Building a fully custom container for SageMaker Processing

We'll reuse the news headlines example from Chapter 6, Training Natural Processing Models:

1. We start with a Dockerfile based on a minimal Python image. We install dependencies, add our processing script, and define it as our entry point:

   FROM python:3.7-slim
   RUN pip3 install --no-cache gensim nltk sagemaker
   RUN python3 -m nltk.downloader stopwords wordnet
   ADD preprocessing-lda-ntm.py /
   ENTRYPOINT ["python3", "/preprocessing-lda-ntm.py"]

2. We build the image and tag it as sm-processing-custom:latest:

   $ docker build -t sm-processing-custom:latest -f Dockerfile .

   The resulting image is 497 MB. For comparison, it's 1.2 GB if we start from python:3.7 instead of python:3.7-slim. This makes it faster to push and download.

3. Using the AWS CLI, we create a repository in Amazon ECR to host this image, and we log in to the repository:

   $ aws ecr create-repository --repository-name sm-processing-custom...