an example where knn is super slow & model is not converging

[Fil]

I'm trying to port https://beta.observablehq.com/@fil/tsne-js (made with https://github.com/karpathy/tsnejs) to tfjs-tsne, and I hit two problems:

1. the kNN procedure takes forever
2. then the model doesn't converge to anything meaningful

I'm not sure what I'm doing wrong… posting here as @Nicola17 seems to think the issue might be with tfjs-tsne.

Example under @observablehq: https://beta.observablehq.com/@fil/hello-tfjs-tsne

and as a standalone page: https://bl.ocks.org/Fil/ee79c135db0c3451d4f44c60925e9466

[yinshanyang]

I have a very similar issue, both that it is on Observable and it relates to porting from karpathy/tsnejs to @tensorflow/tfjs-tsne.

karpathy/tsnejs implementation: https://beta.observablehq.com/@yinshanyang/graphic-sg-t-sne

@tensorflow/tfjs-tsne implementation: https://beta.observablehq.com/@yinshanyang/graphic-sg-t-sne-gpu

[Nicola17]

@yinshanyang we are aware of a problem that occurs when the number of dimensions is higher than the number of points (#60) A fix will be pushed soon.

[yinshanyang]

@nicola17 cool, and thanks for the response.

[Fil]

I've given it another go, but I'm stuck with errors, either Error: Failed to link vertex and fragment shaders. or Error: Unknown feature WEBGL_MAX_TEXTURE_SIZE.