Some problem in TargetAssigner

[Liheng]

I am trying to train my own lidar dataset(9 classes) on the repo.
And here seems something wrong in tensor shape mismatch which occured in ProposalTargetassigner part.

matches = torch.stack(matches).view(self.anchors.shape[:-1])
match_labels = torch.stack(match_labels).view(self.anchors.shape[:-1])

here is my log:

what's the meaning of full_idx in get_matches func?
@jhultman

[Liheng]

in multi-class training , the target assigner code have some problem.
https://github.com/jhultman/PV-RCNN/blob/ca40082be65531c03dfda7353f6eb2d6d47ebb64/pvrcnn/core/proposal_targets.py#L68

because per frame can't contain all class obstacle, so the matches and matches_label can't be just view as [class_num, yaw_num, x_range, y_range].

matches = torch.stack(matches).view(self.anchors.shape[:-1])
match_labels = torch.stack(match_labels).view(self.anchors.shape[:-1])

[Jacob Hultman]

Hi, thanks for reporting this. I have just pushed a fix -- let me know if it works for you.

full_idx is an indexing trick which maps the masked ground truth boxes for class i back to the full set of ground truth boxes (containing all classes). Sorry the naming is poor. Here is an example which I hope explains it.

import torch
i = 2
class_idx = torch.LongTensor([0, 2, 1, 0, 2])
full_idx = torch.arange(class_idx.shape[0])
class_mask = class_idx == i
matches = torch.LongTensor([0, 1, 0])
matches = full_idx[class_mask][matches]
print(matches)
>>> tensor([1, 4, 1])

[Liheng]

seems like the tensor are not in GPU
the full_idx tensor need to load in GPU, that can resolve this problem.

[Liheng]

the newest repo code can works in my own dataset now.
thanks