waymo-research
There are some issues with scenario_id
I encountered the following error while getting scenoria_id based on the tutorial in the notebooks:
Cannot batch tensors with different shapes in component 5. First element had shape [1,16] and element 1 had shape [1,15]. [Op:IteratorGetNext]
Before this, I passed in my custom data_config and compared it with the official one. The main issue was that the batch_dims were different. When I set batch_dims to 1, the issue was resolved.
Based on the error message, I suspect that the problem occurred when batching the tensors into NumPy arrays, where the lengths didn’t match. I then printed out scenario_id and scenario, and noticed that while the length of scenario was consistent, scenario_id had the shape (1, 1, None), dtype=tf.uint8. Upon further inspection, I saw that scenario_id had 15 characters in some cases and 16 characters in others, which matches the error message exactly!
Therefore, I’d like to remind everyone about potential ID-related errors. You could try setting batch_dims to 1. Also, for better parallelization, I hope the official solution will handle padding for scenario_id lengths. Thank you!
Hi @Zhzdpaaa , thanks for raising this. Do you have a minimal reproduction of the issue that you could share?
