Datasets:

Size:
n<1K
ArXiv:
License:
Guzheng_Tech99 / Guzheng_Tech99.py
monetjoe's picture
Update Guzheng_Tech99.py
cb70f80 verified
raw
history blame
12.4 kB
import os
import csv
import random
import librosa
import datasets
import numpy as np
from tqdm import tqdm
from glob import glob
_NAMES = {
"chanyin": 0,
"dianyin": 6,
"shanghua": 2,
"xiahua": 3,
"huazhi": 4,
"guazou": 4,
"lianmo": 4,
"liantuo": 4,
"yaozhi": 5,
"boxian": 1,
}
_NAME = [
"chanyin", # Vibrato
"boxian", # Plucks
"shanghua", # Upward Portamento
"xiahua", # Downward Portamento
"huazhi/guazou/lianmo/liantuo", # Glissando
"yaozhi", # Tremolo
"dianyin", # Point Note
]
_HOMEPAGE = f"https://www.modelscope.cn/datasets/ccmusic-database/{os.path.basename(__file__)[:-3]}"
_DOMAIN = f"{_HOMEPAGE}/resolve/master/data"
_URLS = {
"audio": f"{_DOMAIN}/audio.zip",
"mel": f"{_DOMAIN}/mel.zip",
"label": f"{_DOMAIN}/label.zip",
}
_TIME_LENGTH = 3 # seconds
_SAMPLE_RATE = 44100
_HOP_LENGTH = 512 # SAMPLE_RATE * ZHEN_LENGTH // 1000
class Guzheng_Tech99(datasets.GeneratorBasedBuilder):
def _info(self):
return datasets.DatasetInfo(
features=(
datasets.Features(
{
"audio": datasets.Audio(sampling_rate=44100),
"mel": datasets.Image(),
"label": datasets.Sequence(
feature={
"onset_time": datasets.Value("float32"),
"offset_time": datasets.Value("float32"),
"IPT": datasets.ClassLabel(num_classes=7, names=_NAME),
"note": datasets.Value("int8"),
}
),
}
)
if self.config.name == "default"
else datasets.Features(
{
"mel": datasets.features.Array3D(
dtype="float32", shape=(128, 258, 1)
),
"cqt": datasets.features.Array3D(
dtype="float32", shape=(88, 258, 1)
),
"chroma": datasets.features.Array3D(
dtype="float32", shape=(12, 258, 1)
),
"label": datasets.features.Array2D(
dtype="float32", shape=(7, 258)
),
}
)
),
homepage=_HOMEPAGE,
license="CC-BY-NC-ND",
version="1.2.0",
)
def _RoW_norm(self, data):
common_sum = 0
square_sum = 0
tfle = 0
for i in range(len(data)):
tfle += (data[i].sum(-1).sum(0) != 0).astype("float").sum()
common_sum += data[i].sum(-1).sum(-1)
square_sum += (data[i] ** 2).sum(-1).sum(-1)
common_avg = common_sum / tfle
square_avg = square_sum / tfle
std = np.sqrt(square_avg - common_avg**2)
return common_avg, std
def _norm(self, data):
size = data.shape
avg, std = self._RoW_norm(data)
avg = np.tile(avg.reshape((1, -1, 1, 1)), (size[0], 1, size[2], size[3]))
std = np.tile(std.reshape((1, -1, 1, 1)), (size[0], 1, size[2], size[3]))
return (data - avg) / std
def _load(self, wav_dir, csv_dir, groups):
def files(wav_dir, csv_dir, group):
flacs = sorted(glob(os.path.join(wav_dir, group, "*.flac")))
if len(flacs) == 0:
flacs = sorted(glob(os.path.join(wav_dir, group, "*.wav")))
csvs = sorted(glob(os.path.join(csv_dir, group, "*.csv")))
files = list(zip(flacs, csvs))
if len(files) == 0:
raise RuntimeError(f"Group {group} is empty")
result = []
