Meta issue: lssues for possible collaboration with UCL #673
Comments
None of that happened ... well, in Julia. Ultimately, and perhaps not too surprisingly, we also ended up at a point where double dispatch would be great (function space cross-products of distribution defining functions), but R6 doesn't have an easy way for double dispatch. R7 perhaps... On a side note, @aintoha also calculated a larger batch of integrals that might be useful to re-use instead of re-deriving them. |
Really good to know, thanks!! cc @giordano
Ha ha. |
|
@giordano I have reviewed the checklist today, 11 March 2021. The items with a checkbox still look good, more-or-less in the order given. In particular, to start with a review of latency in MLJBase, and to move measures out. Let's talk details in a call. |
|
Some miscellaneous "smaller" issues: cc @giordano |
|
Closing as stale |
Disintegration of MLJModels (medium)
Disintegration of MLJModels MLJModels.jl#244 : priority would be for GLM, with a blank repo at https://github.com/alan-turing-institute/MLJGLMInterface.jl ; you could use https://github.com/alan-turing-institute/MLJNaiveBayesInterface.jl as a template.doneUniversal transformer for wrapping univariate transformers (medium) Universal table transformer combining univariate transformations dispatched on schema MLJModels.jl#288 : more detailed design proposal needed. Familiarity with the logic of existing
Standardizerhelpful. This may already be a good template for what we want to do here (just replaceUnivariateStandardizerby a user-specified one). Need to worry aboutinverse_transformwhen implemented.Disintegration of MLJBase (medium) parts of Disintegration of MLJBase (discussion and tracking issue) MLJBase.jl#416 , in particular Serialization and OpenML, which seem to be hefty. Worth exploring which dependencies are causing most latency. Also, StatisticalMeasures (medium-long). Added note: Measures currently depend on UnivariateFinite, which in turn depends on Distributions, but only the base API. See this issue: Migrate UnivariateFinite (for categorical distributions) out to new package MLJBase.jl#504
pdfnorm for Distributions.jl (??) Feature request: provide L2 norm of each distribution's pdf JuliaStats/Distributions.jl#806: this is one I believe Mose discussed with @fkiraly but was not completed, in an earlier engagement. Would be good to know what the status of that work is.
investigate source of package compiler issues (medium) (MLJBase + PackageCompiler + Distributed => error MLJBase.jl#427). Suggest commenting out src/composition/ for a start.
[ ]
Review/Redesign of model registry (long) Review of model registry MLJModels.jl#321
Test new API proposal to improve data resampling performance (medium) taking performance issue more seriously MLJBase.jl#309 (comment)doneAdd visualisation to model tuning results (medium) Add plot recipe for visualising multiple hyperparameter tuning outcomes MLJTuning.jl#41
Populate model metadata with good default hyperparameter ranges (short-medium) Populate default hyperparameter ranges in the model metadata. MLJModels.jl#322
Allow use of sample and class weights in sk-learn models (medium) Add sample weights support for sk-learn models MLJScikitLearnInterface.jl#17 (and the related https://github.com/alan-turing-institute/MLJModels.jl/issues/127)
Add control over logging level (short) [FR] allow disabling of logging #255
added mid November
cleanup of measures (short) name, aliases cleanup for measures MLJBase.jl#450doneadded early January 2020
roll out data front-ends for models (medium) Implement the optional data front-end that models will be able to implement after Realize performance improvements for models implementing new data front-end MLJBase.jl#501 .
TLC for DataScienceTutorials (short - medium) The tutorials need updating to latest version of MLJ, and some contributors have made PR's that are languishing.
The text was updated successfully, but these errors were encountered: