def from_medleydb(_session, limit=None):
"""Import data from medleydb
Parameters
----------
session : SQLAlchemy session
The session
"""
import itertools
with open(medleydb.INST_TAXONOMY) as f_handle:
data = yaml.load(f_handle)
walk_taxonomy(data, _session)
_session.commit()
if limit is None:
entries = medleydb.load_all_multitracks()
else:
entries = itertools.islice(medleydb.load_all_multitracks(), limit)
for mt in entries:
data = model.Track.from_medleydb(mt, _session)
_session.add(data)
_session.commit()
def get_tracks(self, min_sources=10, instruments=None, rm_silence=False, trim=None, count=None, **kwargs):
if instruments:
valid_instruments = instruments
else:
valid_instruments = self.get_valid_instruments(min_sources)
logging.info('Valid instruments: ' + str(valid_instruments) + '\n')
counts = defaultdict(lambda: 0)
multitrack_list = mdb.load_all_multitracks()
for track in multitrack_list:
if not track.has_bleed:
for stem in track.stems:
if stem.instrument in valid_instruments:
if count and counts[stem.instrument] >= count:
continue
counts[stem.instrument] += 1
if rm_silence or trim:
ext = os.path.splitext(stem.file_path)[1]
dest = tempfile.mktemp(suffix=ext, dir=self.tempdir)
logging.info('Processing (sox): ' + str(stem.file_path))
sox_args = ['sox', stem.file_path, dest]
if rm_silence:
sox_args.extend(['silence', '1', '0.1', '0.1%', '-1', '0.1', '0.1%'])
if trim:
sox_args.extend(['trim', '0', str(trim)])
subprocess.Popen(sox_args, stderr=subprocess.PIPE)
stem.file_path = dest
stem.label = stem.instrument
yield self.track_to_dict(stem)
logging.info(counts)
def main(args):
mtracks = mdb.load_all_multitracks()
mix_coefs = dict()
for mtrack in mtracks:
mix_coefs[mtrack.track_id] = analyze_mix(mtrack)
with open(args.output_path, 'w') as fdesc:
yaml.dump(mix_coefs, fdesc)
def get_tracks(self, min_sources=10, instruments=None, rm_silence=False, trim=None, count=None, **kwargs):
import medleydb as mdb
if instruments:
valid_instruments = instruments
else:
valid_instruments = self.get_valid_instruments(min_sources)
logging.info('Valid instruments: ' + str(valid_instruments) + '\n')
counts = defaultdict(lambda: 0)
multitrack_list = mdb.load_all_multitracks()
for track in multitrack_list:
if not track.has_bleed:
for stem in track.stems:
if stem.instrument in valid_instruments:
if count and counts[stem.instrument] >= count:
continue
counts[stem.instrument] += 1
if rm_silence or trim:
ext = os.path.splitext(stem.file_path)[1]
dest = tempfile.mktemp(suffix=ext, dir=self.tempdir)
logging.info('Processing (sox): ' + str(stem.file_path))
sox_args = ['sox', stem.file_path, dest]
if rm_silence:
sox_args.extend(['silence', '1', '0.1', '0.1%', '-1', '0.1', '0.1%'])
if trim:
sox_args.extend(['trim', '0', str(trim)])
subprocess.Popen(sox_args, stderr=subprocess.PIPE)
stem.file_path = dest
stem.label = stem.instrument
yield self.track_to_dict(stem)
logging.info(counts)
def get_tracks(self):
multitrack_list = mdb.load_all_multitracks()
for track in multitrack_list:
if track.genre:
track.file_path = track.mix_path
track.label = track.genre
yield track
def get_valid_genres(self, min_sources):
"""Get set of genres with at least min_sources different sources"""
logging.info('Determining valid genres...\n')
multitrack_list = mdb.load_all_multitracks()
genre_counts = defaultdict(lambda: 0)
for track in multitrack_list:
if track.genre:
genre_counts[track.genre] += 1
logging.info(genre_counts)
return {g for g in genre_counts if genre_counts[g] >= min_sources}
def get_valid_genres(self, min_sources):
"""Get set of genres with at least min_sources different sources"""
import medleydb as mdb
logging.info('Determining valid genres...\n')
multitrack_list = mdb.load_all_multitracks()
genre_counts = defaultdict(lambda: 0)
for track in multitrack_list:
if track.genre:
genre_counts[track.genre] += 1
logging.info(genre_counts)
return {g for g in genre_counts if genre_counts[g] >= min_sources}
def main(args):
if args.use_mdb2:
dataset_version = ['V1', 'V2', 'EXTRA']
else:
dataset_version = ['V1']
mtracks = mdb.load_all_multitracks(dataset_version=dataset_version)
Parallel(n_jobs=args.n_jobs, verbose=5)(delayed(compute_features_mtrack)(
mtrack, args.save_dir, args.option, args.precomputed_hcqt_path)
for mtrack in mtracks)
print("Done!")
