• (breaking) Use optimisers from Optimisers.jl instead of Flux.jl (#251). Note that the new optimisers are immutable.
• (RNG changes.) Change the default value of the model field rng fromRandom.GLOBAL_RNG to Random.default_rng(). Change the seeded RNG, obtained by specifying an integer value for rng, from MersenneTwister to Xoshiro (#251)
• (RNG changes.) Update the Short builder so that the rng argument of build(::Short, rng, ...)
  is passed on to the Dropout layer, as these layers now support this on a GPU, at
  least for rng=Random.default_rng() (#251)
• (weakly breaking) Change the implementation of L1/L2 regularization from explicit loss penalization to weight/sign decay (internally chained with the user-specified optimiser). The only breakage for users is that the losses reported in the history will no longer be penalized, because the penalty is not explicitly computed (#251)

Merged pull requests:

• Fix metalhead breakage (#250) (@ablaom)
• Omnibus PR, including switch to explicit style differentiation (#251) (@ablaom)
• 🚀 Instate documentation for MLJFlux (#252) (@EssamWisam)
• Update examples/MNIST Manifest, including Julia 1.10 (#254) (@ablaom)
• ✨ Add 7 workflow examples for MLJFlux (#256) (@EssamWisam)
• Add binary classifier (#257) (@ablaom)
• For a 0.5.0 release (#259) (@ablaom)
• Add check that Flux optimiser is not being used (#260) (@ablaom)

Closed issues:

• Stop using implicit style differentiating (#221)
• Update examples/MNIST (#234)
• MultitargetNeuralNetworkRegressor has surprising target_scitype (#246)
• Error fit GPU (#247)
• Metalhead has broken MLJFlux (#249)

v0.4.0

MLJFlux v0.4.0

Diff since v0.3.1

Merged pull requests:

• Bump Metalhead to 0.9 - making CUDA optional (#240) (@mohamed82008)
• For a 0.4.0 release (#243) (@ablaom)

v0.3.1

MLJFlux v0.3.1

Diff since v0.3.0

• Improve the error message for faulty builders (#238)

Merged pull requests:

• Rethrow the error caught when builder fails (#238) (@ablaom)
• For a 0.3.1 release (#239) (@ablaom)

v0.3.0

MLJFlux v0.3.0

Diff since v0.2.10

Merged pull requests:

• Off-by-one error in clean! method (#227) (@MarkArdman)
• CompatHelper: bump compat for Flux to 0.14, (keep existing compat) (#228) (@github-actions[bot])
• Actions node 12 => node 16 (#229) (@vnegi10)
• Bump compat for Metalhead (#232) (@ablaom)
• For a 0.3 release (#235) (@ablaom)

v0.2.10

MLJFlux v0.2.10

Diff since v0.2.9

Closed issues:

• Builder does not appear to build an architecture compatible with supplied data (#216)
• Calling predict on matrix throws error (#218)

Merged pull requests:

• Sorting out issue with calling predict on matrix (#219) (@pat-alt)
• Dispatch fit! and train! on model for greater extensibility (#222) (@pat-alt)
• For a 0.2.10 release (#223) (@ablaom)
• Fix target scitype for multitarget regressor (#224) (@ablaom)

v0.2.9

MLJFlux v0.2.9

Diff since v0.2.8

• (bug fix) Address improper scaling of l2/l1 loss penalty with batch size (#213) @mohamed82008

Closed issues:

• Penalty (wrongly) not multiplied by the relative batch size (#213)

Merged pull requests:

• Divide penalty by n_batches (#214) (@mohamed82008)
• For a 0.2.9 release (#215) (@ablaom)

v0.2.8

MLJFlux v0.2.8

Diff since v0.2.7

• #212

Closed issues:

• Define a working default Flux chain builder for ImageClassifier (#162)
• Broken link in MNIST example (#204)
• MNIST example has wrong axis label: (#210)

Merged pull requests:

• Add preliminary Metalhead.jl integration (#208) (@ablaom)
• Add docstrings! (#209) (@josephsdavid)
• Fix broken link in examples (#211) (@ablaom)
• For a 0.2.8 release (#212) (@ablaom)

v0.2.7

MLJFlux v0.2.7

Diff since v0.2.6

Merged pull requests:

• CompatHelper: bump compat for Flux to 0.13, (keep existing compat) (#202) (@github-actions[bot])
• For a 0.2.7 release (#203) (@ablaom)