Resent50 performance question

I have a performance issue question for a Jetson TX2.
I am running a MATLAB GPU coder example that uses resnet50 to classify streaming video from a webcam.

When running the code, I am only able to classify video at a rate of about 1.5 frames per second.

A colleague runs the same code on a Jetson Nano and is able to classify video at about 15 frames per second.

Are there some settings that I could check to help improve performance? This seems … real slow.

After removing the display-over-the-image code and adding in a ‘fprintf()’ statement, I’m showing about 1.6 frames per second.
This leads me to think that the issue is with the resnet50 network running on my TX2.


Based on the performance table, TX2+ResNet50 can reach 84fps with the TensorRT library.

It seems there is some slowness that comes from the MATLAB library or your implementation.
Have you checked this issue with the MATLAB team before?


I was able to change a couple of parameters to specify that the ‘TensorRT’ library be used in the " coder.DeepLearningConfig()" parameter and specify that this is a “cfg.GpuConfig.ComputeCapability = ‘6.2’;”.
This allowed me to fun my code at ~38FPS.