def test_interp_harmonics_multi_static(s_multi):
S, sr = s_multi
freqs = librosa.fft_frequencies(sr=sr)
Hall = librosa.interp_harmonics(S, freqs=freqs, harmonics=[0.5, 1, 2])
H0 = librosa.interp_harmonics(S[0], freqs=freqs, harmonics=[0.5, 1, 2])
H1 = librosa.interp_harmonics(S[1], freqs=freqs, harmonics=[0.5, 1, 2])
assert np.allclose(Hall[0], H0)
assert np.allclose(Hall[1], H1)
assert not np.allclose(H0, H1)
def relacion_fundamental_harmonicos(file):
y, sr = librosa.load(file)
# y, sr = librosa.load("../audios/clash cymbals/clash-cymbals__long_forte_undamped.mp3")
h_range = [1, 2, 3, 4, 5, 6, 7, 8, 9, 10]
S = np.abs(librosa.stft(y))
fft_freqs = librosa.fft_frequencies(sr=sr)
S_harm = librosa.interp_harmonics(S, fft_freqs, h_range, axis=0)
return np.sum(S_harm[1]) / np.sum(S_harm[2:])
def interp_hcqt(audio_fpath=None, y=None, fs=None):
"""Compute the harmonic CQT from a given audio file
Parameters
----------
audio_fpath : str
path to audio file
Returns
-------
hcqt : np.ndarray
Harmonic cqt
time_grid : np.ndarray
List of time stamps in seconds
freq_grid : np.ndarray
List of frequency values in Hz
"""
if y is None:
y, fs = librosa.load(audio_fpath, sr=SR)
else:
y = librosa.resample(y, fs, SR)
fs = SR
# How many bins do we need?
# n_bins_max = 2**(ceil(log2(max(HARMONICS))) * BPO + n_bins)
n_bins_plane = N_OCTAVES * BINS_PER_OCTAVE
n_bins_master = int(np.ceil(np.log2(np.max(HARMONICS))) * BINS_PER_OCTAVE) + n_bins_plane
cqt_master = np.abs(librosa.cqt(y=y, sr=fs,
hop_length=HOP_LENGTH,
fmin=FMIN,
n_bins=n_bins_master,
bins_per_octave=BINS_PER_OCTAVE))
freq_grid = librosa.cqt_frequencies(N_OCTAVES * BINS_PER_OCTAVE,
FMIN,
bins_per_octave=BINS_PER_OCTAVE)
freq_master = librosa.cqt_frequencies(n_bins_master, FMIN,
bins_per_octave=BINS_PER_OCTAVE)
hcqt = librosa.interp_harmonics(cqt_master,
freq_master,
HARMONICS)[:, :N_OCTAVES * BINS_PER_OCTAVE]
log_hcqt = ((1.0/80.0) * librosa.core.amplitude_to_db(hcqt, ref=np.max)) + 1.0
time_grid = librosa.core.frames_to_time(
np.arange(log_hcqt.shape[-1]), sr=SR, hop_length=HOP_LENGTH
)
return log_hcqt, freq_grid, time_grid
def test_interp_harmonics_multi_vary(tfr_multi):
times, freqs, mags = tfr_multi
# Force slinear mode here to deal with non-unique frequencies
Hall = librosa.interp_harmonics(mags,
freqs=freqs,
harmonics=[0.5, 1, 2],
kind="slinear")
H0 = librosa.interp_harmonics(mags[0],
freqs=freqs[0],
harmonics=[0.5, 1, 2],
kind="slinear")
H1 = librosa.interp_harmonics(mags[1],
freqs=freqs[1],
harmonics=[0.5, 1, 2],
kind="slinear")
assert np.allclose(Hall[0], H0)
assert np.allclose(Hall[1], H1)
assert not np.allclose(H0, H1)
def test_harmonics_2d():
x = np.arange(16)
y = np.linspace(-8, 8, num=len(x), endpoint=False)**2
y = np.tile(y, (5, 1)).T
h = [0.25, 0.5, 1, 2, 4]
yh = librosa.interp_harmonics(y, x, h, axis=0)
eq_(yh.shape[1:], y.shape)
eq_(yh.shape[0], len(h))
for i in range(len(h)):
if h[i] <= 1:
# Check that subharmonics match
step = int(1./h[i])
vals = yh[i, ::step]
assert np.allclose(vals, y[:len(vals)])
else:
# Else check that harmonics match
step = h[i]
vals = y[::step]
assert np.allclose(vals, yh[i, :len(vals)])
def test_harmonics_badshape_2d():
freqs = np.zeros((5, 5))
obs = np.zeros((5, 10))
librosa.interp_harmonics(obs, freqs, [1])
def harmonic_extract(y, sr):
h_range = [1]
S = np.abs(librosa.stft(y))
fft_freqs = librosa.fft_frequencies(sr=sr)
S_harm = librosa.interp_harmonics(S, fft_freqs, h_range, axis=0)[0]
return S_harm
def __getitem__(self, i):
"""Return a sample from the dataset."""
