Results look different on v3 than v2

[stu-blair]

Hi, I recently tried upgrading from v2.1 to v3.0.1, and when re-running the same code on the same data with the same settings, the results are still good, but when visualized, look quite different -- small isolated islands, rather than the typically more continuous shape that fills the space. I reproduced this on a variety of different datasets.

I know it's not expected to be deterministic, but I was wondering if you have any idea or suggestion on how to get the results looking the way they were previously.

Here are some examples:

v2: → v3:

---

v2: → v3:

---

v2: → v3:

---

And this is the code I call to use the library:

from tsnecuda import TSNE
results = TSNE(perplexity=perplexity,
               n_iter=iterations,
               learning_rate=learning_rate).fit_transform(input_data)

Many thanks!

[stu-blair]

Hi @DavidMChan , I found reducing perplexity seems to help the visuals look more like they used to. However, this does not work consistently for all datasets, unlike tsne-cuda v2, where a perplexity of 30 almost always produced good-looking results for a variety of datasets.

Any idea why the perplexity setting's effects have changed in the new version?