FR: add projector option to skip dimensionality reduction

[aman-tiwari]

It would be great if the Embedding Projector supported precomputed dimensionality reductions.
For instance, if I wanted to use UMAP I could just compute it serverside to map to 3d/2d and have that be directly visualised rather than make a feature request here and wait for it to be implemented. This could also be useful if I don't want to compute a huge T-SNE clientside.

[jart]

Maybe something like TensorFlow.js could help it run faster in the browser cc: @jameswex. We don't have the infrastructure in place right now to do heavy lifting on the serving path. It might be possible to precompute that and log it as a summary, which the backend can then serve as read-only content.

[aman-tiwari]

That is what I meant - logging it in a special summary or having an option in the projector that just displays them as-is (i.e without PCA/TSNE)

[jart]

You could put JSON in a DT_STRING scalar tf.summary.tensor_summary. If it makes more sense to do it with a file, then that's probably fine too; the projector plugin already does special things with files.

[nikste]

I'm having 2D embedding of images that i want to visualize. I'm surprised that there is not an option to not do dimensionality reduction.
Is there any progress on this ?
Also @jart would that json file solution fit in my case? Is there any documentation how that json would need to look like? (also I'm using pytorch)