Really nice work!
I tried to read the code, and found an issue that the implementation is a little different from that in equation (10).
In the repo, it's
def _std(self, t):
# This is a full solution to the ODE for P(t) in our derivations, after choosing g(s) as in self.sde()
sigma_min, theta, logsig = self.sigma_min, self.theta, self.logsig
# could maybe replace the two torch.exp(... * t) terms here by cached values **t
return torch.sqrt(
(
sigma_min**2
* torch.exp(-2 * theta * t)
* (torch.exp(2 * (theta + logsig) * t) - 1)
* logsig
)
/
(theta + logsig)
)```
in `sgmse/sdes.py`.
Can you explain the differences? Thanks!
Dear authors,
Really nice work! I tried to read the code, and found an issue that the implementation is a little different from that in equation (10).
In the repo, it's