Provide SageMaker compatible docker container about gluonts

[jaheba]

One thing I’m unsure about is how we dispatch between different models.

Also, do we want to support training of more than one model per run?

/cc @aalexandrov @lostella

[lostella]

Also, do we want to support training of more than one model per run?

Not sure why we would want to do that: the way I see it, each job would be one training, and should produce one model artifact that should be servable

[jaheba]

Although I think there is a point to be made that it could be simpler to use, SageMaker has the concept of one training-configuration is one job.

[aalexandrov]

One thing I’m unsure about is how we dispatch between different models.

The approach used in #151 is to dispatch the Estimator based on an environment variable SWIST_ESTIMATOR_CLASS (should really be GLUONTS_ESTIMATOR_CLASS) or, if that variable is not set, through a special estimator_class hyperparamter.

[jaheba]

through a special estimator_class hyperparamter

Which is not available during inference. Except, if we store it as part of the model. And that would imply that all models have to go through training. Not sure I like that.

[jaheba]

Another thought:

Since we have models which don't require offline training, we can make the shell accept Forecasters, where:

Forecaster = Union[Predictor, Estimator]

That way we would skip training if a Predictor was specified and just start with the test-step.

[lostella]

through a special estimator_class hyperparamter

Which is not available during inference. Except, if we store it as part of the model. And that would imply that all models have to go through training. Not sure I like that.

@jaheba at inference time there should be no Estimator involved; the container should be given instructions on how to instantiate a Predictor: this will be whatever type of Predictor with its own settings, or a GluonPredictor (either a SymbolBlockPredictor or a RepresentableBlockPredictor, see https://github.com/awslabs/gluon-ts/blob/master/src/gluonts/model/predictor.py#L170 and all the lines that follow) which will be created using the network that was written to S3 after some training job.

[jaheba]

the container should be given instructions on how to instantiate a Predictor

That was exactly my point. During training we can pass a hyper-parameter to select the right Estimator/Predictor. But during inference we would depend on the generated model, which not all models require.