for audio_path, csv_path in files:
result.append((audio_path, csv_path))
return result
def logMel(y, sr=_SAMPLE_RATE):
# 帧长为32ms (1000ms/(16000/512) = 32ms), D2的频率是73.418
mel = librosa.feature.melspectrogram(
y=y,
sr=sr,
hop_length=_HOP_LENGTH,
fmin=27.5,
)
return librosa.power_to_db(mel, ref=np.max)
# Returns the CQT of the input audio
def logCQT(y, sr=_SAMPLE_RATE):
# 帧长为32ms (1000ms/(16000/512) = 32ms), D2的频率是73.418
cqt = librosa.cqt(
y,
sr=sr,
hop_length=_HOP_LENGTH,
fmin=27.5,
n_bins=88,
bins_per_octave=12,
)
return (
(1.0 / 80.0) * librosa.core.amplitude_to_db(np.abs(cqt), ref=np.max)
) + 1.0
def logChroma(y, sr=_SAMPLE_RATE):
# 帧长为32ms (1000ms/(16000/512) = 32ms), D2的频率是73.418
chroma = librosa.feature.chroma_stft(
y=y,
sr=sr,
hop_length=_HOP_LENGTH,
)
return (
(1.0 / 80.0) * librosa.core.amplitude_to_db(np.abs(chroma), ref=np.max)
) + 1.0
def chunk_data(f):
x = []
xdata = np.transpose(f)
s = _SAMPLE_RATE * _TIME_LENGTH // _HOP_LENGTH
length = int(np.ceil((int(len(xdata) / s) + 1) * s))
app = np.zeros((length - xdata.shape[0], xdata.shape[1]))
xdata = np.concatenate((xdata, app), 0)
for i in range(int(length / s)):
data = xdata[int(i * s) : int(i * s + s)]
x.append(np.transpose(data[:s, :]))
return np.array(x)
def load_all(audio_path, csv_path, hop=_HOP_LENGTH, n_IPTs=7, technique=_NAMES):
# Load audio features: The shape of cqt (88, 8520), 8520 is the number of frames on the time axis
y, sr = librosa.load(audio_path, sr=_SAMPLE_RATE)
mel = logMel(y, sr)
cqt = logCQT(y, sr)
chroma = logChroma(y, sr)
# Load the ground truth label
n_steps = cqt.shape[1]
IPT_label = np.zeros([n_IPTs, n_steps], dtype=int)
with open(csv_path, "r", encoding="utf-8") as f: # csv file for each audio
reader = csv.DictReader(f, delimiter=",")
for label in reader: # each note
onset = float(label["onset_time"])
offset = float(label["offset_time"])
IPT = int(technique[label["IPT"]])
left = int(round(onset * _SAMPLE_RATE / hop))
frame_right = int(round(offset * _SAMPLE_RATE / hop))
frame_right = min(n_steps, frame_right)
IPT_label[IPT, left:frame_right] = 1
return dict(
audio_path=audio_path,
csv_path=csv_path,
mel=mel,
cqt=cqt,
chroma=chroma,
IPT_label=IPT_label,
)
data = []
# print(f"Loading {len(groups)} group{'s' if len(groups) > 1 else ''} ")
for group in groups:
for input_files in files(wav_dir, csv_dir, group):
data.append(load_all(*input_files))
for i, dic in tqdm(enumerate(data), total=len(data), desc="Feature extracting"):
x_mel = chunk_data(dic["mel"])
x_cqt = chunk_data(dic["cqt"])
x_chroma = chunk_data(dic["chroma"])
y_i = dic["IPT_label"]
y_i = chunk_data(y_i)
if i == 0:
Xtr_mel = x_mel
Xtr_cqt = x_cqt
Xtr_chroma = x_chroma
Ytr_i = y_i
else:
Xtr_mel = np.concatenate([Xtr_mel, x_mel], axis=0)
Xtr_cqt = np.concatenate([Xtr_cqt, x_cqt], axis=0)
Xtr_chroma = np.concatenate([Xtr_chroma, x_chroma], axis=0)
