Amazon Sagemaker Training, Deployment, and Inference API's

[queueburt]

Pull request for an initial implementation of streamlined API's for Amazon Sagemaker. Includes 3 functions: "fit", "deploy", and "predict". The primary algorithms this was built around were XGBoost and Linear Learner, but it should be compatible with any built in algorithm that accepts 'text/csv' as a content-type.

Two additional environment variables are required to run these flows.

• METAFLOW_SAGEMAKER_REGION is the region of your Sagemaker run.
• METAFLOW_SAGEMAKER_IAM_ROLE is a Sagemaker execution role with permissions to access all appropriate cloud resources, most notably S3.

Usage consists of from metaflow import Sagemaker.

Sagemaker.fit(data, image, hyperparameters, stopping_condition, resource_config)

Returns a string object with the S3 URI of the model artifact generated by the fit.

• data is a dictionary with keys that reference Sagemaker "channel names" found here, and values that consist of CSV data with no headers or indexes.

• image is a string consisting of a Sagemaker built in algorithm container registry path, also found here. Automatic mapping will be coming soon.

• hyperparameters is a dictionary with hyperparameters for the specific algorithm referenced by image. An example for XGBoost can be found here.

• stopping_condition and resource_config are optional dictionaries for overriding some defaults, specifically that of a single ml.m4.xlarge training instance with a 5 GB volume, and a 1 hour max runtime. Syntax for those overrides can be found here

Sagemaker.deploy(model_uri, image, instanceType, instanceCount, instanceWeight, variantName)

Returns a string object with the endpoint name generated by the model deployment.

• model_uri is a string object with the S3 path for the model. This string is returned by Sagemaker.fit

• image is a string object. It should be the same image used for training.

• instanceType, instanceCount, instanceWeight, and variantName are all optional parameters for overriding the defaults of, respectively, "ml.m4.xlarge", 1, 1, and "AllTraffic".

Sagemaker.predict(data, endpoint_name)

Returns a list of predictions.

• data is a CSV object with no headers or indexes representing the features for inference.

• endpoint_name is a string object with the Sagemaker endpoint to be inferred against. This value is returned by Sagemaker.deploy.

A short example of the usage can be found here. This PR also brings in an 08-sagemaker tutorial for use with metaflow tutorials pull that demonstrates the above sample flow.

[iyerr3]

This is a legacy XGBoost image (0.72) which is not synced with upstream open-source XGboost. I suggest using
246618743249.dkr.ecr.us-west-2.amazonaws.com/sagemaker-xgboost:0.90-2-cpu-py3. This image is based on open-source XGBoost and is pinned to v0.90 of XGBoost, with the -2 indicating a SageMaker version. For a given version, backwards-compatibility is guaranteed, so better to use that instead of :latest.

[tobias-gp]

@queueburt Thanks for this great proposal! Do you have any news on the topic?

[savingoyal]

Closing this PR in favor of native support for hosting models on Sagemaker with Metaflow