--- comments: true description: Detailed guide on deploying trained models on NVIDIA Jetson using TensorRT and DeepStream SDK. Optimize the inference performance on Jetson with Ultralytics. keywords: TensorRT, NVIDIA Jetson, DeepStream SDK, deployment, Ultralytics, YOLO, Machine Learning, AI, Deep Learning, model optimization, inference performance --- # 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. UPDATED 18 November 2022. ## Hardware Verification We have tested and verified this guide on the following Jetson devices - [Seeed reComputer J1010 built with Jetson Nano module](https://www.seeedstudio.com/Jetson-10-1-A0-p-5336.html) - [Seeed reComputer J2021 built with Jetson Xavier NX module](https://www.seeedstudio.com/reComputer-J2021-p-5438.html) ## 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, 1. SD Card Image Method 2. NVIDIA SDK Manager Method You can find a very detailed installation guide from NVIDIA [official website](https://developer.nvidia.com/jetpack-sdk-461). You can also find guides corresponding to the above-mentioned [reComputer J1010](https://wiki.seeedstudio.com/reComputer_J1010_J101_Flash_Jetpack) and [reComputer J2021](https://wiki.seeedstudio.com/reComputer_J2021_J202_Flash_Jetpack). ## Install Necessary Packages - **Step 1.** Access the terminal of Jetson device, install pip and upgrade it ```sh sudo apt update sudo apt install -y python3-pip pip3 install --upgrade pip ``` - **Step 2.** Clone the following repo ```sh git clone https://github.com/ultralytics/yolov5 ``` - **Step 3.** Open **requirements.txt** ```sh cd yolov5 vi 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 ```sh # torch>=1.8.0 # torchvision>=0.9.0 ``` **Note:** torch and torchvision are excluded for now because they will be installed later. - **Step 6.** install the below dependency ```sh sudo apt install -y libfreetype6-dev ``` - **Step 7.** Install the necessary packages ```sh pip3 install -r requirements.txt ``` ## 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](https://forums.developer.nvidia.com/t/pytorch-for-jetson) to access all the PyTorch and Torchvision links. Here are some of the versions supported by JetPack 4.6 and above. **PyTorch v1.10.0** 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 **file_name:** torch-1.10.0-cp36-cp36m-linux_aarch64.whl **URL:** [https://nvidia.box.com/shared/static/fjtbno0vpo676a25cgvuqc1wty0fkkg6.whl](https://nvidia.box.com/shared/static/fjtbno0vpo676a25cgvuqc1wty0fkkg6.whl) **PyTorch v1.12.0** 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 **file_name:** torch-1.12.0a0+2c916ef.nv22.3-cp38-cp38-linux_aarch64.whl **URL:** [https://developer.download.nvidia.com/compute/redist/jp/v50/pytorch/torch-1.12.0a0+2c916ef.nv22.3-cp38-cp38-linux_aarch64.whl](https://developer.download.nvidia.com/compute/redist/jp/v50/pytorch/torch-1.12.0a0+2c916ef.nv22.3-cp38-cp38-linux_aarch64.whl) - **Step 1.** Install torch according to your JetPack version in the following format ```sh wget -O pip3 install ``` For example, here we are running **JP4.6.1**, and therefore we choose **PyTorch v1.10.0** ```sh cd ~ sudo apt-get install -y libopenblas-base libopenmpi-dev wget https://nvidia.box.com/shared/static/fjtbno0vpo676a25cgvuqc1wty0fkkg6.whl -O torch-1.10.0-cp36-cp36m-linux_aarch64.whl pip3 install torch-1.10.0-cp36-cp36m-linux_aarch64.whl ``` - **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** ```sh sudo apt install -y libjpeg-dev zlib1g-dev git clone --branch v0.11.1 https://github.com/pytorch/vision torchvision cd torchvision sudo python3 setup.py install ``` 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 ```sh cd ~ git clone https://github.com/marcoslucianops/DeepStream-Yolo ``` - **Step 2.** Copy **gen_wts_yoloV5.py** from **DeepStream-Yolo/utils** into **yolov5** directory ```sh cp DeepStream-Yolo/utils/gen_wts_yoloV5.py yolov5 ``` - **Step 3.