discard negative pad

Author: yxlllc

models/nsf_HiFigan/models.py

@@ -129,7 +129,7 @@ def _f02sine(self, f0, upp):
         rad = f0 / self.sampling_rate * torch.arange(1, upp + 1, device=f0.device)
         rad2 = torch.fmod(rad[..., -1:].float() + 0.5, 1.0) - 0.5
         rad_acc = rad2.cumsum(dim=1).fmod(1.0).to(f0)
-        rad += F.pad(rad_acc, (0, 0, 1, -1))
+        rad += F.pad(rad_acc[:, :-1, :], (0, 0, 1, 0))
         rad = rad.reshape(f0.shape[0], -1, 1)
         rad = torch.multiply(rad, torch.arange(1, self.dim + 1, device=f0.device).reshape(1, 1, -1))
         rand_ini = torch.rand(1, 1, self.dim, device=f0.device)
@@ -249,7 +249,7 @@ def fastsinegen(self, f0):
         rad = s0 * n + 0.5 * ds0 * n * (n - 1) / self.upp
         rad2 = torch.fmod(rad[..., -1:].float() + 0.5, 1.0) - 0.5
         rad_acc = rad2.cumsum(dim=1).fmod(1.0).to(f0)
-        rad += F.pad(rad_acc, (0, 0, 1, -1))
+        rad += F.pad(rad_acc[:, :-1, :], (0, 0, 1, 0))
         rad = rad.reshape(f0.shape[0], 1, -1)
         sines = torch.sin(2 * np.pi * rad)
         return sines