Fixing EPS application when applying sqrt

[FedericoDiMarzo]

Problem

• When using the pytorch_stoi loss in some training sessions some NaN values are were appearing
• One of the possible cause would be related to the sqrt operation since the gradient will lead to a singularity
• An EPS is applied in the current implementation but the order of the operations can allow for exploding gradients

Solution

• Sum EPS just before applying sqrt

[mpariente]

Thanks, that's perfect !