search

myc634 / UltraLiDAR_nusc_waymo

MIT License
45 stars 3 forks source link

lidar_rec_diff explaination #11

Closed Pamphlett closed 3 months ago

Pamphlett commented 3 months ago

Hi,

Thanks for open-sourcing this work.

I am training on NuScenes and have a question regarding the lidar_rec_diff defined here. As the two variable lidar_rec_prob and voxels have totally different scales, making the lidar_rec_diff very big during the whole training (as shown below). Am I missing something here?

2024-08-27 13:50:18,358 - mmdet - INFO - Epoch [100][2300/2345] lr: 9.972e-07, eta: 0:00:25, time: 0.542, data_time: 0.003, memory: 15178, loss_lidar_rec: 0.0924, loss_emb: 0.0656, lidar_rec_diff: 8519.0909, lidar_rec_iou: 0.5504, code_util: 1.0000, code_uniformity: 0.0516, loss: 0.1579, grad_norm: 0.0333
myc634 commented 3 months ago

Hi,

lidar_rec_diff has different scale with other metrics, no worries. Basically we check lidar_rec_iou to make sure it is converged.