YOLO Implementation on Nvidia GPU - for TensorRT and DeepStream SDK

Installing Pre-requisites:

If the target platform is a dGPU, download and install DeepStream 2.0. For Tegra platforms, flash your device with Jetpack 3.3 and install Deepstream 1.5.

Install GStreamer pre-requisites using: sudo apt-get install libgstreamer-plugins-base1.0-dev libgstreamer1.0-dev

Install Google flags using: sudo apt-get install libgflags-dev

To use just the stand alone trt-yolo-app, Deepstream Installation can be skipped. However CUDA 9.2 and TensorRT 4.x should be installed.

Setup

Update all the parameters in Makefile.config file present in the root directory

Building NvYolo Plugin

1. Go to the data directory and add your yolo .cfg and .weights file.

   For yolo v2, Download the config file from the darknet repo located at https://github.com/pjreddie/darknet/blob/master/cfg/yolov2.cfg Download the weights file by running the command wget https://pjreddie.com/media/files/yolov2.weights

   For yolo v2 tiny, Download the config file from the darknet repo located at https://github.com/pjreddie/darknet/blob/master/cfg/yolov2-tiny.cfg Download the weights file by running the command wget https://pjreddie.com/media/files/yolov2-tiny.weights

   For yolo v3,
   Download the config file from the darknet repo located at https://github.com/pjreddie/darknet/blob/master/cfg/yolov3.cfg Download the weights file by running the command wget https://pjreddie.com/media/files/yolov3.weights

   For yolo v3 tiny, Download the config file from the darknet repo located at https://github.com/pjreddie/darknet/blob/master/cfg/yolov3-tiny.cfg Download the weights file by running the command wget https://pjreddie.com/media/files/yolov3-tiny.weights

2. Set the right macro in the network_config.h file to choose a model architecture

3. [OPTIONAL] Update the paths of the .cfg and .weights file and other network params in network_config.cpp file if required.

4. Add absolute paths of images to be used for calibration in the calibration_images.txt file within the data directory.

5. Run the following command from sources/plugins/gst-yoloplugin-tesla for dGPU's or from sources/plugins/gst-yoloplugin-tegra for tegra devices to build and install the plugin make && sudo make install

Building and running the trt-yolo-app

Go to the sources/apps/trt-yolo directory and run the following command to build and install the trt-yolo-app make && sudo make install

The TRT Yolo App located at sources/apps/trt-yolo is a sample standalone app, which can be used to run inference on test images. This app does not have any deepstream dependencies and can be built independently. Add a list of absolute paths of images to be used for inference in the test_images.txt file located at data and run trt-yolo-app from the root directory of this repo. Additionally, the detections on test images can be saved by setting kSAVE_DETECTIONS config param to true in network_config.cpp file. The images overlayed with detections will be saved in the data/detections/ directory.

This app has three command line arguments(optional) that you can pass. One is the batch_size to be used for the TRT engine which is set to 1 by default. The second one is a boolean argument representing if the detections have to be decoded or not which is set to true by default. The last one is an argument to set the seed of random number generators used in the application. It is set to time(0) by default. To run the app with the default options run the following command from the root directory of this repo trt-yolo-app

To change the batch_size of the TRT engine use the following command trt-yolo-app --batch_size=4

Building and running deepStream-yolo-app

Go to the sources/apps/deepstream-yolo directory and run the following command to build and install the deepstream-yolo-app make && sudo make install

The deepStream-yolo-app located at sources/apps/deepstream_yolo is a sample app similar to the Test-1 & Test-2 apps available in the DeepStream SDK. Using the yolo app we build a sample gstreamer pipeline using various components like H264 parser, Decoder, Video Converter, OSD and Yolo plugin to run inference on an elementary h264 video stream.

Once you have built the deepstream-yolo-app as described above, go to the root directory of this repo and run the command deepstream-yolo-app /path/to/sample_video.h264

Running the deepStream-app [TESLA platform only]

Following steps describe how to run the YOLO plugin in the deepstream-app

1. The section below in the config file corresponds to ds-example(yolo) plugin in deepstream. The config file is located at config/deepstream-app_yolo_config.txt. Make any changes to this section if required.

   [ds-example]
   enable=1
   processing-width=1280
   processing-height=720
   full-frame=1
   unique-id=15
   gpu-id=0
   

2. Update path to the video file source in the URI field under [source0] group of the config file uri=file://relative/path/to/source/video

3. Go to the root folder of this repo and run deepstream-app -c config/deepstream-app_yolo_config.txt

Running the nvgstiva-app [TEGRA platform only]

1. The section below in the config file corresponds to ds-example(yolo) plugin in deepstream. The config file is located at config/nvgstiva-app_yolo_config.txt. Make any changes to this section if required.

   [ds-example]
   enable=1
   processing-width=640
   processing-height=480
   full-frame=1
   unique-id=15
   

2. Update path to the video file source in the URI field under [source0] group of the config file uri=file://path/to/source/video

3. Go to the root folder of this repo and run nvgstiva-app -c config/nvgstiva-app_yolo_config.txt

Note

1. If you are using the plugin with deepstream-app (located at /usr/bin/deepstream-app), register the yolo plugin as dsexample. To do so, replace line 671 in gstyoloplugin.cpp with return gst_element_register(plugin, "dsexample", GST_RANK_PRIMARY, GST_TYPE_YOLOPLUGIN);

This registers the plugin with the name dsexample so that the deepstream-app can pick it up and add to it's pipeline. Now go to sources/gst-yoloplugin/ and run make && sudo make install to build and install the plugin.

2. Tegra users working with Deepstream 1.5 and Jetpack 3.3 will have to regenerate the .cache files to use the standard caffe models available in the SDK. This can be done by deleting all the .cache files in /home/nvidia/Model directory and all its subdirectories and then running the nvgstiva-app using the default config file.