def __init__(self,
roots,
sample_rate,
max_time,
target_level=-25,
noise_proportion=0,
noise_type='gaussian',
snrs=[3],
eps=1e-8,
**kwargs):
self.sample_rate = sample_rate
self.max_time = max_time
self.target_level = target_level
self.eps = eps
self.filepths = []
for root in roots:
self.filepths += find_files(root)
assert len(self.filepths) > 0, 'No audio file detected'
self.noise_proportion = noise_proportion
self.snrs = snrs
if noise_type == 'gaussian':
self.noise_sampler = torch.distributions.Normal(0, 1)
else:
self.noise_wavpths = find_files(noise_type)
def SaveSTFT_Arg(pitch_shift, time_stretch, argtime):
targetlist = find_files(C.target_path, ext="wav")
noiselist = find_files(C.noise_path, ext="wav")
noise_num = len(noiselist)
target_index = 0
for targetfile in tqdm(targetlist):
target_mag, _ = LoadAudio_Arg(targetfile)
norm = target_mag.max()
skip_count = 0
if target_mag.shape[0] > C.PATCH_LENGTH:
step = target_mag.shape[0] // C.PATCH_LENGTH
for i in tqdm(range(step), leave=False):
target_mag_p = target_mag[:, i * C.PATCH_LENGTH:(i + 1) *
C.PATCH_LENGTH]
target_mag_p /= norm
noise_file = noiselist[random.randint(0, noise_num - 1)]
noise_mag, _ = LoadAudio(noise_file)
noise_mag = LengthAdjuster(noise_mag)
noise_mag /= norm
addnoise_mag = target_mag_p + noise_mag
addnoise_mag /= norm
fname = str(target_index) + "_" + str(i) + str(argtime)
np.savez(os.path.join(C.PATH_FFT, fname + "_arg.npz"),
speech=target_mag_p,
addnoise=addnoise_mag)
else:
skip_count += 1
print("SKIP:", skip_count)
def get_ann_audio(directory):
'''Get a list of annotations and audio files from a directory.
This also validates that the lengths match and are paired properly.
Parameters
----------
directory : str
The directory to search
Returns
-------
pairs : list of tuples (audio_file, annotation_file)
'''
audio = find_files(directory)
annos = find_files(directory, ext=['jams', 'jamz'])
paired = list(zip(audio, annos))
if len(audio) != len(annos) or any(
[base(aud) != base(ann) for aud, ann in paired]):
raise DataError(
'Unmatched audio/annotation data in {}'.format(directory))
return paired
def sample_wavs_and_dump_txt(root,dev_ids, numbers, meta_data_name):
wav_list = []
count_positive = 0
for _ in range(numbers):
prob = random.random()
if (prob > 0.5):
dev_id_pair = random.sample(dev_ids, 2)
# sample 2 wavs from different speaker
sample1 = "/".join(random.choice(find_files(os.path.join(root,dev_id_pair[0]))).split("/")[-3:])
sample2 = "/".join(random.choice(find_files(os.path.join(root,dev_id_pair[1]))).split("/")[-3:])
label = "0"
wav_list.append(" ".join([label, sample1, sample2]))
else:
dev_id_pair = random.sample(dev_ids, 1)
# sample 2 wavs from same speaker
sample1 = "/".join(random.choice(find_files(os.path.join(root,dev_id_pair[0]))).split("/")[-3:])
sample2 = "/".join(random.choice(find_files(os.path.join(root,dev_id_pair[0]))).split("/")[-3:])
label = "1"
count_positive +=1
wav_list.append(" ".join([label, sample1, sample2]))
f = open(meta_data_name,"w")
for data in wav_list:
f.write(data+"\n")
f.close()
return wav_list
def makeCQTexOvertone():
TORIDASUBASHO = PATH_pwCQT
HOZONBASYO = PATH_pwov2CQT # n倍音を保存するディレクトリ
OVTONE = 2
audiolist = find_files(PATH_AUDIO, ext="wav")
cqtlist = find_files(TORIDASUBASHO, ext="npy")
itemlist = len(audiolist)
i_counter = 1
for audiofile, cqtfile in zip(audiolist, cqtlist):
print("{}/{}".format(i_counter, itemlist))
filename = audiofile.split('/')[-1]
albname = audiofile.split('/')[-2]
# foldname = audiofile.split('/')[-3]
cqt_filename = cqtfile.split("/")[-1]
if (filename.split(".")[0] != cqt_filename.split(".")[0]):
print("file_not_match", filename, cqt_filename)
# ディレクトリチェック
if not (os.path.exists(HOZONBASYO + '/' + albname)):
os.makedirs(HOZONBASYO + '/' + albname)
if not (os.path.exists(HOZONBASYO + '/' + albname + '/' + filename +
'.npy')):
wav, sr = load(audiofile, sr=SR)
cqt = np.load(cqtfile)
excqt = exOvertone(cqt, wav, overtone=OVTONE)
np.save(HOZONBASYO + '/' + albname + '/' + filename + '.npy',
np.array(excqt, dtype="float32"))
i_counter += 1
def TrainConvnetExtractorDeepChroma(trainidx,
epoch=20,
saveas="convnet.model"):
cqtfilelist = np.array(find_files(const.PATH_HCQT, ext="npy"))[trainidx]
labfilelist = np.array(
find_files(const.PATH_CHORDLAB, ext=["lab", "chords"]))[trainidx]
#midifilelist = find_files(const.PATH_MIDI,ext="mid")[:filecnt]
config.train = True
config.enable_backprop = True
convnet = networks.ConvnetFeatExtractor()
model = networks.ConvnetPredictor(convnet)
model.to_gpu(0)
opt = optimizers.MomentumSGD()
opt.setup(model)
print("DeepChroma Convnet Training...")
