annoy-java
annoy-java copied to clipboard
spotify
Approximate nearest neighbors in Java
This Java implementation defines ANGULAR distance as "cosine similarity", while the C++/python implementation defines ANGULAR distance as "approximate theta". Evidence: Java implementation -- dot(u,v), the value of cosine https://github.com/spotify/annoy-java/blob/aef275bd5b872b0931bde557bf8627c0476246af/src/main/java/com/spotify/annoy/ANNIndex.java#L189 https://github.com/spotify/annoy-java/blob/aef275bd5b872b0931bde557bf8627c0476246af/src/main/java/com/spotify/annoy/ANNIndex.java#L319...
I did some testing comparing the results from this library and the C++ library and they are consistently different. I needed to get it to work so I went through...
The Java bindings for `annoy` currently only support the `ANGULAR` and `EUCLIDEAN` metrics. It would be ideal if support was also added for `MANHATTAN`, `HAMMING` and `DOT` metrics, for which...
Hello, I have faced the issue: 1. index is created using default metric in python. 2. index is read in java with Euclidean metric without error. On search raised following...
Sorry not sure where to put this, but I thought I should mention I wrote a C#/.NET port of this as part of my fork of Word2vec.Tools. Pros / features:...
Because the API for MappedByteBuffer, or whatever it's called, uses Java int for offsets, you cannot address past 2GB. To work around this, annoy-java would have to open multiple windows...
Something that was very useful in the C++ version was to let the user manually provide a search_k limit. Cf https://github.com/spotify/annoy/blob/master/src/annoylib.h#L541 https://github.com/spotify/annoy-java/blob/master/src/main/java/com/spotify/annoy/ANNIndex.java#L147 That way you can ask for the 10...
