Hanny

gem 'hanny'

require 'hanny'

# Prepare vector data for search targets and queries with Numo::DFloat (shape: [n_samples, n_features]).
targets = Numo::DFloat.new(5000, 512).rand
queries = Numo::DFloat.new(10, 512).rand

# Build a search index with 256-bit binary code.
index = Hanny::LSHIndex.new(code_length: 256)
index.build_index(targets)

# Obtain the Array<Integer> that has the data indices of 10-nearest neighbors for each query.
candidates = index.search_knn(queries, n_neighbors: 10)

# Obtain the Array<Integer> that has the data indices whithin Hamming radius of 4 for each query.
candidates = index.search_radius(queries, radius: 4)

# Calculate pairwise euclidean distances between the query and its neighbors.
query_id = 0
distances = Hanny::Utils.euclidean_distance(queries[query_id, true], targets[candidates[query_id], true])

# Add new data to the search index.
appended_data_ids = index.append_data(new_data)

# Remove the data from the search index.
removed_data_ids = index.remove_data([0, 1, 2])

# Save and load the search index with Marshal.
File.open('index.dat', 'wb') { |f| f.write(Marshal.dump(index)) }
index = Marshal.load(File.binread('index.dat'))

require 'benchmark'
require 'rumale'
require 'hanny'

# Load MNIST data set.
samples, labels = Rumale::Dataset.load_libsvm_file('mnist')
queries = samples[0..5, true]
targets = samples[6..-1, true]
qlabels = labels[0..5]
tlabels = labels[6..-1]

# Build LSH search index.
index = Hanny::LSHIndex.new(code_length: 128, random_seed: 1)
index.build_index(targets)

# Run a benchmark test for finding 5-nearest neighbors.
n_queries = queries.shape[0]
n_neighbors = 5
Benchmark.bm 50 do |r|
  r.report 'LSH' do
    candidates = index.search_knn(queries, n_neighbors: n_neighbors)
    n_queries.times do |m|
      STDERR.write("\nquery label: %d, neighbors label: " % qlabels[m])
      candidates[m].each { |n| STDERR.write("%d, " % tlabels[n]) }
    end
    STDERR.write("\n")
  end
  r.report 'Brute-force' do
    distance_mat = Hanny::Utils.euclidean_distance(queries, targets)
    candidates = Array.new(n_queries) do |n|
      distance_mat[n, true].to_a.map.with_index.sort_by(&:first).map(&:last)[0...n_neighbors]
    end
    n_queries.times do |m|
      STDERR.write("\nquery label: %d, neighbors label: " % qlabels[m])
      candidates[m].each { |n| STDERR.write("%d, " % tlabels[n]) }
    end
    STDERR.write("\n")
  end
end