Deploy on NVIDIA Jetson using TensorRT and DeepStream SDK
This guide explains how to deploy a trained model into NVIDIA Jetson Platform and perform inference using TensorRT and DeepStream SDK. Here we use TensorRT to maximize the inference performance on the Jetson platform.
We have tested and verified this guide on the following Jetson devices
Before You Start
Make sure you have properly installed JetPack SDK with all the SDK Components and DeepStream SDK on the Jetson device as this includes CUDA, TensorRT and DeepStream SDK which are needed for this guide.
JetPack SDK provides a full development environment for hardware-accelerated AI-at-the-edge development. All Jetson modules and developer kits are supported by JetPack SDK.
There are two major installation methods including,
SD Card Image Method
NVIDIA SDK Manager Method
Install Necessary Packages
- Step 1. Access the terminal of Jetson device, install pip and upgrade it
- Step 2. Clone the following repo
- Step 3. Open requirements.txt
- Step 5. Edit the following lines. Here you need to press i first to enter editing mode. Press ESC, then type :wq to save and quit
Note: torch and torchvision are excluded for now because they will be installed later.
- Step 6. install the below dependency
- Step 7. Install the necessary packages
Install PyTorch and Torchvision
We cannot install PyTorch and Torchvision from pip because they are not compatible to run on Jetson platform which is based on ARM aarch64 architecture. Therefore we need to manually install pre-built PyTorch pip wheel and compile/ install Torchvision from source.
Visit this page to access all the PyTorch and Torchvision links.
Here are some of the versions supported by JetPack 4.6 and above.
Supported by JetPack 4.4 (L4T R32.4.3) / JetPack 4.4.1 (L4T R32.4.4) / JetPack 4.5 (L4T R32.5.0) / JetPack 4.5.1 (L4T R32.5.1) / JetPack 4.6 (L4T R32.6.1) with Python 3.6
Supported by JetPack 5.0 (L4T R34.1.0) / JetPack 5.0.1 (L4T R34.1.1) / JetPack 5.0.2 (L4T R35.1.0) with Python 3.8
- Step 1. Install torch according to your JetPack version in the following format
For example, here we are running JP4.6.1 and therefore we choose PyTorch v1.10.0
- Step 2. Install torchvision depending on the version of PyTorch that you have installed. For example, we chose PyTorch v1.10.0, which means, we need to choose Torchvision v0.11.1
Here a list of the corresponding torchvision version that you need to install according to the PyTorch version:
PyTorch v1.10 - torchvision v0.11.1
PyTorch v1.12 - torchvision v0.13.0
DeepStream Configuration for YOLOv5
- Step 1. Clone the following repo
- Step 2. Copy gen_wts_yoloV5.py from DeepStream-Yolo/utils into yolov5 directory
- Step 3. Inside the yolov5 repo, download pt file from YOLOv5 releases (example for YOLOv5s 6.1)
- Step 4. Generate the cfg and wts files
Note: To change the inference size (defaut: 640)
- Step 5. Copy the generated cfg and wts files into the DeepStream-Yolo folder
- Step 6. Open the DeepStream-Yolo folder and compile the library
- Step 7. Edit the config_infer_primary_yoloV5.txt file according to your model
- Step 8. Edit the deepstream_app_config file
- Step 9. Change the video source in deepstream_app_config file. Here a default video file is loaded as you can see below
Run the Inference
The above result is running on Jetson Xavier NX with FP32 and YOLOv5s 640x640. We can see that the FPS is around 30.
If you want to use INT8 precision for inference, you need to follow the steps below
- Step 1. Install OpenCV
- Step 2. Compile/recompile the nvdsinfer_custom_impl_Yolo library with OpenCV support
Step 3. For COCO dataset, download the val2017, extract, and move to DeepStream-Yolo folder
Step 4. Make a new directory for calibration images
- Step 5. Run the following to select 1000 random images from COCO dataset to run calibration
Note: NVIDIA recommends at least 500 images to get a good accuracy. On this example, 1000 images are chosen to get better accuracy (more images = more accuracy). Higher INT8_CALIB_BATCH_SIZE values will result in more accuracy and faster calibration speed. Set it according to you GPU memory. You can set it from head -1000. For example, for 2000 images, head -2000. This process can take a long time.
- Step 6. Create the calibration.txt file with all selected images
- Step 7. Set environment variables
- Step 8. Update the config_infer_primary_yoloV5.txt file
- Step 9. Run the inference
The above result is running on Jetson Xavier NX with INT8 and YOLOv5s 640x640. We can see that the FPS is around 60.
The following table summarizes how different models perform on Jetson Xavier NX.
|Model Name||Precision||Inference Size||Inference Time (ms)||FPS|
This tutorial is written by our friends at seeed @lakshanthad and Elaine