2 Released

  1. Added level 2 pruning and standalone pruning program "prune". Level 1 and level 2 pruning are independent, and can be used together to make index even smaller. Level-2 pruning also improves the performance for most datasets (need to specify a larger M than default). Pruning can be also achieved by specifying "--prune 1", "--prune 2" or "--prune 3" (1+2) when constructing the graph.
  2. Added search parameters M and T.
    • M: default is 0. If M is bigger than the inferred M of a point, the given M will be used, to the best when neighbors are available.
    • T: run this number of independent search processes and merge the results. Default is 1 to be backward compatible. If a larger T is specified, P can be reduced to achieve the same recall.