plot_sample (a:numpy.ndarray, b:numpy.ndarray, gt_audio:numpy.ndarray, sample_rate:int=16000, plt_ioff:bool=True)
Plot a sample from the model.
db20 (x:numpy.ndarray)
Plot a dummy image
b = np.array([[[1.0, -1.60, 0.80]]])
a = np.array([[[1.0, -0.90, 0.27]]])
print(a.shape, b.shape)
impulse = np.zeros(int(0.3 * 16000))
impulse[0] = 1.0
gt_audio = np.random.randn(int(0.3 * 16000) + 1)
fig = plot_sample(a, b, gt_audio)
fig(1, 1, 3) (1, 1, 3)(<Figure size 600x300 with 1 Axes>,
array([ 1.e+000, -7.e-001, -1.e-001, ..., -5.e-324, -5.e-324, -5.e-324]))save_and_display_audio (audio:numpy.ndarray, name:str, sample_rate:int=44100)
Save audio file and display in a Jupyter notebook.
fft_loss (pred_fft:torch.Tensor, target_fft:torch.Tensor, lin_l1:float=1.0, lin_l2:float=0.0, log_l1:float=0.0, log_l2:float=0.0)
Compute the loss between the target and predicted fft
| Type | Default | Details | |
|---|---|---|---|
| pred_fft | Tensor | magnitude of the predicted fft | |
| target_fft | Tensor | magnitude of the target fft | |
| lin_l1 | float | 1.0 | weight of the linear l1 loss |
| lin_l2 | float | 0.0 | weight of the linear l2 loss |
| log_l1 | float | 0.0 | weight of the log l1 loss |
| log_l2 | float | 0.0 | weight of the log l2 loss |
| Returns | Tensor | scalar loss |
FFTLoss (lin_l1:float=1.0, lin_l2:float=0.0, log_l1:float=0.0, log_l2:float=0.0)
Compute the loss between the target and predicted fft
| Type | Default | Details | |
|---|---|---|---|
| lin_l1 | float | 1.0 | weight of the linear l1 loss |
| lin_l2 | float | 0.0 | weight of the linear l2 loss |
| log_l1 | float | 0.0 | weight of the log l1 loss |
| log_l2 | float | 0.0 | weight of the log l2 loss |
MelScaleLoss (n_mels:int=128, sample_rate:int=16000, n_fft:int=0, lin_l1:float=1.0, lin_l2:float=0.0, log_l1:float=0.0, log_l2:float=0.0, f_min:float=0.0, f_max:Optional[float]=8000.0)
Compute the loss between the target and predicted mel scale
| Type | Default | Details | |
|---|---|---|---|
| n_mels | int | 128 | |
| sample_rate | int | 16000 | |
| n_fft | int | 0 | |
| lin_l1 | float | 1.0 | weight of the linear l1 loss |
| lin_l2 | float | 0.0 | weight of the linear l2 loss |
| log_l1 | float | 0.0 | weight of the log l1 loss |
| log_l2 | float | 0.0 | weight of the log l2 loss |
| f_min | float | 0.0 | |
| f_max | typing.Optional[float] | 8000.0 | |
| Returns | None |
to_zpk (x:torch.Tensor, n_parallel:int, n_biquads:int, gain_factor:float=1.0)
Convert the parameters of the filter to zeros, poles and gain
Input: x: (B, n_parallel, n_biquads, 5) Output: z: (B, n_parallel, n_biquads) p: (B, n_parallel, n_biquads) k: (B, n_parallel, n_biquads)