What's the correct approach to port a point cloud TensorFlow model to Jetson HW?

Robotics & Edge Computing Jetson & Embedded Systems Jetson Nano
,
1

Description

Hi guys,

I am trying to port a point cloud model (developed in Tensorflow): GitHub - yulequan/PU-Net: PU-Net: Point Cloud Upsampling Network, CVPR, 2018 (https://arxiv.org/abs/1801.06761)
to run using TensorRT on Jetson hardware.

  1. What’s the correct approach?
  2. Should I use TF-TRT or try to port to pure TRT?
  3. The model is using custom ops written in c++ via Tensorflow’s tf.load_op_library()

I have seen the tutorials on TensorRT like: NVIDIA-AI-IOT/tf_to_trt_image_classification but most of them are for well-known image models, not sure if this approach will work with point cloud model that I mentioned?

Relevant Files

2

Hi,

We are moving this post to the Jetson Nano forum to get better help.

Thank you.

4

Hi,

Pure TensorRT can give you a much better performance on Jetson.
However, all the layers used in the model need to be supported on the TensorRT:
Or you will need to implement it manually with plugin API.

If there are some non-supported layers, you can try TF-TRT since it can leverage the implementation from the TensorFlow side.

Thanks.

6

This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.