def get_valid_instruments(self, min_sources):
"""Get set of instruments with at least min_sources different sources"""
logging.info('Determining valid instruments...\n')
multitrack_list = mdb.load_all_multitracks()
instrument_counts = defaultdict(lambda: 0)
for track in multitrack_list:
if not track.has_bleed:
instruments = set()
for stem in track.stems:
instruments.add(stem.instrument)
for instrument in instruments:
instrument_counts[instrument] += 1
logging.info(instrument_counts)
return {i for i in instrument_counts if instrument_counts[i] >= min_sources}
def get_valid_instruments(self, min_sources):
"""Get set of instruments with at least min_sources different sources"""
import medleydb as mdb
logging.info('Determining valid instruments...\n')
multitrack_list = mdb.load_all_multitracks()
instrument_counts = defaultdict(lambda: 0)
for track in multitrack_list:
if not track.has_bleed:
instruments = set()
for stem in track.stems:
instruments.add(stem.instrument)
for instrument in instruments:
instrument_counts[instrument] += 1
logging.info(instrument_counts)
return {i for i in instrument_counts if instrument_counts[i] >= min_sources}
def main(args):
mtracks = mdb.load_all_multitracks(dataset_version=['V1', 'V2', 'EXTRA', 'BACH10'])
mix_coefs = dict()
for mtrack in mtracks:
print(mtrack.track_id)
# compute mixing weights on both stft and squared audio
coeffs_stft = analyze_mix_stft(mtrack)
coeffs_audio = analyze_mix_audio(mtrack)
mix_coefs[mtrack.track_id] = {'stft': coeffs_stft, 'audio': coeffs_audio}
print(mix_coefs[mtrack.track_id])
print("")
with open(args.output_path, 'w') as fdesc:
yaml.dump(mix_coefs, fdesc)
def main(args):
mtracks = mdb.load_all_multitracks(dataset_version=['V1', 'V2', 'EXTRA'])
for mtrack in mtracks:
print("Processing {}...".format(mtrack.track_id))
if os.path.exists(
os.path.join(args.save_dir, "{}_training_pairs.json".format(
mtrack.track_id))):
print(" already done!")
continue
json_path = get_all_audio_annot_pairs(mtrack, args.save_dir,
args.resynth_path,
args.replace_path)
print("...saved to {}".format(json_path))
print("")
print("done!")
def get_tracks(self, min_sources=10, genres=None, count=None, **kwargs):
if genres:
valid_genres = genres
else:
valid_genres = self.get_valid_genres(min_sources)
logging.info('Valid genres: ' + str(valid_genres) + '\n')
counts = defaultdict(lambda: 0)
multitrack_list = mdb.load_all_multitracks()
for track in multitrack_list:
if track.genre and track.genre in valid_genres:
if count and counts[track.genre] >= count:
continue
counts[track.genre] += 1
track.file_path = track.mix_path
track.label = track.genre
yield self.track_to_dict(track)
logging.info(counts)
def get_tracks(self, min_sources=10, genres=None, count=None, **kwargs):
import medleydb as mdb
if genres:
valid_genres = genres
else:
valid_genres = self.get_valid_genres(min_sources)
logging.info('Valid genres: ' + str(valid_genres) + '\n')
counts = defaultdict(lambda: 0)
multitrack_list = mdb.load_all_multitracks()
for track in multitrack_list:
if track.genre and track.genre in valid_genres:
if count and counts[track.genre] >= count:
continue
counts[track.genre] += 1
track.file_path = track.mix_path
track.label = track.genre
yield self.track_to_dict(track)
logging.info(counts)
def main(args):
mtracks = mdb.load_all_multitracks(
dataset_version=['V1', 'V2', 'EXTRA', 'BACH10'])
mix_coefs = dict()
for mtrack in mtracks:
print(mtrack.track_id)
# compute mixing weights on both stft and squared audio
coeffs_stft = analyze_mix_stft(mtrack)
coeffs_audio = analyze_mix_audio(mtrack)
mix_coefs[mtrack.track_id] = {
'stft': coeffs_stft,
'audio': coeffs_audio
}
print(mix_coefs[mtrack.track_id])
print("")
with open(args.output_path, 'w') as fdesc:
yaml.dump(mix_coefs, fdesc)
import utils
import medleydb as mdb
import json
# Load all multitracks
mtrack_generator = mdb.load_all_multitracks()
all_tracks_id = [mtrack.track_id for mtrack in mtrack_generator]
# Create a dictionary with all vocal pieces and its sources
mtrack_generator = mdb.load_all_multitracks()
soma = 0
lista = []
dicionario = {}
for mtrack in mtrack_generator:
if not mtrack.is_instrumental:
soma += 1
lista.append(mtrack.track_id)
dicionario[mtrack.track_id] = mtrack.stem_instruments
# Save dictionary to a json file
with open('vocal_pieces.json', 'w') as json_file:
json.dump(dicionario, json_file)
print(" >> Generated vocal_pieces.json!")
