Conversion ResNet50.onnx to TensorRT Engine on Jetson Nano

[darekdev]

Description

I need ResNet50 without top_layer to only feature extraction purpose. I create ResNet50 model (include_top=false) in keras and then export it to ONNX format. All steps are done regards to this tutorial Nvidia TensorRT Notebook

Then I move model to the Jetson Nano and do the engine creation:

/usr/src/tensorrt/bin/trtexec --onnx=/home/jetson/model/resnet50_without_top.onnx --saveEngine=/home/jetson/model/resnet_engine.trt --explicitBatch  --workspace=1024`

but when I do the prediction on jetson nano I get different result then from PC version.

Environment

TensorRT Version:
Jetson Version: nvidia-jetpack Version: 4.4-b144
Operating System + Version: Jetson and Fedora on PC
Python Version (if applicable): 3.7
TensorFlow Version (if applicable): keras

[darekdev]

Solved. I need to set up input shape as below:

/usr/src/tensorrt/bin/trtexec --onnx=/home/jetson/cash15/resnet50_without_top.onnx --saveEngine=/home/jetson/cash15/resnet_engine.trt --explicitBatch --shapes=input_1:1x224x224x3

before this I always get warning:

[06/22/2021-14:57:54] [W] Dynamic dimensions required for input: input_1, but no shapes were provided. Automatically overriding shape to: 1x1x1x3

What is courious that network get image with shape (1, 224, 224, 3) on input and works (with wrong output data).

Best regards.

[AastaLLL]

Hi,

Please add the dynamic shape information to trtexec as well.
For example:

$ /usr/src/tensorrt/bin/trtexec --minShapes=input_1:1x224x224x3 --optShapes=input_1:1x224x224x3 --maxShapes=input_1:1x224x224x3 ...

Thanks.