Embedding Projector: UMAP and TSNE projections broken for embeddings that are not normalized

[alicialics]

Chrome 111.0.5563.64

Issue description

When using embeddings that are not normalized and sphereized, the UMAP and T-SNE are incorrect or not simply loading. See #5547 for a previous bug report.

the reason is that knn expects normalized vectors for cosine distance (cosDistNorm) rather than arbitrary vectors.

• Build and launch projector (must be from master, not https://projector.tensorflow.org)
• Select Iris demo tensor dataset
• Keep "Sphereize data" unchecked
• change projection type to T-SNE
• The end result is different from what happens when using https://projector.tensorflow.org/
• change projection tyoe to UMAP
• See "Initialize UMAP..." modal loading forever

Alternative repo:

• Build and launch projector (must be from master, not https://projector.tensorflow.org)
• Uncheck "Sphereize data" on the default Word2Vec 10k dataset
• Switch projection from "PCA" to either t-SNE or UMAP
• See the UI breaks with "Initializing t-SNE..."/"Initialize UMAP..." modal loading forever

related to: #2421

[dmfolgado]

I'm currently facing this issue. Is there any workaround? Can I pass a normalized embedding to the checkpoint? If so, what would be the correct normalization? I'm trying some approaches, but I could not work with them.

EDIT: This issue is still persistent in v2.13.0. I'm using Windows 11 and Google Chrome.

[alicialics]

@dmfolgado you can either use the built-in "Sphereize data" option or normalize the embedding yourself. I think all you need to make sure each one is a unit vector in euclidean space. Sphereizing does this and in addition normalizes the centroid to origin so it might even work better

[dmfolgado]

Thank you for the suggestion. I had already tried but I continue to get different projections between the online Projector and the offline.

Local version. Clustering the CBF time series dataset (X_test)

Online version. Clustering the CBF time series dataset (X_test)

I attach the data and metadata for reproducibility. data.zip

[dmfolgado]

I just found what was causing the discrepancy. The data I was using for the online projection was standard scaled. It seems that using standard scale data and the spherization yields to the same results.

[delale]

Any update on the stuck on "Initialize UMAP..." message issue? I am currently facing the exact same problem which I cannot recreate with the online embedding projector using the same data. Unless "Spherize data" is checked before clicking on UMAP, TensorBoard embedding projector remains stuck on "Initialising UMAP...".