Hi,

When I use 'cosine' as metric for sklearn.manifold.tsne I get the following error:
"ValueError: All distances should be positive, either the metric or precomputed distances given as X are not correct".

As the codomain of cosine distance is [-1,1] all output from pairwise comparison will not be positive.

For euclidean distance there is a hard coded solution for keeping distances positive by squaring them (passed as parameter).

For cosine distance I suspect that a simply 'absoluting' the values should be a fix to make it run, however seems like there should be a proper understanding of the algorithm before doing a fix (which I currently do not have unfortunately)