JVSデータでの話者re-identifierの精度評価

analyze.py

from pathlib import Path

import numpy
from scipy.spatial.distance import cdist


def load_vector(path_format: Path):
    predicted_vectors = []
    true_speaker_nums = []
    for speaker_num in range(100):
        if speaker_num in (5, 27):
            continue
        for i in range(5):
            true_speaker_nums.append(speaker_num)
            predicted_vectors.append(numpy.load(str(path_format).format(speaker_num=speaker_num, i=i)))

    return (
        numpy.asarray(predicted_vectors),
        numpy.asarray(true_speaker_nums),
    )


def main():
    path_format = Path('/path/to/vectors/{speaker_num}-{i}.npy')
    predicted_vectors, true_speaker_nums = load_vector(path_format=path_format)

    # top1
    distance = cdist(predicted_vectors, predicted_vectors)
    top1_index = numpy.argsort(distance, axis=1)[:, 1]
    top1_speaker_nums = true_speaker_nums[top1_index]

    # result
    rate = (top1_speaker_nums == true_speaker_nums).mean()
    print('rate', rate)

    breakpoint()


if __name__ == '__main__':
    main()