TensorRT for Non-Classification Deep Learning


I want to convert a pb model for a custom superresolution algorithm to TensorRT and run it on Nvidia AGX device. Every tutorial I come by is about “custom classification” examples. I need guidance on:

  • how to convert a (preferably superresolution) model to tensorRT, I can only convert models with input sizes equals to the first layer filter size, what does it even mean?
  • how to wrap the .engine so that I can feed input/output correctly, do I need to feed 5x5 chunks or what else?
  • how to work with a TRT model without needing “deepstream-app”, is it even possible?

Thanks in advance,


You can find below a super-resolution example:

1. For a frozen pb file from TensorFlow v1.x, please convert it via .pb->uff->TRT flow.
Other, please convert the file into ONNX format first.

**2.**In general, the input size is identical to the network input size.
For example, YOLO uses 608 or 416 as input.

3. Yes. You can do it with other multimedia libraries, ex. OpenCV.
Below is an example of TensorRT+OpenCV for your reference.


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