CenterPose
CenterPose copied to clipboard
Question about obj_scale_loss
Hello, thank you for the nice work!
I have a question about obj_scale_loss
.
Why do you use different forms of the scale loss in training and validation phase?
Specifically, in the training phase:
https://github.com/NVlabs/CenterPose/blob/6c89d420b33bd01c14c13f509af08bfe3d8b2fe7/src/lib/trains/object_pose.py#L116-L117
https://github.com/NVlabs/CenterPose/blob/6c89d420b33bd01c14c13f509af08bfe3d8b2fe7/src/lib/models/losses.py#L167
and in the validation phase:
https://github.com/NVlabs/CenterPose/blob/6c89d420b33bd01c14c13f509af08bfe3d8b2fe7/src/lib/trains/object_pose.py#L126-L128
https://github.com/NVlabs/CenterPose/blob/6c89d420b33bd01c14c13f509af08bfe3d8b2fe7/src/lib/models/losses.py#L174-L176
torch.abs(target * mask - pred * mask)
and torch.abs((1 * mask - pred * mask) / target_rmzero)
does not produce same values.
I want to know the meaning of the "relative loss" in the validation phase and why it is only used in the validation phase.