Result in the paper

[Kguo-cs]

Could you provide the pretrained model for reproducing your close-loop result in the paper? Thank you.

[MCZhi]

Sorry, due to restrictions of the Waymo Dataset License Agreement, we are not allowed to share the trained model. If you have any questions when reproducing the training process, please feel free to open an issue and contact us.

[Kguo-cs]

Thank you for your answer. Could you tell me which 100 replay scenes do you use for testing?

[Kguo-cs]

And the 10156 driving scnes for training.

[MCZhi]

The replay scenes are from uncompressed_scenario_training_20s_training_20s.tfrecord-00299-of-01000 and uncompressed_scenario_training_20s_training_20s.tfrecord-00306-of-01000 files. For training, we randomly select 90 files from the datasets (around 10k scenes and 113k data points), and for validation, we randomly select 20 files other than the training files.

[Kguo-cs]

Thank you for your help again. But I find there are 75 scenes from uncompressed_scenario_training_20s_training_20s.tfrecord-00299-of-01000 and 58 scenes from uncompressed_scenario_training_20s_training_20s.tfrecord-00306-of-01000. The sum is 133 (more than 100). How do you handle it?

[MCZhi]

I manually removed some scenes that are not of interest such as waiting at a red light and keep 100 scenes for evaluation.