Bug when calculating pre-emphasis filter for plotting

[ivanpondal]

Hello!

When plotting, the following function is not calculating correctly the pre-emphasis filter:

PedalNetRT/plot_wav.py

Line 24 in 238aca1

return np.concatenate([x, np.subtract(x, np.multiply(x, coeff))])

The correct implementation is being used in the model:

PedalNetRT/model.py

Line 106 in 238aca1

return torch.cat((x[:, :, 0:1], x[:, :, 1:] - coeff * x[:, :, :-1]), dim=2)

At high level this is given a signal x, in order to calculate x[t] you do:
x[t] = x[t] - 0.95*x[t - 1]

btw, great project! 👍

[GuitarML]

@ivanpondal Thanks for catching that, I will check it out, and I'll also check that in the GuitarLSTM repo because I might be making the same mistake there.

[GuitarML]

@ivanpondal Sorry for the delay in responding to this, but if you compare the two implementations, aren't they doing the same thing? The one in plot code is using numpy arrays and the one in the model is using pytorch tensors.

[ivanpondal]

Hey! The problem is it's not using the previous element for the substraction. The pytorch implementation does:
x[:, :, 1:] - coeff * x[:, :, :-1]
That list slicing produces the desired x[t] = x[t] - 0.95*x[t - 1] while the numpy implementation does the substraction over the same list element.

[GuitarML]

Got it now, good catch!