jiwoongim / ft-sne Goto Github PK

I don't think that the final momentum value is ever used. The momentum-decay line uses a hard-coded value of 0.85:

Appendix C of the paper mentions the learning rate and momentum decay but not the momentum switch, which would seem to force the momentum back up to a relatively large value for most of the optimization. Is the momentum decay useful in practice?

Hi, this was a very interesting paper.

Forgive me if I didn't read closely enough, but I couldn't see how many observations were in the Swiss Roll or 3 Gaussians dataset used in Figure 2 of the paper (and on the README of this repo).

Hello, thanks for this great work!
I was trying to run it with my data (dim (2331, 41376)). However, I get this error:

kl, Perplexity 100 Iteration: 1. Perplexities in [nan, nan]. Traceback (most recent call last): File "./run.py", line 85, in <module> color=color, divtype=divtypet, datatype=datatype) File "/src/tsne.py", line 96, in tsne find_sigma(X_shared, sigma_shared, N, perplexity, sigma_iters, metric, verbose) File "/src/core.py", line 204, in find_sigma raise Exception('Invalid sigmas. The perplexity is probably too low.') Exception: Invalid sigmas. The perplexity is probably too low.

I tried to specify an higher perplexity value but didn't work.
I am also sure I don't have NaNs in my data.
Thanks a lot in advance for your help!