Open idontlikelongname opened 2 years ago
Could you briefly summarize what you changed so that I can help you find the potential problems? And the config you provide and the log are inconsistent, i.e. you set num_proposals=100
in config but the log shows num_proposals=300.
. This is a critical hyperparameter for TransFusion because we need the number of proposals to be larger than that of the ground truth, otherwise, the label assignment step will be quite noisy. Please use 300 or more num_proposals for waymo dataset.
Why nuscenes loss_bbox weights is 0.25, but waymo is 2.0, its make the waymo box loss very big, over 10.0 at last.
We set the loss_bbox of waymo to 2 following CenterPoint, because on Waymo the network should care more about the bbox prediction for higher mAP calculated using 3D IoU (and also the classification is relatively easy). The absolute value of bbox loss does not matters.
Thanks for your great works.
Im training TransFusion_L (lidar only) with waymo datasets. but the loss is abnormal, in particular the box size loss. Also the box size in test reults is weird. Why nuscenes loss_bbox weights is 0.25, but waymo is 2.0, its make the waymo box loss very big, over 10.0 at last.
the following is my codes, config and training logs, I simplify the TransFusionHead to make it easy to understand for me.