![]() ![]() We can use the following command to start a deepstream container, mount our local folder and start making config files. We’ll mount projectto /opt/nvidia/deepstream/deepstream/sources/project in the container. You can already put a video in datathat you want to test with, make sure it is a h264 mp4. I made 2 subfolders too: configs and data Make a folder on your local filesystem to use for this project ( I just called mine project ). However, we do want some sort of persistence of the configs and model files we’ll make, so we can volume mount a folder in which to work. Now that we have docker with GPU support in order, we can just run the NVIDIA deepstream container and do everything from in there, so we don’t have to deal with all the packages and such. We can now use -gpus=all to pass through all your gpus to the container.įor reference: official nvidia-docker installation (still nvidia-docker2). Since this is docker 19.03 though, you should install nvidia-docker-toolkit( link) and restart docker. There’s a lot of ways to do this! You’ll find nvidia-docker and nvidia-docker2 in combination with the docker -runtime=nvidiaflag when searching for ways to do this. ![]() In order to access the GPU from within the docker container we have to install the nvidia-container-toolkit. Make sure you can run the following successfully to make sure docker is installed. I know, I don’t like it either, but there’s truly no way around this one, gpu drivers are… difficult creatures. $ sudo apt install nvidia-driver-460Īnd then… reboot. The gpu driver is backwards compatible with cuda and cudnn versions, so you should almost always choose the most recent one. Jetson users on Jetpack just have to run sudo apt install deepstream-5.1 and you’re good to go! 1. If you have the GPU driver and docker already, you can skip the respective steps. We’re starting here from a virtual machine on google cloud with a GPU, aka a clean ubuntu install. So we recommend using their docker containers instead, since they have the dependencies already installed. It can be a serious hassle to find and install NVIDIA’s deepstream packages and dependencies (if not on a jetson board). There are containers though, if you like that more. NOTE: the install section is for a desktop-gpu system, the jetson boards have no need for the driver as it’s already included in Jetpack and can run the config files just fine without a docker environment. ![]()
0 Comments
Leave a Reply. |