Skip to content

Latest commit

 

History

History
167 lines (125 loc) · 8.79 KB

README.org

File metadata and controls

167 lines (125 loc) · 8.79 KB
Raw

Basic Set up

Installing Jetpack

Install the SDK Manager .deb file first on a host machine (Ubuntu 18.04) through Nvidia SDK Manager Install

  1. Make sure your host machine is Ubuntu 18.04, Jetpack ~sdkmanager~` will decide Jetpack version based on the board it detects and the host machine’s version
  2. after the file is installed, run sudo dpkg -i your-sdk-manager-name.deb
  3. now you can run it using sdkmanager

Set up the Jetson board for flashing

connect 5V power through barrel jack, USB cable to host laptop with Ubuntu 18.04, HDMI/Display port to a monitor, some mouse/keyboard to USB, ethernet

The ports should look something like this:

images/ports.jpg

on the back of the Jetson, connect a jumper between GND and FC REC The FC REC and GND jumper on the back of the jetson under the heat sink should look like this:

images/pins.jpg

Flashing the Jetson

type sdkmanager in terminal to run

unselect Deepstream for now (eMMC jetson only has 16 GB storage, cannot install anything other than the OS for now) and ensure the Jetpack version is 4.6.4

The selection screen initially should look like this:

images/install.png click continue

unselect all software other than the OS

The selection screen for software should look like this:

images/software.png click continue, set the username/password, and start the install (takes around 30 minutes)

Now remove the jumper across GND and FC REC otherwise Jetson will not boot since it still thinks it’s in flashing mode

Expanding Jetson Storage using USB Drive

Jetson nano (eMMC version especially) only comes with 16 GB of storage. Cannot do anything as this is without any packages or models installed. Thus, need to expand storage.

Take any USB stick with a large storage capacity (preferable reliable/not cheap so your files don’t die and also metal since they do run hot)

Plug in the USB stick into the Jetson on a free port

Boot up the Jetson and log in

Open gnome disks to format the USB drive

format to GPT and wipe all memory on the drive as well (this will take a while) using gnome disks looks something like this:

images/format.png

After formatting, make sure the partitioning is GUID Paritioning Table (gnome disks shows this by default)

Create a partition

  • Create a partition (click “+” in gnome disks) and allocate 10 GB less than the max size of the drive
    • we will use the extra space for a swapfile later
  • name the volume whatever you want
  • partition type should be Ext4

using gnome disks, the partition creation should look like this:

images/part.png

the final partitions should look like this (extra space for swapfile later):

images/final_part.png

From now all everything is ran on the Jetson unless otherwise specified. Clone this repo on the Jetson so you have access to the files

Open a terminal and type ls /dev/sd* to get what the name of the USB drive is to the Jetson (e.g. mine is /dev/sda1)

cd /boot/extlinux

Change where the Jetson boots from by editing extlinux.conf

make a backup first incase something goes wrong by sudo cp extlinux.conf extlinux.conf.original

edit extlinux.conf by typing vi extlinux.conf and first copy the LABEL block and create another called sdcard or eMMC as a backup boot. (you should now have primary and eMMc/sdcard. It will boot from primary by default, which will be your USB drive)

images/ext.png

chmod +x partUUID.sh and then run ./partUUID.sh in a different terminal and copy and paste the root=xxxx-xxxx-xxxx string and replace the root=/dev/mmcblk0p1 for LABEL primary

images/ext.png Now your default boot is from your USB drive and backup boot is from eMMC

Copy the root and all Jetson files to USB drive to boot from it

Open up the file browser to “mount” the usb drive

Now just copy the root and all your files to the USB drive by running chmod +x copyRootToUSB.sh and then run ./copyRootToUSB.sh -p /dev/sda1 (note: /dev/sda1 was found from earlier by running ls /dev/sd* to find the name of the USB drive folder)

  • this might take a while

After the files are done copying over, restart and it should boot from the USB drive

Environment Set up

Now we will install Jetpack SDKs, install Python 3.8, YOLO, and Pytorch

Install Jetpack SDKs

Remember earlier how we skipped installing Deepstream and Jetpack SDKs because we didn’t have enough storage? Now go back to host computer with Ubuntu 18.04 that you installed sdkmanager on. Start sdkmanager:

  • connect Jetson via USB
  • select Deepstream, click continue, and select all the SDKs (BUT ENSURE YOU UNSELECT THE OS IMAGE, OTHERWISE WILL FLASH AGAIN AND YOU WILL HAVE TO REPEAT EVERYTHING)
  • click install and let it run

Now you will have CUDA 10.2, Deepstream, TensorRT, and related Nvidia software

Install Python 3.8

Jetpack 4.6.4 is based on Ubuntu 18.04, which means it comes only with Python 3.6. This is bad because YOLOv8 requires >= Python 3.8.

Luckily, the apt repo has Python 3.8, install everything with a simple oneliner: sudo apt-get install python3.8 python3.8-dev python3.8-distutils python3.8-venv

Now since you have both Python 3.6 and 3.8, you need to edit your .bashrc to make sure you are running the right Python version. If you don’t do this, you will install everything on Python 3.6 and it will all break…

Edit your .bashrc by typing vi ~/.bashrc add the following to the end of the file:

  • alias python=”python3.8”
  • alias python3=”python3.8”
  • alias pip=”python3.8 -m pip”
  • alias pip3=”python3.8 -m pip”

:wq to write and quit and close the current terminal

Now you have Python 3.8 installed and you will always run Python 3.8 when you use any python commands

Create Python 3.8 venv and keep it always active

Make sure you are in root folder

Create venv using python -m venv “vision”

Now edit your .bashrc again, same steps as before but now add the following line to the end of the file:

  • source ~/vision/bin/activate

~:wq:

Now for every terminal you start, you will always be in the vision venv you created with Python 3.8. Venv is nice because if you mess up, you can still restart without bricking your all Python on your computer

Install YOLOv8

pip install ultralytics==8.0.231

  • this is the current newest version, but for future, this documentation worked with this version of ultralytics and pytorch
  • FYI, as of now, ultralytics works with >= Python 3.8 and >= torch 1.8

if any dependency errors come up from install (e.g. missing packages) just pip install <missing package> and rerun pip install ultralytics==8.0.231

  • this might change in future

Uninstall default Pytorch and Torchvision

Default pip ultralytics install comes with torch 2.1.x and the corresponding torchvision –> this is wayyy too new and is not for aarch64 architecture on the nano

Therefore, uninstall both by running:

  • pip uninstall torch
  • pip uninstall torchvision

Install correct versions of Torch and Torch vision

In this Torch Folder, you can find the aarch64 Python 3.8 compiled torch 1.11.0 and the matching torchvision 0.12.0 wheels specifically for the Jetson nano. Download them and install them via pip:

  • python -m pip install torch-*.whl torchvision-*.whl

To check if everything worked correctly, ensure the torch package recognizes the cuda GPU on the Jetson by:

  • enter a python shell using python
  • import torch
  • type torch.cuda.device_count(), if this returns 1, it means everything went correctly

Verify Installation

Run YOLOv8 inference by:

ONNX GPU Runtime Install

Default ONNX is for Python 3.6 on Jetpack 4.6. To install for Python 3.8, ensure your venv is activated.

Then download the ONNX 1.11 wheel for Python 3.8 here

Install the wheel using pip install onnxruntime_gpu-1.11.0-cp38-cp38-linux_aarch64.whl

Now your ONNX export models should run fine through YOLOv8. Note ONNX alone doesn’t do that much to speed up inference (around 2 or 3 FPS increase on 600x600)

Credits