** Inside the yolov5 repo, download **pt file** from YOLOv5 releases (example for YOLOv5s 6.1) ```sh cd yolov5 wget https://github.com/ultralytics/yolov5/releases/download/v6.1/yolov5s.pt ``` - **Step 4.** Generate the **cfg** and **wts** files ```sh python3 gen_wts_yoloV5.py -w yolov5s.pt ``` **Note**: To change the inference size (default: 640) ```sh -s SIZE --size SIZE -s HEIGHT WIDTH --size HEIGHT WIDTH Example for 1280: -s 1280 or -s 1280 1280 ``` - **Step 5.** Copy the generated **cfg** and **wts** files into the **DeepStream-Yolo** folder ```sh cp yolov5s.cfg ~/DeepStream-Yolo cp yolov5s.wts ~/DeepStream-Yolo ``` - **Step 6.** Open the **DeepStream-Yolo** folder and compile the library ```sh cd ~/DeepStream-Yolo CUDA_VER=11.4 make -C nvdsinfer_custom_impl_Yolo # for DeepStream 6.1 CUDA_VER=10.2 make -C nvdsinfer_custom_impl_Yolo # for DeepStream 6.0.1 / 6.0 ``` - **Step 7.** Edit the **config_infer_primary_yoloV5.txt** file according to your model ```sh [property] ... custom-network-config=yolov5s.cfg model-file=yolov5s.wts ... ``` - **Step 8.** Edit the **deepstream_app_config** file ```sh ... [primary-gie] ... config-file=config_infer_primary_yoloV5.txt ``` - **Step 9.** Change the video source in **deepstream_app_config** file. Here a default video file is loaded as you can see below ```sh ... [source0] ... uri=file:///opt/nvidia/deepstream/deepstream/samples/streams/sample_1080p_h264.mp4 ``` ## Run the Inference ```sh deepstream-app -c deepstream_app_config.txt ```
The above result is running on **Jetson Xavier NX** with **FP32** and **YOLOv5s 640x640**. We can see that the **FPS** is around **30**. ## INT8 Calibration If you want to use INT8 precision for inference, you need to follow the steps below - **Step 1.** Install OpenCV ```sh sudo apt-get install libopencv-dev ``` - **Step 2.** Compile/recompile the **nvdsinfer_custom_impl_Yolo** library with OpenCV support ```sh cd ~/DeepStream-Yolo CUDA_VER=11.4 OPENCV=1 make -C nvdsinfer_custom_impl_Yolo # for DeepStream 6.1 CUDA_VER=10.2 OPENCV=1 make -C nvdsinfer_custom_impl_Yolo # for DeepStream 6.0.1 / 6.0 ``` - **Step 3.** For COCO dataset, download the [val2017](https://drive.google.com/file/d/1gbvfn7mcsGDRZ_luJwtITL-ru2kK99aK/view?usp=sharing), extract, and move to **DeepStream-Yolo** folder - **Step 4.** Make a new directory for calibration images ```sh mkdir calibration ``` - **Step 5.** Run the following to select 1000 random images from COCO dataset to run calibration ```sh for jpg in $(ls -1 val2017/*.jpg | sort -R | head -1000); do \ cp ${jpg} calibration/; \ done ``` **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 ```sh realpath calibration/*jpg > calibration.txt ``` - **Step 7.** Set environment variables ```sh export INT8_CALIB_IMG_PATH=calibration.txt export INT8_CALIB_BATCH_SIZE=1 ``` - **Step 8.** Update the **config_infer_primary_yoloV5.txt** file From ```sh ... model-engine-file=model_b1_gpu0_fp32.engine #int8-calib-file=calib.table ... network-mode=0 ... ``` To ```sh ... model-engine-file=model_b1_gpu0_int8.engine int8-calib-file=calib.table ... network-mode=1 ... ``` - **Step 9.** Run the inference ```sh deepstream-app -c deepstream_app_config.txt ```
The above result is running on **Jetson Xavier NX** with **INT8** and **YOLOv5s 640x640**. We can see that the **FPS** is around **60**. ## Benchmark results The following table summarizes how different models perform on **Jetson Xavier NX**. | Model Name | Precision | Inference Size | Inference Time (ms) | FPS | |------------|-----------|----------------|---------------------|-----| | YOLOv5s | FP32 | 320x320 | 16.66 | 60 | | | FP32 | 640x640 | 33.33 | 30 | | | INT8 | 640x640 | 16.66 | 60 | | YOLOv5n | FP32 | 640x640 | 16.66 | 60 | ### Additional This tutorial is written by our friends at seeed @lakshanthad and Elaine