High latency while run TensorFlow with keras on Jetson Tx2

Im currently running TensorFlow with Keras on Jetson Tx2 but the image i get has very long delay.
Anything change in front of camera will show on the screen about 2-3 seconds later.
Did anyone have the same issue?
The elapsed time of each frame is about 300 ms

Screenshot from 2017-11-14 16-25-30.png

I have tried the following command to speed things up but it was still slow. Before this was even slower

$ sudo nvpmodel -m 0
$ sudo ./jetson_clocks.sh

Hi,

Could you check the execution time for the TensorFlow model first?
Thanks.

Hi thanks for reply
The execution time for tensorflow model is about 250 ms

Hi,

From your profiling result, the bottleneck is the deep network inferencing.

There two possible alternatives:
1. Use TensorRT for inferencing. The acceleration ratio can be found here:

2. Use a more light-weight network.

Thanks.