n_components=2

[MonkeyChap]

trafficstars

Hi

Is there any roadmap for allowing higher dimensionality ? tSNE can also be used to reduce the dimension in datasets - e.g from 200 down to 10. Being able to do this with something much faster than sklearn would be really cool.

Thanks

Ian

[lbugnon]

(Edited) Mark as resolved, this is already working with param n_components

[drcege]

The package in pypi is not updated?
Still AssertionError: n_components should be 2

[smt-HS]

I encountered the same issue

[DmitryUlyanov]

Updated the package. Please try now.

[aCampello]

Adding to the discussion, although this is possible in practice, TSNE might not be the method you are looking for to reduce to higher dimensions. A word of caution was provided by the T-SNE authors themselves in the original paper VISUALIZING DATA USING T-SNE:

It is not obvious how t-SNE will perform on the more general task of dimensionality reduction (i.e., when the dimensionality of the data is not reduced to two or three, but to d > 3 dimensions). To simplify evaluation issues, this paper only considers the use of t-SNE for data visualization. The behavior of t-SNE when reducing data to two or three dimensions cannot readily be extrapolated to d > 3 dimensions because of the heavy tails of the Student-t distribution. In high-dimensional spaces, the heavy tails comprise a relatively large portion of the probability mass under the Student-t distribution, which might lead to d-dimensional data representations that do not preserve the local structure of the data as well. Hence, for tasks in which the dimensionality of the data needs to be reduced to a dimensionality higher than three, Student t-distributions with more than one degree of freedom10 are likely to be more appropriate.