zjulearning / efanna Goto Github PK

Hi, thanks for your efforts!

Have you considered supporting fixed-radius nearest neighbor search (in addition to kNN)?

Best,
Tim

Hi,
I found your code very interesting and I would like to use for my research (SIFT-like descriptors testing), but I have two questions:

1. Does EFANNA support L1 (i.e. Manhattan) distance?
2. If descriptor elements are integer in the range 0-15 (i.e. nibbles), does the code need a specific tuning of some parameters (such as those related to the hash function)?

Thanks in advance

What should I do to get the distance between the sample and the neighbor? （returned by EFANNA）

Hi guys,

nice work. It's great to see that the idea of using proximity graphs is becoming more popular. However, you don't cover all important state of the art. Please, consider comparing to at least the following libraries that beat FALCONN by an order of magnitude or more:

1. https://github.com/searchivarius/nmslib
2. https://github.com/DBWangGroupUNSW/nns_benchmark
   Thank you!

kbriggs:/Downloads/efanna-master> cd samples/
kbriggs:/Downloads/efanna-master/samples> ./efanna_index_buildgraph sift_base.fvecs sift.graph 8 8 8 30 25 10 10
open file error

If I build index with not AVX distance, aligned_cols in matrix constructor will be recalculated to be divided by 8 anyway, that leads to segmentation faults in meanSplit() when k ~= points_num (end of array), because points_num stays the same, when dim was changed and algorithm wants more points then we have

When I run EFANNA on Yale face B dataset (2174 dimensions), The accuracy was about 90%.
However, if I pick 2 random dimensional from the same dataset, the accuracy reduces to 0.5%.

I also didn't work on some toy datasets that I generated in 2D (e.g. swiss roll data).
Is there any bugs in the algorithm?