Np.unwrap function in pytorch

1

Is there same function like numpy.unwrap in pytorch?

I want to implement hilbert transform and extract envelope & Temporal fine structure in pytorch with GPU.

I failed to find the same pytorch function like “scipy.hilbert”, “np.phase”, “np.abs(complex value)”. Now, I successfully made these three function myself.
But, still, I cant realize “np.unwrap” function efficiently.
Because of this, my model waste lots of time in data pre-processing.

Anyway I need a help. thanks

2

I’m not sure if I understand the np.unwrap function correctly, but does something like

def unwrap(x):
    y = x % (2 * pi)
    return torch.where(y > pi, 2*pi - y, y)

work?

3

‘unwrap’ can be great to have when working with sin()'s and cos()'s. ‘diff’ can be useful in general. Both ideally should eventually be included in pytorch.

Here is a pytorch port of numpy’s implementation:

def unwrap(phi, dim=-1):
assert dim is -1, ‘unwrap only supports dim=-1 for now’
dphi = diff(phi, same_size=False)
dphi_m = ((dphi+np.pi) % (2 * np.pi)) - np.pi
dphi_m[(dphi_m==-np.pi)&(dphi>0)] = np.pi
phi_adj = dphi_m-dphi
phi_adj[dphi.abs()<np.pi] = 0
return phi + phi_adj.cumsum(dim)

def diff(x, dim=-1, same_size=False):
assert dim is -1, ‘diff only supports dim=-1 for now’
if same_size:
return F.pad(x[…,1:]-x[…,:-1], (1,0))
else:
return x[…,1:]-x[…,:-1]

4

NOTE: the call to ‘diff’ should have same_size=True