The T-SNE latent graph is a bit difficult to read (as expected, we're squashing 512 dimensions into 2...) I've upgraded the function to a 3d plot, which gives more room to the latents representations but it's still abit murky to be honest.
I'm trying to see if it's possible to make it clearer. maybe switching to cosing function instead of the euclidean standard, but I feel I need some different approach to highlight the vectors more and make it useful, otherwise we can just drop this graph.
The T-SNE latent graph is a bit difficult to read (as expected, we're squashing 512 dimensions into 2...) I've upgraded the function to a 3d plot, which gives more room to the latents representations but it's still abit murky to be honest.
I'm trying to see if it's possible to make it clearer. maybe switching to cosing function instead of the euclidean standard, but I feel I need some different approach to highlight the vectors more and make it useful, otherwise we can just drop this graph.