README.md

WIP

Instalation

require openblas, openmp and cmake. conda install openblas openmp cmake

then install with: pip install . or pip install git+https://github.com/Museum7432/ftsearch.git

Usage

it is similar to Faiss, but designed more for searching on videos (a sequence of encoded vectors of keyframes with OpenClip for example).

import ftsearch
import numpy as np

# 512 is vector dimension
ft = ftsearch.FTSearch(512)

for i in range(100):
    # a video with 1000 keyframes
    # this number can be arbitrary
    data = np.random.random((1000, 512)).astype("float32")
    
    # input the sequence along with the video name (might change later)
    ft.add_seq(data, f"vid_{i}")


query = data[[1,3,4]]

# batch query search
# similar to Faiss search
# it only supports dot product for now
re = ft.search(query, 2)

# the score
print(re[0])
# the indices of the scores
print(re[1])


# sequental search (or temporal search)
# it find the best subset of vectors within each sequence
# that best match the input vector
# min_item_dist is the minimum distance in indice between two selected vectors (consecutively)
# discount_rate: penalize large gap between two selected vectors
re = ft.seq_search(query, topk=1, min_item_dist=1, discount_rate=0.9)

print(re[0])
print(re[1])

# you can also get infos from the returned indices
print(ft.get_info(23123))