[Tip] Remember epsilon when you make RMSE using Pytorch

[jinmyeongAN]

RMSE problem

• RMSE is supposed to be like this.

  rmse = torch.sqrt(torch.nn.MSELoss(logits, y.float()))

  • But, torch.sqrt(0) return NaN

Solution

• So, you should add any extremely small number aka epsilon Here's the solution

eps = 1e-6
rmse = torch.sqrt(torch.nn.MSELoss(logits, y.float()) + eps)