def test_all():
assert next(mdb.load_all_multitracks())
nargs='*',
default=None,
help='List of instruments to extract')
parser.add_argument('-k',
'--keep_silence',
action='store_true',
help="Don't remove silence from audio files")
args = parser.parse_args()
if args.instruments:
valid_instruments = args.instruments
else:
valid_instruments = get_valid_instruments(args.min_sources)
logging.info('Valid instruments: ' + str(valid_instruments) + '\n')
multitrack_list = mdb.load_all_multitracks()
for track in multitrack_list:
if not track.has_bleed:
for stem in track.stems:
if stem.instrument in valid_instruments:
filename = os.path.basename(stem.file_path)
dest = os.path.join(args.destination, stem.instrument,
filename)
# Create directory if necessary
try:
os.makedirs(os.path.dirname(dest))
except OSError:
pass
if args.keep_silence:
def compute_feature(n_fft, hop_length, genres, chunk, rsr, unvoice, melody, bass,
n_random, n_val, get_attr):
# init
genres = genres.split(",")
t_slice = int(chunk * rsr)
tracks = mdb.load_all_multitracks()
instruments_dict = {}
train_set = {}
val_set = {}
tracks_meta = [[] for i in range(len(genres))]
idx_cnt = 0
print("Reading the metadata and build dictionary for instruments ...")
for t in tracks:
if t.genre in genres:
# get the mix audio as well as build a table for instrumental attributes
metadata = get_metadata(t, unvoice, melody, bass)
if get_attr:
for instrument in metadata["instruments"]:
if instrument not in instruments_dict.keys():
instruments_dict[instrument] = idx_cnt
idx_cnt += 1
tracks_meta[get_gid(t.genre, genres)].append(metadata)
print("Done! Start generating features ...")
target_path = "mix_path" if not unvoice else "unvoice_path"
# generate data for each genre
for gid, g in enumerate(genres):
start = time.time()
audio_chunks = []
instruments_attrs = []
print("Generating features for {} ...".format(g))
file_path = "{}.hdf5".format(g)
for md in tracks_meta[gid]:
print("Converting {} ...".format(md[target_path]))
audio = (md[target_path],)
audio += (md["melody_path"],) if melody else () # append the melody part if needed.
audio += (md["bass_path"],) if bass else ()
# get the atual mix, and turn them into stft.
chunks = get_mix_chunks(audio, t_slice, n_fft, hop_length, n_random[gid], rsr)
audio_chunks.extend(chunks)
print("{} clips for {}.".format(len(chunks), md[target_path]))
if get_attr:
attrs = get_instruments_attrs(md["instruments"], instruments_dict, len(chunks))
instruments_attrs.extend(attrs)
# get random indices
print("We have {} clips for {}.".format(len(audio_chunks), g))
idx = np.linspace(0, len(audio_chunks)-1, len(audio_chunks), dtype=int)
np.random.shuffle(idx)
audio_chunks = np.asarray(audio_chunks, dtype=np.float32)
print("shape: {}, type: {}".format(audio_chunks.shape, audio_chunks.dtype))
if audio_chunks.shape[1] == 1:
audio_chunks = np.squeeze(audio_chunks, axis=1)
print(audio_chunks.shape)
audio_chunks = (audio_chunks - audio_chunks.mean()) / audio_chunks.std()
train_set["audio"] = audio_chunks[idx][n_val:, ...]
val_set["audio"] = audio_chunks[idx][:n_val, ...]
print(train_set["audio"].shape)
print(val_set["audio"].shape)
if get_attr:
train_set["attrs"] = np.asarray(instruments_attrs)[idx][n_val:, ...]
val_set["attrs"] = np.asarray(instruments_attrs)[idx][:n_val, ...]
print("Saving ...")
for k in train_set.keys():
np.save("./output/{}_{}_val.npy".format(g, k), val_set[k])
np.save("./output/{}_{}_train.npy".format(g, k), train_set[k])
print("Generation for {} is complete, {} sec elapsed.".format(g, time.time()-start))
parser.add_argument('-s', '--min_sources', type=int, default=10,
help='Minimum number of seperate sources in medleyDB '
'for an instrument to be valid')
parser.add_argument('-i', '--instruments', nargs='*', default=None,
help='List of instruments to extract')
parser.add_argument('-k', '--keep_silence', action='store_true',
help="Don't remove silence from audio files")
args = parser.parse_args()
if args.instruments:
valid_instruments = args.instruments
else:
valid_instruments = get_valid_instruments(args.min_sources)
logging.info('Valid instruments: ' + str(valid_instruments) + '\n')
multitrack_list = mdb.load_all_multitracks()
for track in multitrack_list:
if not track.has_bleed:
for stem in track.stems:
if stem.instrument in valid_instruments:
filename = os.path.basename(stem.file_path)
dest = os.path.join(args.destination,
stem.instrument,
filename)
# Create directory if necessary
try:
os.makedirs(os.path.dirname(dest))
except OSError:
pass
