Jetson nano UFF convert error and non-supported layer

Autonomous Machines Jetson & Embedded Systems Jetson Nano
1

Dear All,

My jetson nano’s working information is below:

  1. SD card image : jetson-nano-sd-r32.1-2019-03-18
  2. python3.6 + tensorflow
  3. SD card image has Tensorrt : /usr/src/tensorrt

My DNN model is,
Tensorflow frozen graph : has a tf.slim(). It contained conv2 node has ‘Abs’ operator.

I converted it with Python UFF.

After using TensorRT inference, error occurred.
“ERROR: UFFParser: Parser error: pnet/conv1/Abs: Unary not supported for other non-constant node”

I found answer that error message on this forum, “NVidia: Please wait for our announcement for the future release.”

So, I use custom plugin for UFF parser as TensorRT doucment (4.1.2.Example 2: Adding a Custom Layer that is not supported in UFF Using C++)

Problem:

  1. On Jetson nano, no where ‘convert-to-uff’ binary file. How can I get this ?
  2. If not exist it, must use Python ? and then, Can I use C++ code for inference ?

And, When is the release for ‘Unary not supported…’ ?

Thanks.

2

Hi,

This is known issue and you can get more information in this topic:
https://devtalk.nvidia.com/default/topic/1038303/jetson-tx2/unary-op-support-from-uffparser-issue-under-tensorrt-4-0/

In summary, the fix in included in the TensorRT5.1 and will be integrated into JetPack4.2.1.
Please pay attention to our announcement for the release.

1. convert_to_uff.py is located at /usr/lib/python3.6/dist-packages/uff/bin/.
2. Please use python to convert .pb file into .uff.
But you can create TensorRT engine from uff with either python or C++.

Thanks.