expose nmslib HNSW method parameters for recall fine-tuning

[labroskokkalas]

I have build a hnsw index ( cosine distance ) with about 6M points. Using the same 6M points as test dataset, the 1st neighbor recall is about 0.979 which is high but not sufficient for my application. Setting M,post,efConstruction and efSearch parameters (https://github.com/nmslib/nmslib/blob/master/manual/methods.md) in tensorflow_similarity/search/nmslib_search.py

self._search_index.setQueryTimeParams({'efSearch': 2000}) self._search_index.createIndex({'post': 2,'efConstruction':'2000','M':64},print_progress=show)

results to a 1st neighbor recall of 0.999. These parameters provide a trade off between recall and indexing_time/memory_usage and would be very useful in production systems for recall fine-tuning.

[owenvallis]

Thanks for sharing the results. We're looking at making the storage and indexing layer more flexible and I'll be sure to include this in the changes.