related_track_idxs = self.related_track_idxs.iat[i]
# set random seed for all sampling
seed = np.random.randint(0, 10000000)
# load STFT of related stems, sample and add
for j, track_idx in enumerate(related_track_idxs):
# load metadata and temp stft tensor
stft_path = self.metadata.at[track_idx, 'stft_path']
instrument = self.metadata.at[track_idx, 'instrument']
volume = self.metadata.at[track_idx, 'trackVolume']
tmp = self._load(stft_path, volume, seed)
# initialize tensors
if j == 0:
X = torch.zeros_like(tmp)
y_all_complex = [
torch.zeros_like(tmp) for _ in range(self.n_classes)
]
c_all = [0] * self.n_classes
# add tensors
X.add_(tmp)
y_all_complex[instrument].add_(tmp)
c_all[instrument] = 1
# take magnitude of combined stems and scale
X_complex = torch.zeros_like(X).copy_(X)
X = self._stft_mag(X)
if self.harmonics:
X = X.numpy()
if X.shape[1] == 1025:
fft_freqs = librosa.fft_frequencies(sr=22050, n_fft=2048)
elif X.shape[1] == 2049:
fft_freqs = librosa.fft_frequencies(sr=44100, n_fft=4096)
X = librosa.interp_harmonics(X, fft_freqs, [1, 2, 3], axis=1)
X = torch.from_numpy(X)
X = X.view(6, X.size(2), self.n_frames)
# stack targets and add metadata for collate function
y_all_complex = torch.stack(y_all_complex)
if X.dim() == 2:
X_complex = X_complex.unsqueeze(0)
X = X.unsqueeze(0)
y_all_complex = y_all_complex.unsqueeze(1)
t = torch.tensor([y_all_complex.size(-2)])
c = torch.tensor(c_all).long()
track_idx = torch.tensor(track_idx)
return {
'X': X,
'X_complex': X_complex,
'y_complex': y_all_complex,
'c': c,
't': t,
'track_idx': track_idx
}
def __getitem__(self, i):
"""Return a sample from the dataset."""