print("start epochs...")
S = []
T = []
for cqtfile, labfile in zip(cqtfilelist, labfilelist):
cqt = np.load(cqtfile)
spec = utils.PreprocessSpec(cqt[:const.CQT_H, :, :])
targ = voc.LoadChromaTarget(labfile)
minlen = min([cqt.shape[1], targ.shape[0]])
S.append(spec[:, :minlen, :])
T.append(targ[:minlen, :])
S = np.concatenate(S, axis=1)
T = np.concatenate(T, axis=0)
assert (S.shape[1] == T.shape[0])
for ep in range(epoch):
sum_loss = 0
randidx = np.random.randint(0,
S.shape[1] - const.CONV_TRAIN_SEQLEN - 1,
S.shape[1] // const.CONV_TRAIN_SEQLEN * 4)
for i in range(0, randidx.size - const.CONV_TRAIN_BATCH,
const.CONV_TRAIN_BATCH):
x_batch = np.stack([
S[:, randidx[j]:randidx[j] + const.CONV_TRAIN_SEQLEN, :]
for j in range(i, i + const.CONV_TRAIN_BATCH)
])
t_batch = np.stack([
T[randidx[j]:randidx[j] + const.CONV_TRAIN_SEQLEN, :]
for j in range(i, i + const.CONV_TRAIN_BATCH)
])
x_in = cp.asarray(x_batch)
t_in = cp.asarray(t_batch)
model.cleargrads()
loss = model(x_in, t_in)
loss.backward()
opt.update()
sum_loss += loss.data
convnet.save(saveas)
print("epoch: %d/%d loss:%.04f" %
(ep + 1, epoch, sum_loss / const.CONV_TRAIN_BATCH))
convnet.save(saveas)
def EvaluateChord(idx, verbose=True, sonify=False, cross=False):
lablist = np.array(find_files(const.PATH_CHORDLAB, ext=["lab",
"chords"]))[idx]
est_lablist = np.array(find_files(
const.PATH_ESTIMATE_CROSS, ext="lab"))[idx] if cross else find_files(
const.PATH_ESTIMATE, ext="lab")
scorelist_majmin = np.array([])
scorelist_sevenths = np.array([])
scorelist_majmininv = np.array([])
scorelist_seventhinv = np.array([])
durations = np.array([])
confmatrix = np.zeros((const.N_CHORDS, const.N_CHORDS))
song_durations = np.array([])
for labfile, estfile in zip(lablist, est_lablist):
(ref_intervals,
ref_labels) = mir_eval.io.load_labeled_intervals(labfile)
(est_intervals,
est_labels) = mir_eval.io.load_labeled_intervals(estfile)
est_intervals, est_labels = mir_eval.util.adjust_intervals(
est_intervals, est_labels, ref_intervals.min(),
ref_intervals.max(), mir_eval.chord.NO_CHORD,
mir_eval.chord.NO_CHORD)
(intervals, ref_labels,
est_labels) = mir_eval.util.merge_labeled_intervals(
ref_intervals, ref_labels, est_intervals, est_labels)
durations = mir_eval.util.intervals_to_durations(intervals)
comparisons_sevenths = mir_eval.chord.sevenths(ref_labels, est_labels)
comparisons_majmininv = mir_eval.chord.majmin_inv(
ref_labels, est_labels)
comparisons_seventhinv = mir_eval.chord.sevenths_inv(
ref_labels, est_labels)
comparisons_majmin = mir_eval.chord.majmin(ref_labels, est_labels)
score_majmin = mir_eval.chord.weighted_accuracy(
comparisons_majmin, durations)
scorelist_majmin = np.append(scorelist_majmin, score_majmin)
score_sevenths = mir_eval.chord.weighted_accuracy(
comparisons_sevenths, durations)
scorelist_sevenths = np.append(scorelist_sevenths, score_sevenths)
score_majmininv = mir_eval.chord.weighted_accuracy(
comparisons_majmininv, durations)
scorelist_majmininv = np.append(scorelist_majmininv, score_majmininv)
score_seventhinv = mir_eval.chord.weighted_accuracy(
comparisons_seventhinv, durations)
scorelist_seventhinv = np.append(scorelist_seventhinv,
score_seventhinv)
if verbose:
print("%s --- %.3f" % (labfile.split('/')[-1], score_majmin))
for i in range(len(ref_labels)):
