How did you solve the problem that the loss is always NaN when training SV and MV with ESAM?

[qiufeng2013]

How did you solve the problem that the loss is always NaN when training SV and MV with ESAM?
And did you really successfully reproduce the ESAM paper?

[VERITAS12]

I fixed a bug in the original code related to boundary-case handling, which was causing numerical issues and NaN losses—after the correction, the training no longer produces NaNs. As I mentioned in the README, due to randomness in the data generation process, it’s difficult to reproduce the paper results exactly; I can only achieve performance that is close to the reported numbers.