TensorRT6 and Tensorflow Faster RCNN models

AI & Data Science Deep Learning (Training & Inference) TensorRT
1

The new release of TensorRT supports Tensorflow Faster RCNN models as in [url]https://github.com/NVIDIA/TensorRT/tree/master/samples/opensource/sampleUffFasterRCNN[/url].

Is it possible to use the Faster RCNN models from the object detection model zoo as in this sample?
Could you provide a config.py file that maps the layer in the graphs to the correct plugins?
Thanks.

2

up

3

I have already posted this on the forums. Still no answer…

I would like to add to this question:
If I wanted to use this tensorRT implementation with Python, is the only approach still to use SWIG to create a wrapper around this FasterRCNN engine?

Same pipeline as : https://github.com/AastaNV/ChatBot

Thank you.

4

I don’t know what your specific use case is but, in general, you don’t need SWIG to use the engine.
Check the example engine_refit_mnist in /usr/src/tensorrt/samples/python

Your pipeline should be:

frozen_graph ----> uff -----> engine -----> inference in python with engine

The problem is in the first step, as FasterRCNN is not fully supported yet.

5

Thanks for the reply @nick0lai

You have brought up a good point…To probe a bit more…

I only thought I needed to use SWIG because there isn’t a supported Python API for ARM (jetson devices) as of yet. See this post:

For using the engine, how would you create on with python?
something like? :

engine = trt.utils.uff_to_trt_engine(G_LOGGER, uff_model, parser, 1, 1 << 20)

I grabbed this from the tutorial on the workflow seen at:
https://docs.nvidia.com/deeplearning/sdk/tensorrt-archived/tensorrt_401/tensorrt-api/python_api/workflows/tf_to_tensorrt.html

If I am able to do this all with python, that would be fantastic!

6

That was an old post in the forum.
In the recent Jetpack for Jetson you have the python API.

You should be fine for the second part, the examples in /usr/src/tensorrt/ are pretty straightforward.

7

Issue is fixed in latest version.

8

@SunilJB is there any official examples of this besides the ones that were talked about earlier ? Or should I just look at the tensorrt examples in the latest install?

9

Hi,
Please reference the TensorRT sample support guide for examples:

Regarding TRT 6 support on JetPack:
JetPack 4.3 Developer Preview, which is packaged with TRT6 is available as a beta. JetPack 4.3 Developer Preview | NVIDIA Developer

Stay tuned for official production release of Jetpack 4.3.

Thanks