confmatrix[voc.GetChordIDSign(ref_labels[i]),
voc.GetChordIDSign(est_labels[i])] += durations[i]
song_durations = np.append(song_durations, np.sum(durations))
return scorelist_majmin, scorelist_sevenths, scorelist_majmininv, scorelist_seventhinv, confmatrix, song_durations
def __init__(self,idx,rand_shift=False):
self.list_labfile = np.array(find_files(C.PATH_CHORDLAB,ext=["lab","chords"]))[idx]
#self.list_cqtfile = np.array(find_files(C.PATH_CQT,ext="npy"))[idx]
self.list_featfile = np.array(find_files(C.PATH_FEAT,ext="npy"))[idx]
#self.list_mfccfile = np.array(find_files(C.PATH_MFCC,ext="npy"))[idx]
#self.list_cqt = [U.normalize_spec(np.load(f)) for f in self.list_cqtfile]
self.list_feat = [np.load(f)[:,:C.N_DIMS_FEAT] for f in self.list_featfile]
self.labs_list,self.lab_intervals_list = LoadLabelSet(self.list_labfile,C.label_shifts[idx])
#self.list_mfcc = [np.load(f).T for f in self.list_mfccfile]
self.rand_shift = rand_shift
def __init__(self, root):
seed = random.randint(1, 1000)
random.seed(seed)
wav_root = Path(root) / "wav48"
wav_files = []
metadata = defaultdict(list)
speaker_dirs = [
speaker_dir for speaker_dir in wav_root.iterdir()
if speaker_dir.is_dir()
]
for speaker_dir in speaker_dirs:
if speaker_dir.stem in self._except_folder:
continue
for wav_file in find_files(speaker_dir):
wav_file = str(PurePosixPath(wav_file).relative_to(root))
wav_files.append(wav_file)
speaker_id = self.get_speaker(wav_file)
metadata[speaker_id].append(wav_file)
self.root = root
self.seed = seed
self.wav_files = wav_files
self.metadata = metadata
def CCMixter():
'''
mix : original wav file
source_1 : inst wav file
source_2 : vocal wac file
'''
Audiolist = os.listdir('./data')
spec_dir = "./Spectrogram"
if os.path.exists(spec_dir) is False:
os.mkdir(spec_dir)
for audio in Audiolist:
try:
audio_path = os.path.join('./data/' + audio)
print("Song : %s" % audio)
if os.path.exists(os.path.join(spec_dir, audio + '.npz')):
print("Already exist!! Skip....")
continue
aud = find_files(audio_path, ext="wav")
mix, _ = load(aud[0], sr=None)
inst, _ = load(aud[1], sr=None)
vocal, _ = load(aud[2], sr=None)
print("Saving...")
SaveSpectrogram(mix, inst, vocal, audio)
except IndexError as e:
print("Wrong Directory")
pass
def EstimateChord(idx, dnnmodel, todir=False):
#dnn = networks.FeatureDNN()
#dnn = networks.ConvnetFeatExtractor()
dnn = networks.FullCNNFeatExtractor()
#dnn = networks.NoOperation()
dnn.load(dnnmodel)
dnn.to_gpu(0)
decoder = networks.NBLSTMCRF()
decoder.load()
decoder.to_gpu(0)
cqtfilelist = np.array(find_files(const.PATH_HCQT, ext="npy"))[idx]
i = 0
chainer.config.train = False
chainer.config.enable_backprop = False
for cqtfile in cqtfilelist:
cqt = utils.Embed(utils.PreprocessSpec(np.load(cqtfile)[:, :, :]), 1)
chroma = dnn.GetFeature(cp.asarray(cqt)).data
path = decoder.argmax(chroma)
feat = cp.asnumpy(chroma)
if todir:
fname = cqtfile.split("/")[-1] + ".lab"
alb = cqtfile.split("/")[-2]
utils.SaveEstimatedLabelsFramewise(
path, const.PATH_ESTIMATE_CROSS + alb + "/" + fname, feat)
else:
utils.SaveEstimatedLabelsFramewise(
path, const.PATH_ESTIMATE + "%03d.lab" % i, feat)
i += 1
def visualize(data_dirs, wav2mel_path, checkpoint_path, output_path):
"""Visualize high-dimensional embeddings using t-SNE."""
