Hi,
I’ve been trying for a while to compress Tensorflow Object Detection API models from the model zoo for TF2 to TensorRT and deploy them in my Jetson Nano.
I’ve been using the official conversion repo . I’ve been able to install and run this repo on my PC, but to either run the resulting .ONNX or .TRT files on Jetson I need to have the TFOD Object Detection API installed.
When trying to install, both tensorflow-addons and tensorflow-text have been impossible to install, even from source. I’ve tried to follow both this issue and this issue, but I guess I’m missing some step.
Could you provide a full guide on how to compress TFOD API models to TensorRT on Jetson?
Thanks.
Hi,
Have you tried to convert the ONNX model on a desktop environment, and create the engine file from ONNX on Nano?
Thanks.
Yes, I get this error when I try:
Should I retrain the model somehow?
Hi,
How do you create the engine file?
Could you try it with trtexec binary?
$ /usr/src/tensorrt/bin/trtexec --onnx=[model]
Thanks.
Hi,
I was using this script (as far as I know the official one) for the engine file.
When I use trtexec I get:
Hi,
By compressing again the model to ONNX and using trtexec I managed to build a .trt file. No my issue is that I can’t run the model, neither using trtexec nor the inference script. The error is the following:
Hi @AastaLLL , could you please give me a hand on this? Thanks.
Hi,
Sorry for the late update.
Would you mind running the trtexec tool with --verbose and sharing the output with us?
More, which JetPack version do you use?
If you are not using 4.6.2, would you mind also giving it a try?
Thansk.
Hi,
I opened a new issue and they helped me out here: Can't execute TRT engine on Jetson Nano - #2 by naisy
I was able to make it run. Sadly now I have JetPack 4.6.1 installed and I can’t go back to 4.6 because my access to the Jetson is limited and I need it to work.
Anyway, thanks for everything. You can close the issue if you want.
