Request for VAD CARLA Tiny PyTorch checkpoint and ONNX export script

[83030395]

Hello Autoware maintainers,

I am working with autoware_tensorrt_vad and the official VAD CARLA Tiny model package:

• vad-carla-tiny_backbone.onnx
• vad-carla-tiny_head.onnx
• vad-carla-tiny_head_no_prev.onnx
• vad-carla-tiny.param.json

I understand from the documentation and PR #11578 that this model was trained on the Bench2Drive CARLA dataset and is based on VAD. I would like to reproduce or adapt the
model for research/deployment purposes, especially to retrain or fine-tune it on real-world data.

Could you please clarify whether the following artifacts can be made public?

1. The PyTorch checkpoint used to export vad-carla-tiny_head.onnx and vad-carla-tiny_backbone.onnx

2. The training config for the CARLA Tiny model

3. The ONNX export script used to generate:

   • vad-carla-tiny_backbone.onnx
   • vad-carla-tiny_head.onnx
   • vad-carla-tiny_head_no_prev.onnx

4. Any custom symbolic/export code for TensorRT plugin ops such as MultiScaleDeformableAttentionPlugin and RotatePlugin

I checked the public Bench2DriveZoo VAD checkpoints, but the available vad_b2d_base.pth appears to have a different architecture from the Autoware CARLA Tiny ONNX model. For
example, vad_b2d_base.pth uses a 200 x 200 BEV embedding and 6 transformer layers, while the released vad-carla-tiny_head.onnx appears to use a 106 x 200 BEV embedding and 3
transformer layers.

Because of this, it does not seem possible to reproduce the released ONNX model from the currently public checkpoint.

If the checkpoint or export script cannot be released, could you share the recommended procedure for training and exporting a compatible VAD CARLA Tiny model for Autoware?

Thank you for your help.

Best regards,
long