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
wav2mel = torch.jit.load(wav2mel_path)
dvector = torch.jit.load(checkpoint_path).eval().to(device)
print("[INFO] model loaded.")
n_spkrs = 0
paths, spkr_names, mels = [], [], []
for data_dir in data_dirs:
data_dir_path = Path(data_dir)
for spkr_dir in [x for x in data_dir_path.iterdir() if x.is_dir()]:
n_spkrs += 1
audio_paths = find_files(spkr_dir)
spkr_name = spkr_dir.name
for audio_path in audio_paths:
paths.append(audio_path)
spkr_names.append(spkr_name)
for audio_path in tqdm(paths, ncols=0, desc="Preprocess"):
wav_tensor, sample_rate = torchaudio.load(audio_path)
with torch.no_grad():
mel_tensor = wav2mel(wav_tensor, sample_rate)
mels.append(mel_tensor)
embs = []
for mel in tqdm(mels, ncols=0, desc="Embed"):
with torch.no_grad():
emb = dvector.embed_utterance(mel.to(device))
emb = emb.detach().cpu().numpy()
embs.append(emb)
tsne = TSNE(n_components=2, verbose=1, perplexity=40, n_iter=300)
transformed = tsne.fit_transform(embs)
print("[INFO] embeddings transformed.")
data = {
"dim-1": transformed[:, 0],
"dim-2": transformed[:, 1],
"label": spkr_names,
}
plt.figure()
sns.scatterplot(
x="dim-1",
y="dim-2",
hue="label",
palette=sns.color_palette(n_colors=n_spkrs),
data=data,
legend="full",
)
plt.legend(loc="center left", bbox_to_anchor=(1, 0.5))
plt.tight_layout()
plt.savefig(output_path)
def ConfMatrix_Allclass(idx):
lab_list = np.array(find_files(C.PATH_CHORDLAB, ext=["lab",
"chords"]))[idx]
estimated_lab_list = [
os.path.join(C.PATH_ESTIMATE_CROSS,
p.split("/")[-2],
p.split("/")[-1]) for p in lab_list
]
durations = np.array([])
confmatrix = np.zeros((C.N_VOCABULARY_TRIADS, C.N_VOCABULARY_TRIADS))
for labfile, estfile in zip(lab_list, estimated_lab_list):
(ref_intervals,
ref_labels) = mir_eval.io.load_labeled_intervals(labfile)
(est_intervals,
est_labels) = mir_eval.io.load_labeled_intervals(estfile)
est_intervals, est_labels = mir_eval.util.adjust_intervals(
est_intervals, est_labels, ref_intervals.min(),
ref_intervals.max(), mir_eval.chord.NO_CHORD,
mir_eval.chord.NO_CHORD)
(intervals, ref_labels,
est_labels) = mir_eval.util.merge_labeled_intervals(
ref_intervals, ref_labels, est_intervals, est_labels)
durations = mir_eval.util.intervals_to_durations(intervals)
ref_labels_id = encode_chordseq_hierarchical(ref_labels)
est_labels_id = encode_chordseq_hierarchical(est_labels)
for i in range(len(ref_labels)):
confmatrix[ref_labels_id[i, 0], est_labels_id[i,
0]] += durations[i]
confmatrix /= np.sum(confmatrix, axis=1, keepdims=True)
return confmatrix
def makeCQTData():
SAVEDIR = PATH_pwCQT
audiolist = find_files(PATH_AUDIO, ext="wav")
itemlist = len(audiolist)
for i, audiofile in enumerate(audiolist):
print("{}/{}".format(i + 1, itemlist))
wav, sr = load(audiofile, sr=SR)
filename = audiofile.split('/')[-1]
albname = audiofile.split('/')[-2]
# foldname = audiofile.split('/')[-3]
# ディレクトリチェック
if not (os.path.exists(SAVEDIR + '/' + albname)):
os.makedirs(SAVEDIR + '/' + albname)
if not (os.path.exists(SAVEDIR + '/' + albname + '/' + filename +
'.npy')):
cqt_spec, freqs = cqt(wav, sr, fmin="C1")
# cqt_power = np.abs(cqt_spec) # sqrtモードで行こう # 2019.12.22
cqt_power = np.array(
[np.sqrt(c.real**2 + c.imag**2) for c in cqt_spec],
dtype="float32")
cqt_power = np.power(cqt_power, 2) # power 系1
cqt_power *= 2 # power 系2
cqt_power = cqt_power.reshape(1, cqt_power.shape[0],
cqt_power.shape[1])
np.save(SAVEDIR + '/' + albname + '/' + filename + '.npy',
np.array(cqt_power, dtype="float32"))
def load_npz(target=None, first=None):
npz_files = find_files('../DSD100_Npz/Dev', ext="npz")[:first]
# npz_files = find_files('../numpy', ext="npz")[:first]
for file in npz_files:
npz = np.load(file)
assert (npz["mix"].shape == npz[target].shape)
yield npz['mix'], npz[target]
def __init__(self, file_path, meta_data, max_timestep=None):
self.roots = file_path
self.root_key = list(self.roots.keys())
self.max_timestep = max_timestep
