search

MCG-NJU / SparseOcc

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

loss_single 的疑问 #36

Closed FreqBayes closed 3 months ago

FreqBayes commented 3 months ago

Great Works! 训练时发现 models/sparseocc_head.py 中的 SparseOCCHead.loss_single() 第97行，好像不太对，会导致后面匈牙利匹配中num_instances为空。调查之后可能是因为pytorch-1.10，torch.bincount(voxel_instances[b]) 如果 voxel_instances 中数据为 uint8 则结果异常。这里能不能改成 voxel_instances[b].long() 或者别的格式？

afterthat97 commented 3 months ago

确实，我们也发现了这个bug，预计下个版本修复 建议你这几天先用torch 2.0，新版torch没有这个问题

afterthat97 commented 3 months ago

修复了，感谢指出