searching

Searching

The main functionality of deglib is to search for similar feature vectors.

import numpy as np
import deglib

# build the graph ...

# generate query
query = np.random.random(DIMS).astype(np.float32)

# search query
indices, distances = graph.search(query, eps=0.1, k=16)

print(indices)    # data[result_indices] will show the 16 closest datapoints to "query"
print(distances)  # numpy array with 16 distances to the results

This will give the 16 indices of the datapoints closest to the query as well as their distances to the query.

Multiple queries at once

It is also possible to search for multiple queries at once:

N_QUERIES = 10
query = np.random.random(N_QUERIES, DIMS).astype(np.float32)

indices, distances = graph.search(query, eps=0.1, k=16)

In this case indices and distances will have shape (10, 16) (16 indices for 10 queries).

Parameters

There are a lot of parameters to control the search speed and recall. See SearchGraph for more details.

Filters

Sometimes you want to exclude some samples of your dataset from search. This can be done using filters:

valid_labels = np.array([0, 3, 34, 52, 108])
query = np.random.random(dims).astype(np.float32)

results, distances = graph.search(query, filter_labels=valid_labels, eps=0.01, k=4)

In this case results will only include values in valid_labels.