# extract dev speaker and store in self.black_list_spealers
with open(meta_data, "r") as f:
self.black_list_speakers = f.read().splitlines()
# calculate speakers and support to remove black list speaker (dev)
self.all_speakers = \
[f.path for key in self.root_key for f in os.scandir(self.roots[key]) if f.is_dir()]
self.speaker_num = len(self.all_speakers)
self.necessary_dict = self.processing()
self.label_mapping_spk_id = {}
# speaker id map to speaker num
self.build_label_mapping()
print("search all wavs paths")
start = time.time()
self.dataset = []
for speaker in tqdm.tqdm(self.all_speakers):
wav_list = find_files(speaker)
self.dataset.extend(wav_list)
end = time.time()
print(f"search all wavs paths costs {end-start} seconds")
self.label = self.build_label(self.dataset)
def __init__(self, root):
seed = random.randint(1, 1000)
random.seed(seed)
if (Path(root) / "by_book").exists():
_root = Path(root) / "by_book"
else:
_root = Path(root)
wav_files = []
metadata = defaultdict(list)
speaker_dirs = [
speaker_dir for speaker_dir in (_root / "female").iterdir()
if speaker_dir.is_dir()
]
speaker_dirs += [
speaker_dir for speaker_dir in (_root / "male").iterdir()
if speaker_dir.is_dir()
]
for speaker_dir in speaker_dirs:
for wav_file in find_files(speaker_dir):
wav_file = str(PurePosixPath(wav_file).relative_to(root))
wav_files.append(wav_file)
speaker_id = self.get_speaker(wav_file)
metadata[speaker_id].append(wav_file)
self.root = root
self.seed = seed
self.wav_files = wav_files
self.metadata = metadata
def __init__(self, vad_config, key_list, file_path, meta_data, max_timestep=None):
self.roots = file_path
self.root_key = key_list
self.max_timestep = max_timestep
self.vad_c = vad_config
self.dataset = []
self.all_speakers = []
for index in range(len(self.root_key)):
cache_path = Path(os.path.dirname(__file__)) / 'cache_wav_paths' / f'cache_{self.root_key[index]}.p'
p = Path(self.roots[index])
# loca cache_path if file exists
if os.path.isfile(cache_path):
# cache dict:
# {
# "speaker_id1": ["wav_a_path1", "wav_a_path2", ...],
# "speaker_id2": ["wav_b_path1", "wav_b_path2", ...],
# ...,
# }
cache_wavs_dict = pickle.load(open(cache_path,"rb"))
self.all_speakers.extend(list(cache_wavs_dict.keys()))
for speaker_id in list(cache_wavs_dict.keys()):
for wavs in cache_wavs_dict[speaker_id]:
utterance_id = "/".join(str(p/speaker_id/wavs).split("/")[-3:]).replace(".wav","").replace("/","-")
self.dataset.append([str(p / speaker_id / wavs), utterance_id])
else:
speaker_wav_dict = {}
speaker_dirs = [f.path.split("/")[-1] for f in os.scandir(self.roots[index]) if f.is_dir()]
self.all_speakers.extend(speaker_dirs)
print("search all wavs paths")
start = time.time()
for speaker in tqdm.tqdm(speaker_dirs):
speaker_dir = p / speaker
wav_list=find_files(speaker_dir)
speaker_wav_dict[speaker] = []
for wav in wav_list:
wav_sample, _ = apply_effects_file(str(speaker_dir/wav), EFFECTS)
wav_sample = wav_sample.squeeze(0)
length = wav_sample.shape[0]
if length > self.vad_c['min_sec']:
utterance_id = "/".join(str(speaker_dir/wav).split("/")[-3:]).replace(".wav","").replace("/","-")
self.dataset.append([str(speaker_dir/wav), utterance_id])
speaker_wav_dict[speaker].append("/".join(wav.split("/")[-2:]))
end = time.time()
print(f"search all wavs paths costs {end-start} seconds")
print(f"save wav paths to {cache_path}! so we can directly load all_path in next time!")
pickle.dump(speaker_wav_dict, open(cache_path,"wb"))
self.speaker_num = len(self.all_speakers)
self.necessary_dict = self.processing()
self.label_mapping_spk_id = {}
# speaker id map to speaker num
self.build_label_mapping()
self.label=self.build_label(self.dataset)
def main(data_dirs, out_dir, n_workers, audio_processor_path):
"""Preprocess audio files into features for training."""
audio_paths = chain.from_iterable([find_files(data_dir) for data_dir in data_dirs])
audio_processor_path = Path(audio_processor_path) / "audioprocessor"
audio_processor_path = str(audio_processor_path).replace("/", ".")
