You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
I'm a little confused about the workings of the asymmetricSmoothLoss function. It looks like you compute smoothness losses for the valid (nonOccSmooth) and invalid sections (occSmooth) separately. However, at the end of the function, the sum of the losses is multiplied by the corrected valid mask. Doesn't this essentially eliminate the occSmooth losses? From my initial reading, it seems like the correct output of the function should be nonOccSmooth + (1.-valid) * occSmooth.
I'm a little confused about the workings of the asymmetricSmoothLoss function. It looks like you compute smoothness losses for the valid (nonOccSmooth) and invalid sections (occSmooth) separately. However, at the end of the function, the sum of the losses is multiplied by the corrected valid mask. Doesn't this essentially eliminate the occSmooth losses? From my initial reading, it seems like the correct output of the function should be nonOccSmooth + (1.-valid) * occSmooth.
unsupFlownet/src/util/lossComponents/asymmetricSmoothLoss.py
Line 31 in 3f652d7
The text was updated successfully, but these errors were encountered: