A small question of implementation

[TianQi-777]

Thank you for your sharing, but I have a small question. Why do you 1) use F.softplus for variance(std_dev) every time and 2) add a constant(min_std_dev). Is it to ensure that the square of variance is positive? If so, why not use Relu? (click here 1, click here 2) Example as follows:

prior_std_devs[t + 1] = F.softplus(_prior_std_dev) + self.min_std_dev

Looking forward to your reply.