Increase Jetson Inference speed

usuario1918 · March 26, 2023, 1:16am

I have a Jetson nano 4GB, I use it with 5W.
This code has an SPF of 17

dusty-nv/jetson-inference/blob/master/python/examples/detectnet.py

#!/usr/bin/env python3
#
# Copyright (c) 2020, NVIDIA CORPORATION. All rights reserved.
#
# Permission is hereby granted, free of charge, to any person obtaining a
# copy of this software and associated documentation files (the "Software"),
# to deal in the Software without restriction, including without limitation
# the rights to use, copy, modify, merge, publish, distribute, sublicense,
# and/or sell copies of the Software, and to permit persons to whom the
# Software is furnished to do so, subject to the following conditions:
#
# The above copyright notice and this permission notice shall be included in
# all copies or substantial portions of the Software.
#
# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
# THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
# FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER

This file has been truncated. show original

With the model SSD-Mobilenet-v2

My question is: how do I make it faster? Maybe 22 FPS. Like the one in this video: Jetson AI Fundamentals - S3E4 - Object Detection Inference - YouTube

It occurs to me that you can do two things:

Decrease the size of the image that is read via the camera (/dev/video0)
That it only read/recognize/detect people that is what I want.

Do you know how to do point 1 and 2?

AastaLLL · March 27, 2023, 2:59am

Hi,

Have you tried to maximize the device’s performance first?

$ sudo jetson_clocks

Since Nano is relatively limited in computing resources, a common alternative is to do the inference periodically.
This can be done via Deepstream SDK:
https://docs.nvidia.com/metropolis/deepstream/dev-guide/text/DS_plugin_gst-nvinfer.html?highlight=interval

Thanks.

usuario1918 · March 27, 2023, 3:46am

I’ve seen that command several times but I’ve never run it and I don’t know what it’s for. I thought it’s some clock/time thing, isn’t it? What exactly does it do?

dusty_nv · March 27, 2023, 4:13pm

Hi @usuario1918, the jetson_clocks script will lock their clocks to their maximum values as defined by the current power mode (in your case, 5W mode). Typically there is the Dynamic Voltage & Frequency Scaling (DVFS) governor which adapts the clocks based on workload (jetson_clocks disables this). Once detectnet.py is up & running though, the clocks should already be pretty stable.

In that YouTube video, Jetson Nano was running in 10W mode, which gives more performance. If you require higher FPS, are you able to change from 5W to 10W mode?

usuario1918 · March 27, 2023, 4:17pm

I can use 10W (with the barrel connector and the joint of the two pins), but when I started it (the jetson inference) it closed. Maybe because I didn’t run sudo jetson_clocks?

dusty_nv · March 27, 2023, 4:43pm

Did the window close, or did the system turnoff? My guess is that your power supply wasn’t capable of supplying the load needed for 10W mode. Running jetson_clocks may make that more immediately obvious, since it will draw more power all the time (as opposed to only under considerable workloads) If it continues to be an issue, you may want to look into upgrading your power supply.

usuario1918 · March 31, 2023, 11:42pm

How many times should I run that command ( sudo jetson_clocks)? Every time I turn on the jetson nano or just once?

I am reading that it is used to use the maximum power of Jetson nano.

dusty_nv · April 4, 2023, 1:08pm

Hi @usuario1918, you should run it after each reboot if you want it to remain active. Although changing the nvpmodel to 10W mode will be the main thing that improves your performance with jetson-inference. Running detectnet/detectnet.py on a continuous video stream or camera on Jetson Nano will result in the clocks being maximized by the dynamic load governor anyways.

system · April 26, 2023, 5:21am

This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.