Higher Loss Values in Parametric UMAP Notebooks

lmcinnes / umap

Uniform Manifold Approximation and Projection

BSD 3-Clause "New" or "Revised" License

7.41k stars 805 forks source link

Higher Loss Values in Parametric UMAP Notebooks #1142

Open erenozcelik opened 2 months ago

erenozcelik commented 2 months ago

I've been experimenting with the Parametric UMAP using the example notebooks. However, I'm encountering an issue where the loss values I'm getting are significantly higher than in notebooks. This has occurred with both the second (02.0-parametric-umap-mnist-embedding-convnet.ipynb) and third (03.0-parametric-umap-mnist-embedding-convnet-with-reconstruction.ipynb) notebooks.

timsainb commented 2 months ago

I updated the notebooks to match the keras update in this PR. Have not pulled it in yet:

https://github.com/lmcinnes/umap/pull/1123

erenozcelik commented 2 months ago

Thank you for the quick reply. I tested with UMAP versions 0.5.2 and 0.5.3 and obtained the same results as in the notebooks. Starting from version 0.5.4, the results begin to differ. I got exactly the same results as in your keras-umap branch when using newer versions. I was wondering which change is causing this difference.