It is annoying to have to use Multiscale affinities, because we have to instantiate a separate Multiscale class.
Description of changes
By default, we used PerplexityBasedNN, which -- under the hood -- just converted the perplexity to single-element list and called the same underlying machinery that MultiscaleMixture did. In this PR, I change the default perplexity model to MultiscaleMixture, which allows calls like
tsne = TSNE(perplexity=[50, 500])
# or
embedding.transform(x_new, perplexity=[10, 50])
This doesn't have any breaking changes and will produce the same results as before, but will be more flexible.
The parameter naming is a bit annoying/confusing. Currently, I've made it so that both the TSNE() constructor and transform both only accept perplexity and not perplexities.
Extra
Also adds support for Python 3.10 (fixes #205)
Fixes potential bug on MacOS builds using clang-13, where we need to add -fno-finite-math-only flag
Issue
It is annoying to have to use Multiscale affinities, because we have to instantiate a separate
Multiscaleclass.Description of changes
By default, we used
PerplexityBasedNN, which -- under the hood -- just converted the perplexity to single-element list and called the same underlying machinery thatMultiscaleMixturedid. In this PR, I change the default perplexity model toMultiscaleMixture, which allows calls likeThis doesn't have any breaking changes and will produce the same results as before, but will be more flexible.
The parameter naming is a bit annoying/confusing. Currently, I've made it so that both the
TSNE()constructor andtransformboth only acceptperplexityand notperplexities.Extra
Includes