Can't pass new entry_point to estimator.deploy() #1427
Description
Describe the bug
Can't pass a new entry_point to estimator.deploy(). Only way is to set the entry_point property on the estimator first and then deploy. Unintuitive that entry_point would be ignored in deploy method.
To reproduce
estimator = sagemaker.sklearn.SKLearn(
image_name=str(custom_image),
source_dir=str(Path(current_folder, 'src').resolve()),
entry_point='entry_point.py',
hyperparameters=hyperparameters,
role=config.SAGEMAKER_IAM_ROLE,
train_instance_count=1,
train_instance_type=train_instance_type,
sagemaker_session=session,
output_path='s3://' + str(Path(config.S3_BUCKET, config.OUTPUTS_S3_PREFIX)),
code_location='s3://' + str(Path(config.S3_BUCKET, config.OUTPUTS_S3_PREFIX))
)
estimator.fit(...)
predictor = estimator.deploy(
endpoint_name=endpoint_name,
instance_type=deploy_instance_type,
initial_instance_count=1,
entry_point='another_entry_point.py'
)
Expected behavior
Would expect the deployed endpoint to use another_entry_point.py rather than entry_point.py.
System information
A description of your system. Please provide:
- SageMaker Python SDK version: 1.55.3
- Framework name (eg. PyTorch) or algorithm (eg. KMeans): SKLearn
- Framework version: 0.20.0-cpu-py3
- Python version: Python 3
- CPU or GPU: CPU
- Custom Docker image (Y/N): Y (but extends SKLearn container from SM)
Additional context
Seems like the issue is that the kwarg for entry_point is removed and replaced with the property (from the training) instead. See https://github.com/aws/sagemaker-python-sdk/blob/master/src/sagemaker/sklearn/estimator.py#L167. Should be able to override without having to set property on the estimator. Could effect other frameworks too.