Skip to content

v0.14.0

Latest

Choose a tag to compare

@github-actions github-actions released this 04 Nov 16:47
· 1 commit to main since this release

BetaML v0.14.0

Diff since v0.13.1

New GPU support and NN module refactoring

  • vendor-independent (Cuda, Metal, ...) GPU support for neural networks, via a new device option (:auto, :cpu, :gpu or an MLDataDevices device) in NeuralNetworkEstimator, AutoEncoder and the MLJ models (NeuralNetworkRegressor, MultitargetNeuralNetworkRegressor, NeuralNetworkClassifier, AutoEncoder): with the default :auto, the network trains wherever the data is and predictions come back on the same device as the input.
  • efficiency improvement of all neural network models (cpu included) thanks to batch processing at the inner level (so each layer op is a matrix multiplication rather than a vector by matrix)
  • added ForwardDiff for AD in the GPU in some cases

Breaking changes

  • ScalarFunctionLayer with learnable weights: there is now one weight per node, or a single weight shared by all nodes, and the layer computes yᵢ = f(xᵢ,wᵢ). The user-provided dfx(x,w) and dfw(x,w) are now scalar functions; when omitted, they are computed with ForwardDiff.
  • crossentropy: the default weight is now the scalar 1. The function works on matrices and keeps the element type of ŷ.
  • PerceptronClassifier_lp and PegasosClassifier_lp: the field weigths is renamed to weights.

Issue solved:

  • #68
  • GeneralImputer: fixed for single-record inputs.

Documentation and infrastructure

  • Proofread docs and docstrings, aligned with the code.
  • Updated the tutorials for the current Flux API.
  • New static Benchmarks page, generated by docs/benchmarks/run_benchmarks.jl. It includes CPU vs CUDA comparisons with Flux.
  • The default branch is now main, and the CI actions are updated.

Closed issues:

  • Correct the predict in AutoEncoder to consider non-vector layer outputs (#68)
  • Mixtures.jl Precompilation Failure with Distributions v0.25.129 (#82)