FMA dataset conversion script, genre classification, to individual subdirs

proc_fma.py

#! /usr/env/python3
#
#  FMA conversion script,  genre classification
#  Author: Scott Hawley
#  License: Do as you like
#
# For FMA dataset https://github.com/mdeff/fma
# to be used with panotti https://github.com/drscotthawley/panotti
#
# This will create a directory called Samples/
#   inside which will be directories for each genre, e.g. Rock/ Instrumental/ etc.
#   within these directories will be the audio files converted to .wav format
#
# Requires: fma
#           librosa
#           ffmpeg
#           ...hour(s) to run.  it's slow, runs in serial
import os
import librosa

import utils  # note my fix to fma utils, https://github.com/mdeff/fma/issues/34

AUDIO_DIR = 'fma_small/'
tracks = utils.load('tracks.csv')

os.mkdir('Samples')

small = tracks['set', 'subset'] <= 'small'

y_small = tracks.loc[small, ('track', 'genre_top')]

sr = 44100
for track_id, genre in y_small.iteritems():
    if not os.path.exists('Samples/'+genre):
        os.mkdir('Samples/'+genre)

    mp3_filename = utils.get_audio_path(AUDIO_DIR, track_id)
    out_wav_filename = 'Samples/'+genre+'/'+str(track_id)+'.wav'
    in_wav_filename = out_wav_filename
    cmd = 'ffmpeg -hide_banner -loglevel panic -i ' + mp3_filename + ' ' + in_wav_filename
    print("excuting conversion: "+cmd)
    os.system(cmd)

    ## We could just have ffmpeg do the full conversion, but we'll let librosa
    ## apply its own defaults by reading & writing
    print("reading ",in_wav_filename)
    data, sr = librosa.load(in_wav_filename, sr=sr, mono=True)

    print("writing ",out_wav_filename)
    librosa.output.write_wav(out_wav_filename,data,sr=sr)

hey, Im working with siamese nn for similarity and i need to work with a dataset of pairs positive for same music different exemples and negative different music i couldnt find any but i thought i could make one by fma small i started by augmented each music and make pos pair with itself and it didnt work well, im looking for an idea for this. could you help? please.