I want to train a custom ML model with SageMaker. The model is written in Python and should be shipped to SageMaker in a Docker image. Here is a simplified version of my Dockerfile (the model sits in the train.py file):
FROM amazonlinux:latest
# Install Python 3
RUN yum -y update && yum install -y python3-pip python3-devel gcc && yum clean all
# Install sagemaker-containers (the official SageMaker utils package)
RUN pip3 install --target=/usr/local/lib/python3.7/site-packages sagemaker-containers && rm -rf /root/.cache
# Bring the script with the model to the image
COPY train.py /opt/ml/code/train.py
ENV SAGEMAKER_PROGRAM train.py
Now, if I initialize this image as a SageMaker estimator and then run the fit
method on this estimator I get the following error:
"AlgorithmError: CannotStartContainerError. Please make sure the container can be run with 'docker run train'."
In other words: SageMaker is not able to get into the container and run the train.py file. But why? The way I am specifying the entrypoint with ENV SAGEMAKER_PROGRAM train.py
is recommended in the docs of the sagemaker-containers package (see 'How a script is executed inside the container').
I found a hint in the AWS docs and came up with this solution:
ENTRYPOINT ["python3.7", "/opt/ml/code/train.py"]
With this the container will run as an executable.