Is it possible SageMaker HyperParameter Tuning Job without sagemaker-training tool kit

Question

I had made my custom training image so It can be conducted through CreateTrainingJob, not sagemaker training took kit (requiring "ContainerEntrypoint" option).

But when I'm trying to run HyperParameter Tuning Job, it is not allowed to add "ContainerEntrypoint" option in "AlgorithmSpecification" field.

Is it impossible to run HyperParameter Tuning Job with training images that can not be run without sagemaker training toolkit?

Thanks!

Accepted Answer

Hi there!

For custom training image, you can specify the entrypoint in your Dockerfile.

Below are some code snippet as well as links you can use as reference:

https://github.com/aws/amazon-sagemaker-examples/blob/main/hyperparameter_tuning/rapids_bring_your_own/code/Dockerfile
```
ENTRYPOINT ["./entrypoint.sh"]
```
https://github.com/aws/amazon-sagemaker-examples/blob/main/hyperparameter_tuning/keras_bring_your_own/Dockerfile
```
ENTRYPOINT ["python", "-m", "trainer.start"]
```

Furthermore, SageMaker Training Toolkit is a nicely wrapped up python package for you to use and eases the process of creating a custom training image, it's no different from implementing the logic yourself.

So it is definitely possible to run HyperParameter Tuning Job using custom containers without using our SageMaker Training Toolkit.

Is it possible SageMaker HyperParameter Tuning Job without sagemaker-training tool kit

Contenus pertinents