Losses: Cutoff MSE

class losses.CutoffMSE(cutoffValue=1)

Description:

Cutoff MSE loss has two steps:

  1. Filtering:
    Filter the noisy labels out by comparing their noise variances to a threshold. Remove the label if its noise variance is bigger than a threshold, otherwise, use it to compute the loss.
  2. MSE loss:
    Compute the mse loss using the filtered labels.
Args:
threshold:An upper bound noise variance threshold to filter out the noisy labels.
forward(y_pred, y, lbl_var)
Description:
Compute the forward pass.