Ytr_i = np.concatenate([Ytr_i, y_i], axis=0)
# Transform the shape of the input
Xtr_mel = np.expand_dims(Xtr_mel, axis=3)
Xtr_cqt = np.expand_dims(Xtr_cqt, axis=3)
Xtr_chroma = np.expand_dims(Xtr_chroma, axis=3)
# Normalize
Xtr_mel = self._norm(Xtr_mel)
Xtr_cqt = self._norm(Xtr_cqt)
Xtr_chroma = self._norm(Xtr_chroma)
return [list(Xtr_mel), list(Xtr_cqt), list(Xtr_chroma)], list(Ytr_i)
def _parse_csv_label(self, csv_file):
label = []
with open(csv_file, mode="r", encoding="utf-8") as file:
for row in csv.DictReader(file):
label.append(
{
"onset_time": float(row["onset_time"]),
"offset_time": float(row["offset_time"]),
"IPT": _NAME[_NAMES[row["IPT"]]],
"note": int(row["note"]),
}
)
return label
def _split_generators(self, dl_manager):
audio_files = dl_manager.download_and_extract(_URLS["audio"])
csv_files = dl_manager.download_and_extract(_URLS["label"])
trainset, validset, testset = [], [], []
if self.config.name == "default":
files = {}
mel_files = dl_manager.download_and_extract(_URLS["mel"])
for path in dl_manager.iter_files([audio_files]):
fname: str = os.path.basename(path)
if fname.endswith(".flac"):
item_id = fname.split(".")[0]
files[item_id] = {"audio": path}
for path in dl_manager.iter_files([mel_files]):
fname = os.path.basename(path)
if fname.endswith(".jpg"):
item_id = fname.split(".")[0]
files[item_id]["mel"] = path
for path in dl_manager.iter_files([csv_files]):
fname = os.path.basename(path)
if fname.endswith(".csv"):
item_id = fname.split(".")[0]
files[item_id]["label"] = self._parse_csv_label(path)
for item in files.values():
if "train" in item["audio"]:
trainset.append(item)
elif "validation" in item["audio"]:
validset.append(item)
elif "test" in item["audio"]:
testset.append(item)
else:
audio_dir = os.path.join(audio_files, "audio")
csv_dir = os.path.join(csv_files, "label")
X_train, Y_train = self._load(audio_dir, csv_dir, ["train"])
X_valid, Y_valid = self._load(audio_dir, csv_dir, ["validation"])
X_test, Y_test = self._load(audio_dir, csv_dir, ["test"])
for i in range(len(Y_train)):
trainset.append(
{
"mel": X_train[0][i],
"cqt": X_train[1][i],
"chroma": X_train[2][i],
"label": Y_train[i],
}
)
for i in range(len(Y_valid)):
validset.append(
{
"mel": X_valid[0][i],
"cqt": X_valid[1][i],
"chroma": X_valid[2][i],
"label": Y_valid[i],
}
)
for i in range(len(Y_test)):
testset.append(
{
"mel": X_test[0][i],
"cqt": X_test[1][i],
"chroma": X_test[2][i],
"label": Y_test[i],
}
)
random.shuffle(trainset)
random.shuffle(validset)
random.shuffle(testset)
return [
datasets.SplitGenerator(
name=datasets.Split.TRAIN, gen_kwargs={"files": trainset}
),
datasets.SplitGenerator(
name=datasets.Split.VALIDATION, gen_kwargs={"files": validset}
),
datasets.SplitGenerator(
name=datasets.Split.TEST, gen_kwargs={"files": testset}
),
]
def _generate_examples(self, files):
for i, path in enumerate(files):
yield i, path