search

MCG-NJU / SparseOcc

[ECCV 2024] Fully Sparse 3D Occupancy Prediction & RayIoU Evaluation Metric
https://arxiv.org/abs/2312.17118
Apache License 2.0
235 stars 16 forks source link

questiona about instance_class_ids #6

Closed J6mesQian closed 5 months ago

J6mesQian commented 5 months ago

Hi, I found out that the instance_class_ids during the training are all empty lists, which caused the program to abort. I was wondering if you have encountered this problem before. Thanks!

(Pdb) instance_class_ids [tensor([], device='cuda:0', dtype=torch.int64), tensor([], device='cuda:0', dtype=torch.int64), tensor([], device='cuda:0', dtype=torch.int64), tensor([], device='cuda:0', dtype=torch.int64), tensor([], device='cuda:0', dtype=torch.int64), tensor([], device='cuda:0', dtype=torch.int64), tensor([], device='cuda:0', dtype=torch.int64), tensor([], device='cuda:0', dtype=torch.int64)]

afterthat97 commented 5 months ago

instance_class_ids belongs to the panoptic version. I released this part a few minutes ago, please check it!

YANG-CY-163 commented 5 months ago

This is caused by the incorrect behavior of torch.bincount when computing loss(may occur in PyTorch 1.10), refer to this . Using a higher version of PyTorch(e.g. 2.0.0) can solve this problem, hope it helps you.

J6mesQian commented 5 months ago

update torch does solve this, thx!!