AWS Deep Learning Containers (DLCs) provide pre-built, optimized Docker images for popular deep learning frameworks like TensorFlow, PyTorch, and MXNet. Designed for users deploying machine learning workloads on AWS, these containers simplify the setup and execution of training and inference tasks across services such as Amazon SageMaker, EC2, ECS, and EKS, offering a streamlined path to production.

How It Works

The project offers a set of Docker images pre-configured with deep learning frameworks, NVIDIA CUDA for GPU acceleration, and Intel MKL for CPU optimization. These images are hosted on Amazon ECR, ensuring readily available, tested, and optimized environments. The DLCs serve as the default execution environments for Amazon SageMaker jobs, abstracting away complex dependency management and providing a consistent, high-performance runtime.

Quick Start & Requirements

• Primary Install/Run: Building custom images involves cloning the repository, setting AWS environment variables, logging into ECR, installing requirements (pip install -r src/requirements.txt), and running build scripts (python src/main.py ...).
• Prerequisites: An AWS account with appropriate IAM permissions (e.g., AmazonEC2ContainerRegistryFullAccess, AmazonSageMakerFullAccess), Docker client, and Python 3.
• Setup Time: Initial image builds can be time-consuming due to downloading base layers; subsequent builds are faster.
• Documentation: Links to available images and SageMaker integration details are mentioned.

Highlighted Details

• Optimized environments featuring TensorFlow, PyTorch, MXNet, CUDA, and Intel MKL.
• Seamless integration with Amazon SageMaker for training, inference, and batch transforms.
• Tested compatibility across Amazon EC2, ECS, and EKS.
• Extensive local testing framework using pytest for various AWS deployment scenarios (EC2, ECS, EKS, SageMaker local/remote).
• Supports customization of Dockerfiles, adding artifacts to the build context, and incorporating custom packages.

Licensing & Compatibility

The core project is licensed under the Apache-2.0 License. However, specific components like smdistributed.dataparallel and smdistributed.modelparallel are released under the AWS Customer Agreement. The Apache-2.0 license is generally permissive for commercial use.

Limitations & Caveats

Amazon SageMaker does not support tensorflow_inference py2 images. Building images for the first time requires downloading base layers, which can be time-intensive. Setting up the environment necessitates significant AWS IAM permissions.