This is a straightforward follow-up of my previos JetPack-4.4 for Jetson Nano and Setting up Jetson Xavier NX posts. As NVIDIA formally released JetPack-4.5 a couple of weeks ago, I have tested it and documented the steps of installing various software packages and testing my tensorrt_demos examples.

So far I have only tested JetPack-4.5 on my Jetson Nano DevKit. But I believe the following steps work equally well on Jetson Xavier NX and else.

1. Basic set-up (microSD card)

I recommend using a microSD card of at least 128GB in size.

I downloaded the JetPack-4.5 image, “Jetson Nano Developer Kit SD Card Image”, from the official Getting Started With Jetson Nano Developer Kit page and “etched” the image onto my microSD card. I then inserted the microSD card and boot my Jetson Nano DevKit. During initial Ubuntu set-up, I selected “MAXN” mode for the Nano. And I created an account named “nvidia” (you could replace it with your preferred user name).

Tip #1: When the Jetson Nano boot up to the Desktop, I would hit Ctrl-Alt-T to bring up a terminal and then lock the terminal on the Launcher (on the left-hand side).

Tip #2: In the command terminal, I would type and execute “chromium-browser” to bring up the web browser. Again, I would lock it onto Launcher for easier use later on.

Then I’d run the following commands to set up CUDA related environment variables properly.

$ sudo jetson_clocks
$ sudo apt update
### Set proper environment variables
$ mkdir ${HOME}/project
$ cd ${HOME}/project
$ git clone
$ cd jetson_nano
$ ./
$ source ${HOME}/.bashrc

2. Making sure python3 “cv2” is working

I’d just use the built-in OpenCV-4.1.1 provided by NVIDIA (no need to build from source by myself). I’d do the following for system library dependencies and python module dependencies. NOTE: For “protobuf” libraries, instead of doing apt install, I recommend installing the newer version (3.8.0) using my “” script as shown below. The “protobuf” libraries could have a noticeable effect on the performance (inference speed) of tensorflow or else. The building/installation process of “protobuf” could take a couple of hours, though. So be patient…

In addition, I use an older version of pip3, “20.2.1”, since the latest version does not work (SSLError: “WRONG_VERSION_NUMBER”, which seems related to accessing https pip index url’s behind a firewall/proxy) in our company lab network environment.

### Install dependencies for python3 "cv2"
$ sudo apt update
$ sudo apt install -y build-essential make cmake cmake-curses-gui \
                      git g++ pkg-config curl libfreetype6-dev \
                      libcanberra-gtk-module libcanberra-gtk3-module \
                      python3-dev python3-pip
$ sudo pip3 install -U pip==20.2.1 Cython testresources setuptools
$ cd ${HOME}/project/jetson_nano
$ ./
$ sudo pip3 install numpy==1.19.4 matplotlib==3.2.2

Then I’d test my script with a USB webcam, and make sure the python3 “cv2” module could capture and display images properly.

### Test (using a USB webcam)
$ cd ${HOME}/project
$ wget
$ python3 --usb --vid 0

3. Installing tensorflow-2.4.0

Reference (official documentation from NVIDIA): Installing TensorFlow For Jetson Platform

To install tensorflow, I just followed instructions on the official documentation, but skipped installation of “protobuf”. (I already built and installed “protobuf-3.8.0” in the opencv section.) These steps would take roughly 20~30 minutes to finish.

$ sudo apt install -y libhdf5-serial-dev hdf5-tools libhdf5-dev zlib1g-dev \
                      zip libjpeg8-dev liblapack-dev libblas-dev gfortran
$ sudo pip3 install -U future==0.18.2 mock==3.0.5 h5py==2.10.0 \
                       keras_preprocessing==1.1.1 keras_applications==1.0.8 \
                       gast==0.2.2 futures pybind11
$ sudo pip3 install --pre --extra-index-url \

I tested and made sure “import tensorflow as tf” worked OK in python3.

4. Testing TensorRT GoogLeNet and MTCNN

Please refer to my JetPack-4.4 for Jetson Nano post if you are interested in testing these 2 models.

5. Skipping TensorRT SSD models

My Demo #3: SSD example only works against tensorflow-1.x. So I skipped this testing.

6. Testing TensorRT YOLOv3 and YOLOv4 models

Reference: Demo #4: YOLOv3 and Demo #5: YOLOv4

I installed dependencies and build the TensorRT yolov3/yolov4 engines.

### Clone the tensorrt_demos code
$ cd ${HOME}/project
$ git clone
### Install dependencies and build TensorRT engine
$ cd ${HOME}/project/tensorrt_demos/ssd
$ ./
$ sudo pip3 install onnx==1.9.0
$ cd ${HOME}/project/tensorrt_demos/plugins
$ make
$ cd ${HOME}/project/tensorrt_demos/yolo
$ ./
$ python3 -m yolov3-416
$ python3 -v -m yolov3-416
$ python3 -m yolov4-416
$ python3 -v -m yolov4-416

Next, I tested the TensorRT engines with the USB webcam.

$ cd ${HOME}/project/tensorrt_demos
$ python3 --usb 0 -m yolov3-416
$ python3 --usb 0 -m yolov4-416


JetPack-4.5 contains exactly the same versions of CUDA toolkit, cuDNN and TensorRT libraries as JetPack-4.4. I didn’t feel any performance difference between JetPack-4.5 and the previous version when testing it on Jetson Nano DevKit.

According to NVIDIA’s announcement, one particular new addition in JetPack-4.5 is availability of JetPack 4.5 container on NGC. I probably should find time to test that.

blog built using the cayman-theme by Jason Long. LICENSE