For `Estimator`, `entry_point` is unnecessarily required to be at the root of `source_dir`.

[stevenmanton]

Describe the bug
When using the Estimator with the source_dir and entry_point parameters, the entry_point script is required to be at the root of source_dir. This restriction seems unnecessary and, indeed, not present in the CLI API.

From the docs:

If source_dir is specified, then entry_point must point to a file located at the root of source_dir.

To reproduce
Suppose a code directory structure:

src/
|- code/
|  |- script.py
|- config/
|  |- config.yaml

The following is not allowed: Estimator(source_dir="src", entry_point="code/script.py"). Instead, the sagemaker library will strip the relative path from entry_point and the job will failed because it cannot find script.py.

Expected behavior
The entry_point should support relative paths. Indeed, the CLI API seems to support this without any issue.

Additional context
The solution seems fairly simple: delete this line:

self.entry_point = os.path.basename(self.entry_point)

I'm not sure why it's necessary to strip the relative path. In my local testing, deleting this line seems to fix the issue.

[stevenmanton]

I also validated that this works for remote jobs. I'm happy to open up a PR fix if one of the maintainers agrees it's an improvement.

[stevenmanton]

Any update on this? @mufaddal-rohawala can you take a look? Thanks!

[stevenmanton]

Is there any update on this request? I could put in a one-line PR that would fix this issue.

[cjuracek]

+1 to the issue, this would also be helpful for myself. @mufaddal-rohawala Not sure if you're the right person to tag, but I'll echo Steve