VpTree does not accept float in knncolle
shrit opened this issue · comments
Hi Aaron,
It seems that VpTree does not accept float value as the std::tuple is hard coded to double.
Here is the code:
Here is the error when trying to use float for Umap:
/./knncolle/VpTree/VpTree.hpp:150:29: error: no matching function for call to ‘std::tuple<int, const double*, double>::tuple(int&, const float*, int)’
150 | items.push_back(DataPoint(i, vals + i * num_dim, 0));
It seems that const INTERNAL_t*, is not being propagated correctly from umap, and it is using the default value causing the above error.
I know this is belong to the knn repo, but it is here when the error is happening.
Any easy solution ?
Many thanks
Hopefully fixed by ff42321. Note that VpTree is a decent default for small datasets but you'll probably want to use one of the approximate methods (Annoy or HNSW) for anything larger.
@LTLA Thank you very much for the quick fix, do you know how I can specify Annoy as a template parameter for Umap ? is there an example that shows how this is done ?
Many thanks
The README has an example:
umappp::Umap x;
knncolle::AnnoyEuclidean<> searcher(ndim, nobs, data.data());
x.run(&searcher, 2, embedding.data());You might need to add <float> in various places above.
TBH I don't usually use this method signature for run(), as I compute the neighbor list manually and provide it to Umap::run(). This is because I might end up re-using the neighbor list with different UMAP parameters so I just compute the same set of neighbors once and use it across different UMAP runs.
If you want some more inspiration, check out these bindings:
Perfect, thank you very much, I will keep this open for a while until I have the chance to run it again and let you know if I have any question