Jenson Nano AI / ML Experiments

This repository is mainly for documenting my personal experiments on Jetson platform.

Current setup is Jetson Nano Developement Kit (4GB) version B01 with Intel Dual Band Wireless-AC 3168 card and two IMX219 sensor cameras in neat metal case from Waveshare.

None of existing start points suited my purposed as I want to run experiments in headless environment ie. from Jupyter notebooks. JetPack version when writing these is 4.4.1.

I recommend at least 128Gb sdcard.

Initial setup

Initial setup is done as described in NVIDIAs guide. Getting WIFI working needs following additional steps to be done from console. Everything thereafter can be done over SSH connection. AC 3168 card does not understand power management and don't turn on if it is used.

sudo sed -i -e '/APPEND/s/$/ pcie_aspm=off/' /boot/extlinux/extlinux.conf
sudo reboot

After reboot connect to your WIFI network.

sudo nmcli device wifi connect <SSID> password <PASSWORD>

Bootstrapping environment

Setup tasks needed for nice working environment are collected from history into bootstrap.sh - just running it should work.

./bootstrap.sh

These days I don't want to maintain 'dotfile' configurations among lot of enviroments. Thus I always install OhMyZsh and SpaceVim on all new systems.

curl -sLf https://raw.github.com/ohmyzsh/ohmyzsh/master/tools/install.sh | bash
curl -sLf https://spacevim.org/install.sh | bash

install_jupyterlab.sh installs and enables JupyterLab, needed dependencies and many nice and usable plugins.

./install_jupyterlab.sh

Now base system is ready. At this point it is good to make backup snapshot of the sdcard on computer sdcard was initialized:

sudo /bin/dd if=/dev/sdb bs=1M | zip jetson_backup.zip -

Freeing up resources

As I am using Jetson only in headless mode with Jupyter Lab I don't need native GUI. To remove it do:

sudo systemctl stop lightdm
sudo systemctl disable lightdm
sudo systemctl set-default multi-user.target
sudo apt remove --purge ubuntu-desktop lightdm gdm3

Also Docker can be disabled

sudo systemctl stop containerd
sudo systemctl disable containderd

This frees about 250M memory to more important stuff like ML models.

Hello World experiment

Install libjetson-inference library from the "Hello AI World" This comes with handy pretained models.

git clone --recursive https://github.com/dusty-nv/jetson-inference
cd jetson-inference
mkdir build
cd build
cmake ../
make -j$(nproc)
sudo make install
sudo ldconfig

Then start jupyterlab on terminal and browse to notebooks

jupyter-lab --no-browser

TensorFlow models and NVIDIA TensorRT

TensorFlow maintains nice gallery of model implementations for TensorFlow.

NVIDIA provides higly optimized TensorRT runtime and SDK with the Jetson Nano SDK.

TensorFLow models can be converted directly with included trt_convert

They also provide handy converter to convert PyTorch models to TensorRT. Installation from GitHub:

git clone https://github.com/NVIDIA-AI-IOT/torch2trt
cd torch2trt
sudo python3 setup.py install --plugins

Many pretrained models can be found from PyTorchGub