audioprocessor = getattr(
importlib.import_module(audio_processor_path), "AudioProcessor"
)
save_dir = Path(out_dir)
save_dir.mkdir(parents=True, exist_ok=True)
executor = ProcessPoolExecutor(max_workers=n_workers)
futures = []
for audio_path in audio_paths:
futures.append(
executor.submit(load_process_save, audioprocessor, audio_path, save_dir)
)
infos = {
"sample_rate": audioprocessor.sample_rate,
"hop_len": audioprocessor.hop_len,
"n_mels": audioprocessor.n_mels,
"utterances": [future.result() for future in tqdm(futures, ncols=0)],
}
with open(save_dir / "metadata.json", "w") as f:
json.dump(infos, f, indent=2)
def SearchErrorExample(idx, ref, est):
lab_list = np.array(find_files(C.PATH_CHORDLAB, ext=["lab",
"chords"]))[idx]
estimated_lab_list = [
os.path.join(C.PATH_ESTIMATE_CROSS,
p.split("/")[-2],
p.split("/")[-1]) for p in lab_list
]
list_examples = []
for labfile, estfile in zip(lab_list, estimated_lab_list):
(ref_intervals,
ref_labels) = mir_eval.io.load_labeled_intervals(labfile)
(est_intervals,
est_labels) = mir_eval.io.load_labeled_intervals(estfile)
est_intervals, est_labels = mir_eval.util.adjust_intervals(
est_intervals, est_labels, ref_intervals.min(),
ref_intervals.max(), mir_eval.chord.NO_CHORD,
mir_eval.chord.NO_CHORD)
(intervals, ref_labels,
est_labels) = mir_eval.util.merge_labeled_intervals(
ref_intervals, ref_labels, est_intervals, est_labels)
durations = mir_eval.util.intervals_to_durations(intervals)
ref_labels_id = encode_chordseq_hierarchical(ref_labels)
est_labels_id = encode_chordseq_hierarchical(est_labels)
for i in range(len(ref_labels)):
if durations[i] > 0.5 and ref_labels_id[
i, 0] == ref and est_labels_id[i, 0] == est:
list_examples.append((labfile, ref, est, intervals[i]))
return list_examples
def LoadSpectrogram_batch(batch_size=10) :
filelist = find_files('./Spectrogram', ext="npz")
batch_list = []
for i in range(0, len(filelist), batch_size):
#print(i)
file_list = filelist[i : i + batch_size]
batch_list.append(file_list)
return batch_list
def separate(PATH_INPUT, PATH_OUTPUT, MODEL, SR=16000, FFT_SIZE = 1024, H = 512):
if os.path.isdir( PATH_INPUT):
# 入力がディレクトリーの場合、ファイルリストをつくる
filelist_mixdown = find_files(PATH_INPUT, ext="wav", case_sensitive=True)
else:
# 入力が単一ファイルの場合
filelist_mixdown=[PATH_INPUT]
print ('number of mixdown file', len(filelist_mixdown))
# 出力用のディレクトリーがない場合は 作成する。
_, path_output_ext = os.path.splitext(PATH_OUTPUT)
print ('path_output_ext',path_output_ext)
if len(path_output_ext)==0 and not os.path.exists(PATH_OUTPUT):
os.mkdir(PATH_OUTPUT)
# モデルの読み込み
unet = train.UNet()
chainer.serializers.load_npz( MODEL,unet)
config.train = False
config.enable_backprop = False
# ミックスされたものを読み込み、vocal(speech)の分離を試みる
for fmixdown in filelist_mixdown:
# audioread でエラーが発生した場合は、scipyを使う。
try:
y_mixdown, _ = load(fmixdown, sr=SR, mono=True)
except:
sr_mixdown, y_mixdown = read(fmixdown)
if not sr_mixdown == SR:
y_mixdown = resample(y_mixdown, sr_mixdown, SR)
# 入力の短時間スペクトラムを計算して、正規化する。
spec = stft(y_mixdown, n_fft=FFT_SIZE, hop_length=H, win_length=FFT_SIZE)
mag = np.abs(spec)
mag /= np.max(mag)
phase = np.exp(1.j*np.angle(spec))
print ('mag.shape', mag.shape)
start = 0
end = 128 * (mag.shape[1] // 128) # 入力のフレーム数以下で、networkの定義に依存して 適切な値を選ぶこと。
# speech(vocal)を分離するためのマスクを求める
mask = unet(mag[:, start:end][np.newaxis, np.newaxis, 1:, :]).data[0, 0, :, :]
mask = np.vstack((np.zeros(mask.shape[1], dtype="float32"), mask))
# 入力の短時間スペクトラムにマスクを掛けて、逆FFTで波形を合成する。
mag2=mag[:, start:end]*mask
phase2=phase[:, start:end]
y = istft(mag2*phase2, hop_length=H, win_length=FFT_SIZE)
# 分離した speech(vocal)を出力ファイルとして保存する。
if len(path_output_ext)==0:
# ディレクトリーへ出力
foutname, _ = os.path.splitext( os.path.basename(fmixdown) )
fname= os.path.join(PATH_OUTPUT, (foutname + '.wav'))
else:
# 指定されたファイルへ出力
fname= PATH_OUTPUT
print ('saving... ', fname)
write_wav(fname, y, SR, norm=True)
def __init__(self, file_path, meta_data, max_timestep=None):
self.roots = file_path
self.root_key = list(self.roots.keys())
self.max_timestep = max_timestep
self.dataset = []
self.all_speakers = []
for key in self.root_key:
cache_path = f"./downstream/voxceleb2_amsoftmax/cache_wav_paths/cache_{key}.p"
p = Path(self.roots[key])
# loca cache_path if file exists
if os.path.isfile(cache_path):
# cache dict =
#{"speaker_id1":["wav_a_path1","wav_a_path2",...],"speaker_id2":["wav_b_path1", "wav_b_path2", ....],...}
cache_wavs_dict = pickle.load(open(cache_path, "rb"))
self.all_speakers.extend(list(cache_wavs_dict.keys()))
for speaker_id in list(cache_wavs_dict.keys()):
for wavs in cache_wavs_dict[speaker_id]:
self.dataset.append(str(p / speaker_id / wavs))
else:
speaker_wav_dict = {}
# calculate speakers and support to remove black list speaker (dev)
speaker_dirs = [
f.path.split("/")[-1] for f in os.scandir(self.roots[key])
if f.is_dir()
]
self.all_speakers.extend(speaker_dirs)
print("search all wavs paths")
start = time.time()
for speaker in tqdm.tqdm(speaker_dirs):
speaker_dir = p / speaker
wav_list = find_files(speaker_dir)
speaker_wav_dict[speaker] = []
for wav in wav_list:
self.dataset.append(str(speaker_dir / wav))
speaker_wav_dict[speaker].append("/".join(
wav.split("/")[-2:]))
end = time.time()
print(f"search all wavs paths costs {end-start} seconds")
print(
f"save wav paths to {cache_path}! so we can directly load all_path in next time!"
