Ludwig Schmidt

The C++ library never copies data. For the Python bindings, the story is a bit more complicated. There are two levels of wrapping going on in the Python bindings: A:...

I'm not exactly sure how you would use the generated hash tables on a smaller instance. But #46 is indeed the ideal solution for loading and storing hash tables.

@funtusov Could you control the memory usage of FALCONN in the end?

OK, thanks for the update. We'll investigate this when we update the Python wrapper in #69 .

Have you tried increasing the number of probes?

Hi hxwu, updates are supported in some of the core classes, in particular the DynamicLinearProbingHashTable and the DynamicCompositeHashTable: https://github.com/ludwigschmidt/fth_lsh/blob/master/lsh_table/src/probing_hash_table.h https://github.com/ludwigschmidt/fth_lsh/blob/master/lsh_table/src/composite_hash_table.h What we need next is a dynamic version of the...

Internally, find_k_nearest_neighbors relies on the methods for finding candidates. So in principle what you described shouldn't happen. Just to clarify: what distance function are you using when you find the...

FALCONN can certainly miss the nearest neighbor. But if the nearest neighbor appears in the candidate set, FALCONN should find it. What distance function are you using when you set...

The script could also figure out which version of swig to use (`swig` on Mac vs `swig3.0` on Linux), and which g++ (`g++-5` on Mac with the homebrew g++ install...

Use the right linker flag for C++ (`lc++` vs `lstdc++`).