Zebin Xing
@JunrQ We did not experiment with other rotation counts, so this setting was not specifically optimized.
1. The open-source code is the complete version, and the parameters in the shell script are already configured. 2. When I train on a 4090 GPU with batch_size=2, it already...
@Bryceder, I haven't encountered this. But maybe you can check it from the aspects below. 1. It saved the ckpt every 5 epochs. Have it run 5 epochs? 2. Is...
@ssuralcmu Hi, this will take more than one day.
@wangrocket Hi, you can refer to `/scripts/generate/run_generate_dac_label.sh` to get the dac_score label and `/navsim/agents/goalflow/goalflow_features.py` to find how to add the dac_score label to the feature dict.
@fengjiang5 Yes, in training, we have tried two different settings: using the gt goal point (end point of the gt trajectory) and the goal point predicted by the goal point...
@frkmac3 Yes, you should first infer a goal point, and then the trajectory.
@popov1212 Sorry for the delayed reply. It may happen when training the whole model directly. Have you tried first training the perception, and then training the trajectory decoder?
@popov1212 Is the perception module frozen or not? If the perception module is frozen, such situation could indeed occur.
@mono0826 The vocab is generated by using k-means on endpoints of the gt-trajectory based on navtrain. You can refer to [issue 5](https://github.com/YvanYin/GoalFlow/issues/5) for details.