Sagemaker-python-sdk: How to save model?

[meownoid]

How can I save model after the model.fit()so next time I could load and deploy it using model.deploy()?

[andremoeller]

Hi @meownoid ,

Model artifacts from a training job are uploaded to S3 when training completes, so the model is already saved. You can use attach() with the training job name from fit() to deploy this model:

https://github.com/aws/sagemaker-python-sdk/blob/157d8670977243f7f77327175d40364c885482b3/src/sagemaker/estimator.py#L531-L550

Does this answer your question?

[yangaws]

@meownoid

I will close the issue now. Feel free to reply if you have any further questions.

[ricoms]

@andremoeller, that is true if you did the training on the SageMaker, but in my example, I did not. I train everything in EC2 instances, then saved the architecture file in json format and the weights in h5 format.

Is there any way that I can deploy and estimator to SageMaker from this files?

I'm following this material: https://github.com/samir-souza/ai-workshop/blob/master/lab/03_ImageClassification/ImageClassificationFashionMNIST.ipynb my application is similar.

[KobaKhit]

Follow up here. What format does sagemaker save models in? To be more precise, after you unzip the model.tar.gz what is the resulting format of the model file. It does not seem to be mleap. Was wondering if its possible to load the model.tar.gz produced by sagemaker training job using an mleap like library.

[laurenyu]

@KobaKhit the format of the model depends on how you trained the image, e.g. if you use TensorFlow, then the model is saved in the format specified by your training script.