PyTorch estimatore searches for entry_point in source_dir

[jendrikjoe]

System Information

• PyTorch:
• 0.4.0:
• 3.6:
• GPU:
• 3.6:
• No:

Describe the problem

Hey there everyone,

I am currently utilizing the PyTorch version of the SageMaker estimator and need to add additional files for it.
For that I used the source_dir option in the Estimator. However, different to what the documentation states, the entry_point has to reside in the source_dir. Is it possible to change that? Otherwise, I have to keep lots of files around ;)

Minimal repro / logs

estimator = PyTorch(entry_point="PredictionSageMaker.py",
                    source_dir='/path/to/source_dir_sagemaker',
                    role='aws role',
                    framework_version='0.4.0',
                    train_instance_count=1,
                    train_instance_type='ml.p2.xlarge',
                    hyperparameters={'epochs': 50, 'model': 'SimpleRNNNet', 'batch-size': 2048, 
                                 'learning-rate': 1e-5, 'mode': 'day'})

Traceback (most recent call last):
  File "(*deleted*)/launchSageMaker.py", line 28, in <module>
    estimator.fit({'train': '(*deleted*)', 'test': '(*deleted*)'})
  File "(*deleted*)/anaconda3/lib/python3.6/site-packages/sagemaker/estimator.py", line 176, in fit
    self._prepare_for_training(job_name=job_name)
  File "(*deleted*)/anaconda3/lib/python3.6/site-packages/sagemaker/estimator.py", line 542, in _prepare_for_training
    validate_source_dir(self.entry_point, self.source_dir)
  File "(*deleted*)/anaconda3/lib/python3.6/site-packages/sagemaker/fw_utils.py", line 87, in validate_source_dir
    raise ValueError('No file named "{}" was found in directory "{}".'.format(script, directory))
ValueError: No file named "PredictionSageMaker.py" was found in directory "(*deleted*)/source_dir_sagemaker".

[yangaws]

Hi @jendrikjoe ,

Thanks for bringing this. This is a doc error. The entry_point file has to be in the source_dir if source_dir is used.