Motion prediction along the z axis about pf-track HOT 4 CLOSED

@zen-d Thanks for your question! I only use the motion on the BEV for updating reference points (

from pf-track.

@ziqipang Thanks for your prompt reply! Maybe I would like to further make myself understood. The last layer of future_reg is an FC with the weight matrix of shape $256 \times 3$, but it seems that only the first two dimensions of its output are extracted for loss calculation
https://github.com/TRI-ML/PF-Track/blob/main/projects/tracking_plugin/models/trackers/tracker.py#L243-L247

Based on my experience, PyTorch DDP will raise an error when launching the training, since the weights of future_reg do not all receive gradients (only a $256 \times 2$ matrix do), i.e., "unused parameters" exist. However, I did not meet such an error. The underlying reason is actually what I am curious about.

from pf-track.

@zen-d Yeah, I agree with you. I am also surprised after noticing what you mentioned to me. I tend to think this is a behavior of DDP and not so familiar with it. Perhaps you could have a minimal example of exploring this and ask some experts on the Pytorch forum.

from pf-track.

@ziqipang After second thought, I realize that as a whole tensor, some entry of the FC weight matrix just receives zero gradients, instead of not receiving grad (thereby without triggering any error). It might not be so useful but correct. Thank you.

from pf-track.