Hello, I trained model EfficientDet D0 using Tensorflow Object Detection API using a remote server. Then, I freeze the model, load the trained model into the jetson TX2, and performed the inference of the model. I was able to execute the model at 5 FPS in my Jetson TX2. However, I read a paper in which the author use the same hardware (Jetson TX2) and the same model (EfficientDet D0) but he achieves 13 FPS for model inference. It is important to mention that in both cases we are not using TensorRT, so I am not sure why he achieves more FPS using the same hardware and model.
For perfoming the Inference of the model in my Jetson TX2 I:
- Enable the two Denver CPUS that were disable.
- Activate the Jetson clocks.
- Set the Jetson to mode MAXN.
With the aforementioned configuration my jetson tx2 is working at full capacity? Is there a way to improve the FPS of the model inference using another configuration of the Jetson TX2?
Thank you in advance for the support.
TensorRT acceleration is not applied automatically but needs some conversion.
We have an example for running EfficientDet D0 with TensorRT.
Would you mind giving it a try to see if any performance improvement?
I cannot use TensorRT because my Jetson TX2 has Jetpack 4.5.1 installed and this version has TensorRT 7.1.3, which does not support EfficientDet. I searched on the internet if it is possible to upgrade the TensorRT to 8.0.1, however, in many forums they affirm that the only way to upgrade is to change the version of Jetpack (In my case I would have to replace my Jetpack 4.5.1 with Jetpack 4.6.0). I wouldn’t like to upgrade my current Jetpack because I will lose all the setup (dependencies installation) of my project. It would be great to upgrade my current Jetpack version without losing all my setup.
On the other hand, the question of my previous post is more related to the hardware configuration of the Jetson TX2. In that post I noted that an author of a paper achieves (13 FPS) using the Jetson TX2 and EfficientDet D0. However, I used the same hardware (Jetson TX2) and model (EffcientDet D0) and only achieved 5 FPS. I am not sure why there is a gap of 8 FPS from his implementation and mine if we are using the same model and hardware, the only think that I could imagen is that I am not using the Jetson TX2 at its maximum capacity. Maybe I am not using the Jetson’s GPU at max capacity?
Thank you AastaLLL for the help
Based on your statement, you may not use the full GPU resources for inference.
You can confirm this by running tegrastats to check the GPU loading.
In general, it’s recommended to use TensorRT since it has optimized for Jetson.
For other frameworks, they might only optimize the performance for desktop.
So you will see some performance gaps by using TensorFlow and TensorRT.
This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.