The TSNE class has no fit method. This issue is basically asking whether this
is ok.

IMHO no: we are breaking the API contract here.

I was wondering about this too there is no way to put new data through the existing model ie transform / predict. So although I can create very nice clusters with T-SNE, I cannot use the model to embed new data and use those values for prediction.

Looking in more depth at the algorithm and code it may not be possible to embed new values with an existing model. Could someone confirm??

It's not intended to be inductive, i.e. generalise to new samples. But
other estimators in scikit-learn that are similarly transductive still
support fit.

On 27 January 2015 at 15:13, 8evers notifications@github.com wrote:

I was wondering about this too there is no way to put new data through
the existing model ie transform / predict. So although I can create very
nice clusters with T-SNE, I cannot use the model to embed new data and use
those values for prediction.

Looking in more depth at the algorithm and code it may not be possible to
embed new values with an existing model. Could someone confirm??

—
Reply to this email directly or view it on GitHub
#4133 (comment)
.

It's not intended to be inductive, i.e. generalise to new samples.

I thought that the new Barn-Hut solver could to that.

I confirm that this is fixed in #4025 that is still under review (I plan on finishing my pass on the code that today).

The semantics of fit_transform(X) are not equivalent to fit(X) then transform(X) for that class but this is a mathematical limitation of the model. As long that it's properly documented in the docstrings it's fine with me.

Fixed by #4136, will get a even more of a fix in #4025