Increase fps on Jetson TX2 using Tensorflow and TensorRT

Hi!

I’m a student working on a project and I need to increase my fps for object recognition.
I am using Tensorflow 1.14 on a Jetson TX2 with maximized performances (./jetson_clocks.sh).
I’ve already tried TensorRT for optimizing my graph, that is the output of a 200’000 steps training with Faster RCNN.
For now, the maximum i can achieve in terms of fps is around 4.5.

Is there any chance to increase these value even more?

Thanks.

Hi,

The profiling result is similar to ours.
We got 3.13 fps on FP32 mode and 5.22 fps with FP16 mode.

By the way, FasterRCNN can reach 26.54 fps on Xavier with FP16 mode.

Thanks.

Thanks, we are using FP16 too! Boosting fps to 5.22 would be great for our application.

I’m posting the parametres that i’m using in Tensor RT, maybe there is some new combination that i can try to boost my fps from 4.5 to 5.2!

Converter constructor parameters for optimization

nodes_blacklist = [“detection_boxes”, “detection_scores”, “detection_classes”] #output nodes
max_workspace_size_bytes = 1 << 30
precision_mode = “FP16”
minimum_segment_size = 3
maximum_cached_engines = 1

Thanks again!