Hello,
we have developed a python program which loads models based on vgg19 net and uses a tracking algorithm in opencv. The algorithms are applied to 8 short-duration videos (about 10 seconds) with frames of 800x600 size where the cnn model is applied to the first frame and the tracking algorithm to the others.
Working with the jetson xavier nx we encountered some problems that are difficult to reproduce, but related to the fact that sometimes processing freezes (for example, the screen freezes for a couple of seconds while using a screen sharing program such as Discord).
To try to reproduce this problem, we ran the python code mentioned above 100 times and we noticed that for 93 times the expected computation time is around 300 seconds (which is what we expect) while for 7 times in a random way ( neither at the beginning of the 100 elaborations nor at the end and one after the other) the elaboration times are quite high starting from about 1000 to about 2500 seconds.
How to solve this problem? how to check if something went wrong in the installation of the jetson environment? How to prevent this inconvenience?
This is some useful information about the jetson and the environment we are using.
NVIDIA Jetson Xavier NX (Developer Kit Version) – Jetpack 4.6 [L4t 32.6.1]
cuda 10.2.300
OPENCV 4.5.5 compiled CUDA
Tensorflow 2.3.2+nv21,12
OS Ubuntu 18.04 LTS
Python 3.6.9
Thanks in advance