This is an example Dockerfile template and docker compose for deploying your agent in CVAT.

The general Dockerfile approach is:

1. Select base image to use with your agent.
2. Install necessary packages with a package manager. For example YOLO needs libxcb1, libgl1, libglib2.0-0
3. Install cvat-cli, and if you need any Python dependencies, you may put them to requirements.txt
4. You should have func.py file that acts as adapter between your agent and CVAT. You can refer to existing adapters for YOLO, SAM2 or transformers for inspiration. (/ai-models/detector/yolo/func.py, /ai-models/tracker/sam2/func.py, /ai-models/transformers/func.py)
5. Consider supporting some build args like $USE_GPU to allow users to choose between CPU and GPU versions of the image. For the models that we've implemented as agents we rely on pytorch CPU or GPU wheels.
6. Your image should implement the following functionality
   • Register function that will be powered by your agent in CVAT.
   • Run agent that will be polling CVAT for new tasks (Entrypoint).
   • Deregister function in CVAT when it is not needed anymore.

All three capabilities are powered by cvat-cli that authenticates in CVAT with provided API key and allows to register/deregister functions and run agents. Please refer to scripts entrypoint.sh, function_registration.sh and function_deregistration.sh for examples of how to use cvat-cli to implement these capabilities.

General recommendations

• Run your container with a non-root user.
• Use lightweight base images to reduce the size of your image and speed up deployment.
• Keep the frequently changing layers of your Docker closer to the end of your Dockerfile to take advantage of caching.
• Always pin the versions of your dependencies.

Docker compose

Once your image is built, use Docker Compose to manage the agent lifecycle. All services use the same image.

1. It runs function registration script in the container (cvat-function-register service).
2. cvat-function-register service creates a file on shared volume.
3. cvat-agent service runs agents that connect to CVAT function that was created in step 1.
4. When agent is not needed user should run cvat-function-deregister service manually to delete function from CVAT. (docker compose run --rm cvat-function-deregister)

It's easy to configure docker compose using .env file that should be located in the same folder as docker-compose.yaml