What is called within a sagemaker custom (training) container? - tensorflow

Question

Somewhere this spring the behaviour of the sagemaker docker image changed and I cannot find the way I need to construct it now.
Directory structure
/src/some/package
/project1
/some_entrypoint.py
/some_notebook.ipynb
/project2
/another_entrypoint.py
/another_notebook.ipynb
setup.py
Docker file
Note that I want to shift tensorflow version, so I changed the FROM to the latest version. This was the
breaking change.
# Core
FROM 763104351884.dkr.ecr.eu-west-1.amazonaws.com/tensorflow-training:2.3.0-cpu-py37-ubuntu18.04
COPY . /opt/ml/code/all/
RUN pip install /opt/ml/code/all/
WORKDIR "/opt/ml/code"
Python code
This code should start the entrypoints, for example here we have the code of some_notebook.ipynb. I tried all possible combinations of working directory + source_dir (None, '.', or '..'), entry_point (with or without /), dependencies ('src')...
if setup is present it tries to call my project as a module (python -m some_entrypoint)
if not, it often is not able to find my entrypoint. Which I don't understand because the TensorFlow is supposed to add it to the container, isn't it?
estimator = TensorFlow(
entry_point='some_entrypoint.py',
image_name='ECR.dkr.ecr.eu-west-1.amazonaws.com/overall-project/sagemaker-training:latest',
source_dir='.',
# dependencies=['../src/'],
script_mode=True,
train_instance_type='ml.m5.4xlarge',
train_instance_count=1,
train_max_run=60*60, # seconds * minutes
train_max_wait=60*60, # seconds * minutes. Must be >= train_max_run
hyperparameters=hyperparameters,
metric_definitions=metrics,
role=role,
framework_version='2.0.0',
py_version='py3',
)
estimator.fit({
'training': f"s3://some-data/"}
# , wait=False
)
Ideally I would want to understand the logic within: what is called given what settings?

Answer 1

when the training container runs, your entry_point script will be executed.
Since your notebook file and entry_point script are under the same directory, your source_dir should just be "."
Does your entry_point script import any modules that are not installed by the tensorflow training container by default? Also could you share your stacktrace of the error?