Bash script mode support across all estimators

aws / sagemaker-training-toolkit

Train machine learning models within a 🐳 Docker container using 🧠 Amazon SageMaker.

Apache License 2.0

496 stars 118 forks source link

Bash script mode support across all estimators #75

Closed ehsanmok closed 4 years ago

ehsanmok commented 4 years ago

It'd be very useful to haveentry_point across the estimators handling bash scripts where cmd args are passed as hyperparameters in an *Estimator. It seems TF supports bash script mode but no such support for PyTorch, MXNet.

ehsanmok commented 4 years ago

I've tested this for PyTorch entry_point=“train.sh” and the bash script handling is wrong actually with error

"Error while finding module specification”. wants to run like python -m scripts/train.sh

@laurenyu has pointed to the faulty logic here and here.

In my case, I have a python package with setup.py then so it becomes a PYTHON_PACKAGE but it should check for the entry_point file ext first then decides how to run it not the other way around.