Hello,
our model depends on some 3D-Convolutions (Keras 3DConv). Right now, TensorRT does not support these operations on the DrivePX2, right?
What is a possible workaround to do inference using the Driveworks SDK?
Hi,
You can implement TensorRT non-supported layer with plugin API and this is available from DRIVE SW 8.0.
Here is an example of conv3D plugin for your reference:
[url]https://github.com/NVIDIA-AI-IOT/redtail/tree/master/stereoDNN/lib[/url]
Unfortunately, DRIVE SW8.0 is only available for Xavier based platform.
If this is a possible option for you, it’s recommended to upgrade your system to DRIVE AGX platform to unblock your task.
Thanks.