)
pickle.dump(speaker_wav_dict, open(cache_path, "wb"))
self.speaker_num = len(self.all_speakers)
self.necessary_dict = self.processing()
self.label_mapping_spk_id = {}
# speaker id map to speaker num
self.build_label_mapping()
self.label = self.build_label(self.dataset)
def TrainConvnetExtractor(trainidx, epoch=20, saveas="convnet.model"):
cqtfilelist = np.array(find_files(const.PATH_MIDIHCQT,
ext="npz"))[trainidx]
#midifilelist = find_files(const.PATH_MIDI,ext="mid")[:filecnt]
config.train = True
config.enable_backprop = True
convnet = networks.FullCNNFeatExtractor()
model = networks.ConvnetPredictor(convnet)
model.to_gpu(0)
opt = optimizers.AdaDelta()
opt.setup(model)
print("train set length: %d" % trainidx.size)
print("start epochs...")
S = []
T = []
for cqtfile in cqtfilelist:
dat = np.load(cqtfile)
spec = utils.PreprocessSpec(dat["spec"])[:const.CQT_H, :, :]
targ = GetConvnetTargetFromPianoroll(dat["target"]).astype(np.int32)
assert (spec.shape[1] == targ.shape[0])
S.append(spec)
T.append(targ)
S = np.concatenate(S, axis=1)
T = np.concatenate(T, axis=0)
for ep in range(epoch):
sum_loss = 0
assert (S.shape[1] == T.shape[0])
randidx = np.random.randint(0,
S.shape[1] - const.CONV_TRAIN_SEQLEN - 1,
S.shape[1] // const.CONV_TRAIN_SEQLEN * 4)
for i in range(0, randidx.size - const.CONV_TRAIN_BATCH,
const.CONV_TRAIN_BATCH):
x_batch = np.stack([
S[:, randidx[j]:randidx[j] + const.CONV_TRAIN_SEQLEN, :]
for j in range(i, i + const.CONV_TRAIN_BATCH)
])
t_batch = np.stack([
T[randidx[j]:randidx[j] + const.CONV_TRAIN_SEQLEN, :]
for j in range(i, i + const.CONV_TRAIN_BATCH)
])
x_in = cp.asarray(x_batch)
t_in = cp.asarray(t_batch)
model.cleargrads()
loss = model(x_in, t_in)
loss.backward()
opt.update()
sum_loss += loss.data
convnet.save(saveas)
print("epoch: %d/%d loss:%.04f" %
(ep + 1, epoch, sum_loss / const.CONV_TRAIN_BATCH))
convnet.save(saveas)
def getNoteTemplates(path_notes):
list_templates = []
list_noteaudio = find_files(path_notes,ext="wav")
for noteaudio in list_noteaudio:
S_mag = U.LoadAudio(noteaudio)
init_H = np.ones((1,S_mag.shape[1]))
template,activate = NMF.nmf_sklearn(S_mag,k=1,H=init_H,verbose=False)
list_templates.append(template[:,0]/np.max(template))
templates = np.stack(list_templates)
return templates
def reload_data(path_to_features, part):
matfiles = find_files(path_to_mat + part + '/', ext='mat')
for i in range(len(matfiles)):
if matfiles[i][len(path_to_mat) + len(part) + 1:].startswith('LFCC'):
key = matfiles[i][len(path_to_mat) + len(part) + 6:-4]
lfcc = sio.loadmat(matfiles[i],
verify_compressed_data_integrity=False)['x']
with open(path_to_features + part + '/' + key + 'LFCC.pkl',
'wb') as handle2:
pickle.dump(lfcc, handle2, protocol=pickle.HIGHEST_PROTOCOL)
def get_filenames(dir):
files = find_files(dir, ext='npy')
filenames = []
for f in files:
f = os.path.basename(f)[:-4]
filenames.append(f)
return filenames
def shuffleCQT_baka():
TORIDASUBASHO = PATH_hrCQT
HOZONBASYO = PATH_SAME2
HOZONBASYO2 = PATH_SAME3
OVTONE = 2
audiolist = find_files(PATH_AUDIO, ext="wav")
cqtlist = find_files(TORIDASUBASHO, ext="npy")
itemlist = len(audiolist)
i_counter = 1
for audiofile, cqtfile in zip(audiolist, cqtlist):
print("{}/{}".format(i_counter, itemlist))
filename = audiofile.split('/')[-1]
albname = audiofile.split('/')[-2]
# foldname = audiofile.split('/')[-3]
cqt_filename = cqtfile.split("/")[-1]
if (filename.split(".")[0] != cqt_filename.split(".")[0]):
