Golly の Generations で作った音のクラスタリング

$ 7z x golly_generations_cluster.7z

import numpy
import scipy.signal
import sklearn.cluster
import shutil
import soundfile
from pathlib import Path

def extract_feature(path, n_frame=39):
    """
    スペクトログラムは2次元のデータ。
    sklearn.cluster で使えるように numpy.ravel で1次元にして返す。

    spectrogram.shape = (n_freq, n_frame)

    n_frame のデフォルト値はテストに使ったデータセットを調べて決めた。
    """
    data, samplerate = soundfile.read(str(path))
    frequency, time, spectrogram = scipy.signal.spectrogram(data, samplerate)

    if spectrogram.shape[1] < n_frame:
        zeros = numpy.zeros((spectrogram.shape[0],
                             n_frame - spectrogram.shape[1]))
        spectrogram = numpy.concatenate((spectrogram, zeros), axis=1)
    elif spectrogram.shape[1] > n_frame:
        spectrogram = spectrogram[:][0:n_frame]

    return numpy.ravel(numpy.transpose(spectrogram))

def write_result(output_directory, n_clusters, labels, filepath):
    if output_directory.exists():
        shutil.rmtree(output_directory)

    digits = len(str(abs(n_clusters - 1)))
    output_directories = [
        output_directory / Path(f"{index:0{digits}d}")
        for index in range(n_clusters)
    ]

    for directory in output_directories:
        directory.mkdir(parents=True, exist_ok=True)

    for label, path in zip(labels, filepath):
        shutil.copy(path, output_directories[label])

if __name__ == "__main__":
    directory_path = Path("generations")
    if not directory_path.is_dir():
        print("Invalid path.")

    filepath = [path for path in directory_path.glob("*.wav")]
    features = numpy.array([extract_feature(path) for path in filepath])

    n_clusters = 40
    cluster = sklearn.cluster.KMeans(n_clusters=n_clusters).fit(features)

    write_result(
        Path("cluster_spectrogram"), n_clusters, cluster.labels_, filepath)

    numpy.save("data/filepath.npy", filepath)
    numpy.save("data/features.npy", features)
    numpy.save("data/features_shape.npy", (39, 129))
    numpy.save("data/labels.npy", cluster.labels_)
    numpy.save("data/centers.npy", cluster.cluster_centers_)

import python_speech_features

def extract_feature(path, n_frame=19):
    """
    mfcc.shape = (n_frame, n_cepstrum)
    """
    data, samplerate = soundfile.read(str(path))
    nfft = 1024
    mfcc = python_speech_features.mfcc(
        data,
        samplerate,
        winlen=nfft / samplerate,
        winstep=0.01,
        numcep=26,
        nfilt=52,
        nfft=nfft,
        preemph=0.97,
        ceplifter=22,
    )

    if mfcc.shape[0] < n_frame:
        zeros = numpy.zeros((n_frame - mfcc.shape[0], mfcc.shape[1]))
        mfcc = numpy.concatenate((mfcc, zeros), axis=0)
    elif mfcc.shape[0] > n_frame:
        mfcc = mfcc[0:n_frame]

    return numpy.ravel(mfcc)

import matplotlib.pyplot as pyplot
import numpy
import python_speech_features
import sklearn.cluster
import soundfile
from pathlib import Path

def get_inertia(n_clusters, features):
    cluster = sklearn.cluster.KMeans(n_clusters=n_clusters).fit(features)
    return cluster.inertia_

if __name__ == "__main__":
    features = numpy.load("features.npy")

    x_range = (2, 101)
    errors = [get_inertia(k, features) for k in range(*x_range)]
    k_value = [k for k in range(*x_range)]

    pyplot.plot(k_value, errors, lw=1, color="gray")
    pyplot.plot(k_value, errors, "o", markersize=3, color="black")
    pyplot.xlabel("n_clusters")
    pyplot.ylabel("error")
    pyplot.xlim((x_range[0], x_range[1] - 1))
    pyplot.grid()
    pyplot.show()

import numpy
import sklearn.manifold

features = numpy.load("data/features.npy")
tsne = sklearn.manifold.TSNE(n_components=2).fit(features)
numpy.save("data/embedding.npy", tsne.embedding_)