LEIP Recipes are delivered as part of the LEIP SDK and are made available inside a Docker image. You will need to have Docker installed to continue. A container is provided with all the required dependencies for the SDK while running the image. Refer to the Docker Overview for additional information on Docker images and containers.
We highly recommend that you increase the amount of memory you have allocated to Docker to match the requirements below.
You will need to meet the following specifications for your development environment to follow this recipe:
Your host computer must:
Have support for Docker.
Have a Linux operating system. For Windows support, please contact Latent AI.
Have an NVIDIA graphics processing unit (GPU). This recipe supports Maxwell through Ampere class GPUs.
Have at least 16GB of RAM to run the experiment “as-is.” Refer to the troubleshooting section on how to modify the experiment if the system has less RAM.
Your detection dataset needs to be in a supported format to use the plug-in BYOD feature. LEIP Recipes currently support data in the MS COCO and Pascal VOC formats for the Detection models and the PyTorch ImageFolder format for the Classifier models.
Pulling the Container
Before you begin, you will need your license key and to have been granted access to the Latent AI container registry in order to complete the SDK installation process.
The LEIP SDK is distributed in a Docker container. Please verify you have Docker Version 1.19 or higher installed on your target host and have allocated at least 4GB of memory for Docker. While Latent AI provides several Docker variants of the SDK depending on your hardware target, the initial AGX recipe is only provided with the CUDA and CUDA 10.2 optimized containers.
Choose the container you would like to use for your development environment. There are multiple options depending on your use case. The latest version will always include the latest tag. However, if you need a specific version, the versions are available by replacing the variant’s latest tag with the version number string (for example, registry.latentai.io/leip-sdk/gpu-cuda:2.7.0).
Please contact firstname.lastname@example.org for assistance if you need access to the container registry and a license key.
Docker Hub Namespace
Optimized CUDA GPU Support
# Login to the Latent AI container registry
# (for credentials, refer to the instructions that came with your license key)
docker login registry.latentai.io
# Pull one of the LEIP SDK Docker images from the registry
docker pull $variant
Starting the Container
Start the Docker container by using your LEIP license key.
# Run the container using a valid LEIP license key
# and setting the `--gpus all` directive as needed
docker run -e LICENSE_KEY=$license-key --name leip -it --gpus all --ipc=host $variant
# TIP: after you exit the container, if you want to return to this container state
docker start leip; docker attach leip
# use docker rm leip to remove it
Now we will review the stages that make up a LEIP Recipe.