|
|
import numpy as np |
|
|
import torch |
|
|
import random |
|
|
|
|
|
|
|
|
def frame_shift(mels, labels, net_pooling=4): |
|
|
bsz, n_bands, frames = mels.shape |
|
|
shifted = [] |
|
|
new_labels = [] |
|
|
for bindx in range(bsz): |
|
|
shift = int(random.gauss(0, 90)) |
|
|
shifted.append(torch.roll(mels[bindx], shift, dims=-1)) |
|
|
shift = -abs(shift) // net_pooling if shift < 0 else shift // net_pooling |
|
|
new_labels.append(torch.roll(labels[bindx], shift, dims=-1)) |
|
|
return torch.stack(shifted), torch.stack(new_labels) |
|
|
|
|
|
|
|
|
def mixup(data, target=None, alpha=0.2, beta=0.2, mixup_label_type="soft"): |
|
|
"""Mixup data augmentation by permuting the data |
|
|
|
|
|
Args: |
|
|
data: input tensor, must be a batch so data can be permuted and mixed. |
|
|
target: tensor of the target to be mixed, if None, do not return targets. |
|
|
alpha: float, the parameter to the np.random.beta distribution |
|
|
beta: float, the parameter to the np.random.beta distribution |
|
|
mixup_label_type: str, the type of mixup to be used choice between {'soft', 'hard'}. |
|
|
Returns: |
|
|
torch.Tensor of mixed data and labels if given |
|
|
""" |
|
|
with torch.no_grad(): |
|
|
batch_size = data.size(0) |
|
|
c = np.random.beta(alpha, beta) |
|
|
|
|
|
perm = torch.randperm(batch_size) |
|
|
|
|
|
mixed_data = c * data + (1 - c) * data[perm, :] |
|
|
if target is not None: |
|
|
if mixup_label_type == "soft": |
|
|
mixed_target = torch.clamp( |
|
|
c * target + (1 - c) * target[perm, :], min=0, max=1 |
|
|
) |
|
|
elif mixup_label_type == "hard": |
|
|
mixed_target = torch.clamp(target + target[perm, :], min=0, max=1) |
|
|
else: |
|
|
raise NotImplementedError( |
|
|
f"mixup_label_type: {mixup_label_type} not implemented. choice in " |
|
|
f"{'soft', 'hard'}" |
|
|
) |
|
|
|
|
|
return mixed_data, mixed_target |
|
|
else: |
|
|
return mixed_data |
|
|
|
|
|
|
|
|
def add_noise(mels, snrs=(6, 30), dims=(1, 2)): |
|
|
""" Add white noise to mels spectrograms |
|
|
Args: |
|
|
mels: torch.tensor, mels spectrograms to apply the white noise to. |
|
|
snrs: int or tuple, the range of snrs to choose from if tuple (uniform) |
|
|
dims: tuple, the dimensions for which to compute the standard deviation (default to (1,2) because assume |
|
|
an input of a batch of mel spectrograms. |
|
|
Returns: |
|
|
torch.Tensor of mels with noise applied |
|
|
""" |
|
|
if isinstance(snrs, (list, tuple)): |
|
|
snr = (snrs[0] - snrs[1]) * torch.rand( |
|
|
(mels.shape[0],), device=mels.device |
|
|
).reshape(-1, 1, 1) + snrs[1] |
|
|
else: |
|
|
snr = snrs |
|
|
|
|
|
snr = 10 ** (snr / 20) |
|
|
sigma = torch.std(mels, dim=dims, keepdim=True) / snr |
|
|
mels = mels + torch.randn(mels.shape, device=mels.device) * sigma |
|
|
|
|
|
return mels |
|
|
|
