Docker Image for Sagemaker

The aim of this repository is to provide a custom Docker image for SageMaker Training Job.

Setup

1. Create an ECR repository at first
2. Create a build project on CodeBuild console.
   • Do not forget to check "Privileged" under "Environment" section.
3. Attach the built image to SageMaker Domain. (If you want to use this image on SageMaker Studio.)

Environment variable

Define the following environment variables on CodeBuild console.

• AWS_DEFAULT_REGION: e.g. eu-central-1
• AWS_ACCOUNT_ID: your account ID. You can copy it on the AWS console. (Click your account name.)
• ECR_REPOSITORY_NAME: Repository name in ECR

Policy

The IAM role of the CodeBuild project needs a privilege to push a built Docker image to your ECR repository. Add the following policy to the role after replacing AWS_ACCOUNT_ID with your AWS account ID.

{
    "Version": "2012-10-17",
    "Statement": [
        {
            "Sid": "VisualEditor0",
            "Effect": "Allow",
            "Action": [
                "ecr:CompleteLayerUpload",
                "ecr:UploadLayerPart",
                "ecr:InitiateLayerUpload",
                "ecr:BatchCheckLayerAvailability",
                "ecr:PutImage"
            ],
            "Resource": "arn:aws:ecr:*:AWS_ACCOUNT_ID:repository/*"
        },
        {
            "Sid": "VisualEditor1",
            "Effect": "Allow",
            "Action": "ecr:GetAuthorizationToken",
            "Resource": "*"
        }
    ]
}

TODO

• GPU