kevinmoonlab / grae Goto Github PK

show metrics function should check if df is empty before processing (this will happen when working on a single data set with no supervised metrics)

Partitioning the manifold leads to locally constant time labels because of the lack of time granularity. This raises errors when computing correlations and MI.

Running UMAP score method on train split before running on test split leads to obviously wrong test embeddings.

(This was hot fixed for the paper by scoring test before train and the results are valid, but the bug should still be investigated).

A lot of pairs have 0 distance, but aren't exactly equal. Need to remove duplicates given a small tolerance.

Roll class is bugged when used alone. SwissRoll (which is the children class we use in the paper) is fine.

I tried to install the git repository via clone, and I got an error while running it. Then I tried to find the error and I found that the error occurred due to the installation of comet.ml ==3.1.14. Error note indicated that it's an error of the repo.

Does anyone have the same issue as me? How should I fix it? Can you please guide me on this?

Embedding plots for Rotated Digits are now a bit different from the paper version. Metrics are still good. Maybe something is not seeded properly.

Thank you for your work and for this beautifully constructed software!

I'm looking forward to adapting some of the source code for another project. Naturally, I would like to pay my respects to the team who developed the code, and intended to do so by attaching the project license alongside a reference to this repository and to your publication at the top of the script. However, the project does not specify any license.

How should someone interested in adapting the code available in this repository proceed?

Cheers,
Davi