search

RyanLiGod / hnsw-python

hnsw implemented by python
62 stars 14 forks source link

readme

hnsw-python

HNSW implemented by python.

Supported distances:

Distance parameter Equation
Squared L2 'l2' d = sum((Ai-Bi)^2)
Cosine similarity 'cosine' d = 1.0 - sum(Ai*Bi) / sqrt(sum(Ai*Ai) * sum(Bi*Bi))

examples

import time
from progressbar import *
import pickle
from hnsw import HNSW

dim = 200
num_elements = 10000

data = np.array(np.float32(np.random.random((num_elements, dim))))
hnsw = HNSW('cosine', m0=16, ef=128)
widgets = ['Progress: ',Percentage(), ' ', Bar('#'),' ', Timer(), ' ', ETA()]

# show progressbar
pbar = ProgressBar(widgets=widgets, maxval=train_len).start()
for i in range(len(data)):
    hnsw.add(data[i])
    pbar.update(i + 1)
pbar.finish()

# save index
with open('glove.ind', 'wb') as f:
    picklestring = pickle.dump(hnsw, f, pickle.HIGHEST_PROTOCOL)

# load index
fr = open('glove.ind','rb')
hnsw_n = pickle.load(fr)

add_point_time = time.time()
idx = hnsw_n.search(np.float32(np.random.random((1, 200))), 10)
search_time = time.time()
print("Searchtime: %f" % (search_time - add_point_time))