Skip to content

NEWS.md

Latest commit

 

History

History
31 lines (25 loc) · 1.41 KB

NEWS.md

File metadata and controls

31 lines (25 loc) · 1.41 KB
Raw

NEWS

v0.18

Refactor of GPU training backend

  • Computations are now done through KernelAbstractions.jl instead of CUDA specific kernels. Objective is to eventually have full support for AMD / ROCm in addition to current NVIDIA / CUDA devices.
  • Important performance increase, notably for larger max depth. Training time is now closely increase linearly with depth.

Breaking change: improved reproducibility

  • Training returns exactly the same fitted model for a given learner (ex: EvoTreeRegressor).
  • Reproducibility is respected for both cpu and gpu. However, results may differ between cpu and gpu. Ie: reproducibility is guaranteed only within the same device type.
  • The learner / model constructor (ex: EvoTreeRegressor) now has a seed::Int argument to set the random seed. Legacy rng kwarg will now be ignored.
  • The internal random generator is now Random.Xoshiro (was previously Random.MersenneTwister with rng::Int).

Added node weight information in fitted trees

  • The train weight reaching each of the split/leaf nodes is now stored in the fitted trees. This is accessible via model.trees[i].w for the i-th tree in the fitted model. This is notably inteded to support SHAP value computations.
config = EvoTreeRegressor(; max_depth=3)
m = fit(config; x_train, y_train)
m.trees[2].w

7-element Vector{Float32}:
 8000.0
 5000.0
 3000.0
  750.0
 4250.0
 1250.0
 1750.0