print("file_not_match", filename, cqt_filename)
# ディレクトリチェック
if not (os.path.exists(HOZONBASYO + '/' + albname)):
os.makedirs(HOZONBASYO + '/' + albname)
if not (os.path.exists(HOZONBASYO + '/' + albname + '/' + filename +
'.npy')):
cqt = np.load(cqtfile)
excqt = np.vstack((cqt, cqt))
np.save(HOZONBASYO + '/' + albname + '/' + filename + '.npy',
np.array(excqt, dtype="float32"))
if not (os.path.exists(HOZONBASYO2 + '/' + albname)):
os.makedirs(HOZONBASYO2 + '/' + albname)
if not (os.path.exists(HOZONBASYO2 + '/' + albname + '/' + filename +
'.npy')):
cqt = np.load(cqtfile)
excqt = np.vstack((cqt, cqt, cqt))
np.save(HOZONBASYO2 + '/' + albname + '/' + filename + '.npy',
np.array(excqt, dtype="float32"))
i_counter += 1
def __init__(self, roots, sample_rate, max_time, target_level, noise_proportion, snrs, **kwargs):
self.sample_rate = sample_rate
self.max_time = max_time
self.target_level = target_level
self.noise_proportion = noise_proportion
self.snrs = snrs
self.filepths = []
for root in roots:
self.filepths += find_files(root)
assert len(self.filepths) > 0, 'No audio file detected'
self.noise_sampler = torch.distributions.Normal(0, 1)
def LoadSpectrogram(target="vocal") :
filelist = find_files('./Spectrogram', ext="npz")
x_list = []
y_list = []
for file in filelist :
data = np.load(file)
x_list.append(data['mix'])
if target == "vocal" :
y_list.append(data['vocal'])
else :
y_list.append(data['inst'])
return x_list, y_list
def LoadDataset(target="vocal"):
filelist_fft = find_files(C.PATH_FFT, ext="npz")[:200]
Xlist = []
Ylist = []
for file_fft in filelist_fft:
dat = np.load(file_fft)
Xlist.append(dat["mix"])
if target == "vocal":
assert(dat["mix"].shape == dat["vocal"].shape)
Ylist.append(dat["vocal"])
else:
assert(dat["mix"].shape == dat["inst"].shape)
Ylist.append(dat["inst"])
return Xlist, Ylist
def get_ann_audio(directory):
'''Get a list of annotations and audio files from a directory.
This also validates that the lengths match and are paired properly.
Parameters
----------
directory : str
The directory to search
Returns
-------
pairs : list of tuples (audio_file, annotation_file)
'''
audio = find_files(directory)
annos = find_files(directory, ext=['jams', 'jamz'])
paired = list(zip(audio, annos))
if len(audio) != len(annos) or any([base(aud) != base(ann) for aud, ann in paired]):
raise DataError('Unmatched audio/annotation data in {}'.format(directory))
return paired
# -*- coding: utf-8 -*-
"""
Created on Sat May 7 16:45:29 2016
@author: parallels
"""
import functions
from librosa.util import find_files
from librosa.core import load
audiofilelist = find_files("database/audios/",ext = "wav")
print "saving peaks...."
for audiofile in audiofilelist:
y,sr = load(audiofile)
filename = audiofile.split("/")[-1]+".npy" # -1 means the last name of the directory
functions.save_maximum_array(y,filename)
print "saved:" + filename
"""
Created on Fri Nov 3 10:59:08 2017
@author: wuyiming
"""
import os
from librosa.core import load
from librosa.util import find_files
import yaml
import util
PATH_MENDLEY = "MedleyDB/Audio"
metadatalist = find_files(PATH_MENDLEY, ext="yaml")
all_voctracks = []
all_insttracks = []
for metafile in metadatalist:
print("YAML file: %s" % metafile)
songname = metafile.split("/")[-2]
print("song: %s" % songname)
with open(metafile, "r+") as f:
data = yaml.load(f)
if data["instrumental"] != "no":
print("Instrumental track. Skipped.")
continue
