constants.py

# 生成したファイルの出力先ディレクトリ
OUTPUT_DIR = 'output'

# 学習データとして使う音色
INSTS_FOR_TRAINING = [
    # Piano
    1, 3, 5, 7, 8,
    # Chromatic Percussion
    9, 11, 13, 15,
    # Organ
    17, 19, 21, 23, 24,
    # Guitar
    25, 27, 29, 31,
    # Bass
    33, 35, 37,
    # Strings
    41, 43, 45, 47,
    # Ensemble
    49, 51, 53, 55,
    # Brass
    57, 59, 60, 61, 63,
    # Reed
    65, 66, 68, 70, 72,
    # Pipe
    73, 75, 78, 80,
    # Synth Lead
    81, 85,
    # Ethnic
    105, 107, 110, 112
]

# テストデータとして使う音色
INSTS_FOR_EVALUATING = [
    # Piano
    2, 4, 6,
    # Chromatic Percussion
    10, 12, 16,
    # Organ
    18, 20, 22,
    # Guitar
    26, 28, 30,
    # Bass
    34, 36, 38,
    # Strings
    42, 44,
    # Ensemble
    50, 52,
    # Brass
    58, 62, 64,
    # Reed
    67, 69, 71,
    # Pipe
    74, 76, 79,
    # Synth Lead
    82, 88,
    # Ethnic
    106, 108, 111
]

# C3 ～ B6 の範囲
KEYNAMES = ['C', 'D', 'E', 'F', 'G', 'A', 'B']
OCTAVE_RANGE = range(3, 7)

csvgen.py

from concurrent import futures
import csv
import os
from pathlib import Path
import typing
import numpy
import scipy
from scipy.io import wavfile
from scipy.fftpack import fft
from constants import *

START_FRAME = 110250 # 2.5s
FFTLEN = 1024
WINDOW = scipy.hamming(FFTLEN)

def _iter_notes() -> typing.Iterator[str]:
    for octave in OCTAVE_RANGE:
        for key in KEYNAMES:
            yield key + str(octave)

NOTES = list(_iter_notes())

def create_ft_csvfile(wavfile_path: str) -> None:
    rate, wav_data = wavfile.read(wavfile_path)
    assert rate == 44100

    # 2.5s から 1024 サンプルを取り出してフーリエ変換
    fftinput = wav_data[START_FRAME : START_FRAME + FFTLEN] * WINDOW
    fftresult = fft(fftinput)

    # 対数パワースペクトルにする
    ps = numpy.log10(numpy.absolute(fftresult[:FFTLEN // 2]))
    ps = ps / numpy.max(ps)

    dataname = Path(wavfile_path).stem
    csvpath = os.path.join(OUTPUT_DIR, dataname + '.csv')
    with open(csvpath, 'w', encoding='utf_8', newline='\r\n') as f:
        for x in ps: f.write('{:.10f}\n'.format(x))

    print(dataname)

def create_dataset_csvfile(insts: typing.Sequence[int], filename: str) -> None:
    with open(os.path.join(OUTPUT_DIR, filename), 'w', encoding='utf_8', newline='') as f:
        writer = csv.writer(f)
        writer.writerow(['x:data', 'y:label'])

        for inst in insts:
            for octave in OCTAVE_RANGE:
                for key_index in range(len(KEYNAMES)):
                    datapath = './{:03d}_{}{}.csv'.format(inst, KEYNAMES[key_index], octave)
                    writer.writerow([datapath, key_index])

def main() -> None:
    with futures.ThreadPoolExecutor() as executor:
        # wav ファイルをフーリエ変換した csv ファイルを作成
        for entry in os.scandir(OUTPUT_DIR):
            if entry.is_file() and entry.name.endswith('.wav'):
                executor.submit(create_ft_csvfile, entry.path)

        # 学習データ
        create_dataset_csvfile(INSTS_FOR_TRAINING, 'training.csv')
        # テストデータ
        create_dataset_csvfile(INSTS_FOR_EVALUATING, 'evaluating.csv')

if __name__ == '__main__':
    main()

wavgen.py

from concurrent import futures
import os
import subprocess
import pretty_midi
from constants import *

INSTS = INSTS_FOR_TRAINING.copy()
INSTS.extend(INSTS_FOR_EVALUATING)

os.makedirs(OUTPUT_DIR, exist_ok=True)

def create_wavfile(inst_num: int, octave: int, key: str) -> None:
    note_name = key + str(octave)
    note_num = pretty_midi.note_name_to_number(note_name)
    inst_disp = '{:03d}'.format(inst_num)
    path_prefix = os.path.join(OUTPUT_DIR, inst_disp + '_' + note_name)

    # MIDI の作成
    inst = pretty_midi.Instrument(inst_num)
    # timidity の出力の最初にノイズが含まれてしまうので、最初に 1 音入れておく
    inst.notes.append(pretty_midi.Note(100, note_num, 0, 1))
    inst.notes.append(pretty_midi.Note(100, note_num, 2, 4))

    midi_file = pretty_midi.PrettyMIDI()
    midi_file.instruments.append(inst)
    midi_file.write(path_prefix + '.mid')

    # WAV へ変換
    subprocess.run(
        ['timidity', '-OwM', '-o', path_prefix + '.wav', path_prefix + '.mid'],
        stdout=subprocess.DEVNULL, check=True)

    print(inst_disp + '\t' + note_name)

def main() -> None:
    # それなりに時間がかかるので並列化
    with futures.ThreadPoolExecutor() as executor:
        for inst_num in INSTS:
            for octave in OCTAVE_RANGE:
                for key in KEYNAMES:
                    executor.submit(create_wavfile, inst_num, octave, key)

if __name__ == "__main__":
    main()