xinge008 / Cylinder3D

Rank 1st in the leaderboard of SemanticKITTI semantic segmentation (both single-scan and multi-scan) (Nov. 2020) (CVPR2021 Oral)
Apache License 2.0
859 stars 180 forks source link

In training loss #141

Open akshay-antony opened 2 years ago

akshay-antony commented 2 years ago

Is there a need to convert sparse to dense for loss calculation and backward pass, or can we use the inverse map and calculate loss between two sparse tensors i.e find argmax without using logits.dense() and compare with pointwise labels. Will this work too?