Conversation
add REPL to deps
oops, forgotten figure
|
Super helpful review @OkonSamuel , thank you. Your suggestions will make understanding the code much better. Unfortunately, I'm having trouble committing the suggestions in a batch. Any chance you have a branch that already includes them collectively? If not, let me know and I'll manually commit them. |
Hi Anthony, I'm sorry I didn't commit them in a branch. You can manually commit them and I can go through it once more before you merge it in. |
Co-authored-by: Okon Samuel <[email protected]>
Co-authored-by: Okon Samuel <[email protected]>
Co-authored-by: Okon Samuel <[email protected]>
Co-authored-by: Okon Samuel <[email protected]>
Co-authored-by: Okon Samuel <[email protected]>
Co-authored-by: Okon Samuel <[email protected]>
Co-authored-by: Okon Samuel <[email protected]>
Co-authored-by: Okon Samuel <[email protected]>
Co-authored-by: Okon Samuel <[email protected]>
Co-authored-by: Okon Samuel <[email protected]>
|
@OkonSamuel I was able to commit your suggestions by doing them one at a time in reverse order. There was one instance where I think you inadvertently deleted a little too much, which I have restored, namely lines 76-80 of src/functions.jl. I have reviewed the rendering of the docs and it looks good. Fixed a couple of broken links that I discovered. Ready for your second look. |
Closes #25
Addition of
AveragePrecision(a form of "area under the precision-recall curve") to be provided in a follow-up PR.Implementation notes
As in sk-learn, I have added a more general core method
Functions.confusion_counts_at_threshold(in sk-learn this is calledconfusion_matrix_at_thresholds) and have refactored the existingFunctions.rocfunction to make use of that. The newFunctions.precision_recall_curvemethod is then pretty trivial to implement. When you compute what you need forroc, you basically get all the confusion matrix entries along the way, so there's no real loss in performance for the refactoredroc.There's some considerable refactoring of document strings to lower maintenance burden, which is a little complicated, so reviewer should generate the Documenter.jl manual and review:
Functions.precision_recall_curve,Functions.roc,Functions.confusion_counts_at_thresholds,precision_recall_curve,roc.There were a number of mistakes in the
rocdocstrings around the threshold and associated bins, which hopefully I have correct now.I have also fixed some incorrect logging tests that crept in a recent PR on the Boyce Index.