related_track_idxs = self.related_track_idxs.iat[i]
# load STFT of related stems, sample and add
for j, track_idx in enumerate(related_track_idxs):
# load metadata and temp stft tensor
stft_path = self.metadata.at[track_idx, 'stft_path']
instrument = self.metadata.at[track_idx, 'instrument']
volume = self.metadata.at[track_idx, 'trackVolume']
tmp = self._load(stft_path, volume)
# initialize tensors
if j == 0:
X = torch.zeros_like(tmp)
y_all_complex = [
torch.zeros_like(tmp) for _ in range(self.n_classes)
]
c_all = [0] * self.n_classes
# add tensors
X.add_(tmp)
y_all_complex[instrument].add_(tmp)
c_all[instrument] = 1
# take magnitude of combined stems and scale and split into chunks
X_complex = torch.zeros_like(X).copy_(X)
X = self._stft_mag(X)
# stack targets and add metadata for collate function
y_all_complex = torch.stack(y_all_complex)
if X.dim() == 2:
X_complex, _ = self._split_track(X_complex, self.n_frames, 1)
X_complex = X_complex.unsqueeze(1)
X, ns = self._split_track(X, self.n_frames, 1)
X = X.unsqueeze(1)
y_all_complex = y_all_complex.unsqueeze(1)
else:
X_complex0, _ = self._split_track(X_complex[0], self.n_frames, 1)
X_complex1, _ = self._split_track(X_complex[1], self.n_frames, 1)
X_complex = torch.stack([X_complex0, X_complex1], dim=1)
X0, _ = self._split_track(X[0], self.n_frames, 1)
X1, ns = self._split_track(X[1], self.n_frames, 1)
X = torch.stack([X0, X1], dim=1)
if self.harmonics:
X = X.numpy()
if X.shape[2] == 1025:
fft_freqs = librosa.fft_frequencies(sr=22050, n_fft=2048)
elif X.shape[2] == 2049:
fft_freqs = librosa.fft_frequencies(sr=44100, n_fft=4096)
X = librosa.interp_harmonics(X, fft_freqs, [1, 2, 3], axis=2)
X = torch.from_numpy(X)
X = X.transpose(1, 0).contiguous()
X = X.view(X.size(0), 6, X.size(3), self.n_frames)
t = torch.tensor([y_all_complex.size(-2)])
c = torch.tensor(c_all).long()
track_idx = torch.tensor(track_idx)
return {
'X': X,
'X_complex': X_complex,
'y_complex': y_all_complex,
'c': c,
'ns': ns,
't': t,
'track_idx': track_idx
}
def __getitem__(self, i):
"""Return a sample from the dataset."""
# target track metadata
related_track_idxs = self.related_track_idxs.iat[i]
# sample volume alphas
n_related = len(related_track_idxs)
volume_alphas = self._sample_volume_alphas(n_related)
if self.threshold:
volume_alphas = F.threshold(volume_alphas, self.threshold, 0.0)
if self.instrument_mask:
volume_alphas *= self._sample_instrument_mask(n_related)
# set random seed for all sampling
seed = np.random.randint(0, 10000000)
# load STFT of related stems, sample and add
for j, track_idx in enumerate(related_track_idxs):
# load metadata and temp stft tensor
stft_path = self.metadata.at[track_idx, 'stft_path']
instrument = self.metadata.at[track_idx, 'instrument']
tmp = self._load(stft_path, volume_alphas[j], seed)
if self.chan_swap:
tmp = self._chan_swap(tmp)
if self.gain_slope:
tmp = self._gain_modulation(tmp)
# initialize input tensor and add
if j == 0:
X = torch.zeros_like(tmp)
X.add_(tmp)
# magnitdue and scale target for tensors
tmp = self._stft_mag(tmp)
# initialize target tensor and add
if j == 0:
y_all = [torch.zeros_like(tmp) for _ in range(self.n_classes)]
y_all[instrument].add_(tmp)
# add random interference
X = self._add_interfere(X, volume_alphas, seed)
# take magnitude of combined stems and scale
X = self._stft_mag(X)
if self.harmonics:
X = X.numpy()
if X.shape[1] == 1025:
fft_freqs = librosa.fft_frequencies(sr=22050, n_fft=2048)
elif X.shape[1] == 2049:
fft_freqs = librosa.fft_frequencies(sr=44100, n_fft=4096)
X = librosa.interp_harmonics(X, fft_freqs, [1, 2, 3], axis=1)
X = torch.from_numpy(X)
X = X.view(6, X.size(2), self.n_frames)
# stack target tensors
y_all = torch.stack(y_all)
if X.dim() == 2:
X = X.unsqueeze(0)
y_all = y_all.unsqueeze(1)
return {'X': X, 'y': y_all}
