def encodes(self, sg: AudioSpectrogram) -> AudioSpectrogram:
channel_mean = sg.contiguous().view(sg.size(0), -1).mean(-1)[:, None, None]
mask_val = ifnone(self.val, channel_mean)
c, y, x = sg.shape
# Position of the first mask
start = ifnone(self.start, random.randint(0, y - self.size))
for _ in range(self.num_masks):
mask = torch.ones(self.size, x) * mask_val
if not 0 <= start <= y - self.size:
raise ValueError(
f"Start value '{start}' out of range for AudioSpectrogram of shape {sg.shape}"
)
sg[:, start : start + self.size, :] = mask
# Setting start position for next mask
start = random.randint(0, y - self.size)
return sg
def encodes(self, ai: AudioTensor) -> AudioTensor:
sig = ai.data
orig_samples = ai.nsamples
crop_samples = int((self.duration / 1000) * ai.sr)
if orig_samples == crop_samples:
return ai
elif orig_samples < crop_samples:
ai.data = _tfm_pad_signal(sig, crop_samples, pad_mode=self.pad_mode)
else:
crop_start = random.randint(0, int(orig_samples - crop_samples))
ai.data = sig[:, crop_start : crop_start + crop_samples]
return ai
def _tfm_pad_signal(sig, width, pad_mode=AudioPadType.Zeros):
"""Pad spectrogram to specified width, using specified pad mode"""
c, x = sig.shape
if pad_mode in [AudioPadType.Zeros, AudioPadType.Zeros_After]:
zeros_front = (random.randint(0, width - x)
if pad_mode == AudioPadType.Zeros else 0)
pad_front = torch.zeros((c, zeros_front))
pad_back = torch.zeros((c, width - x - zeros_front))
return torch.cat((pad_front, sig, pad_back), 1)
elif pad_mode == AudioPadType.Repeat:
repeats = width // x + 1
return sig.repeat(1, repeats)[:, :width]
def encodes(self, sg: AudioSpectrogram) -> AudioSpectrogram:
sr, hop = sg.sr, sg.hop_length
w_crop = int((sr * self.duration) / (1000 * hop)) + 1
w_sg = sg.shape[-1]
if w_sg == w_crop:
sg_crop = sg
elif w_sg < w_crop:
sg_crop = _tfm_pad_spectro(sg, w_crop, pad_mode=self.pad_mode)
else:
crop_start = random.randint(0, int(w_sg - w_crop))
sg_crop = sg[:, :, crop_start : crop_start + w_crop]
sg_crop.sample_start = int(crop_start * hop)
sg_crop.sample_end = sg_crop.sample_start + int(self.duration * sr)
sg.data = sg_crop
return sg
def _tfm_pad_spectro(sg, width, pad_mode=AudioPadType.Zeros):
"""Pad spectrogram to specified width, using specified pad mode"""
c, y, x = sg.shape
if pad_mode in [AudioPadType.Zeros, AudioPadType.Zeros_After]:
padded = torch.zeros((c, y, width))
start = random.randint(0, width -
x) if pad_mode == AudioPadType.Zeros else 0
padded[:, :, start:start + x] = sg.data
return padded
elif pad_mode == AudioPadType.Repeat:
repeats = width // x + 1
return sg.repeat(1, 1, repeats)[:, :, :width]
else:
raise ValueError(f"""pad_mode {pad_mode} not currently supported,
only AudioPadType.Zeros, AudioPadType.Zeros_After,
or AudioPadType.Repeat""")