Invoke different entrypoints/modules when training with custom container - google-cloud-ml

Question

I've built a custom Docker container with my training application. The Dockerfile, at the moment, is something like
FROM python:slim
COPY ./src /pipelines/component/src
RUN pip3 install -U ...
...
ENTRYPOINT ["python3", "/pipelines/component/src/training.py"]
so when I run
gcloud ai-platform jobs submit training JOB_NAME \
--region=$REGION \
--master-image-uri=$IMAGE_URI
it goes as expected.
What I'd like to do is to add another module, like /pipelines/component/src/tuning.py; remove the default ENTRYPOINT from Dockerfile; decide which module to call from the gcloud command. So I tried
gcloud ai-platform jobs submit training JOB_NAME \
--region=$REGION \
--master-image-uri=$IMAGE_URI \
--module-name=src.tuning \
--package-path=/pipelines/component/src
It returns Source directory [/pipelines/component] is not a valid directory., because it's searching for the package path on the local machine, instead of the container. How can I solve this problem?

Answer 1

You can use TrainingInput.ReplicaConfig.ContainerCommand field to override the docker image's entrypoint. Here is a sample command:
gcloud ai-platform jobs submit training JOB_NAME \
--region=$REGION
--master-image-uri=$IMAGE_URI
--config=config.yaml
And config.yaml content will be something like this:
trainingInput:
scaleTier: BASIC
masterConfig:
containerCommand: ["python3", "/pipelines/component/src/tuning.py"]
This link has more context about config flag.
Similarly, you can override docker image's command with containerArgs field.