Open ehsanmok opened 3 years ago
The current SageMaker module wrapping process makes debugging very hard for both training and deployment. Now for inference, decorator is the simplest kind of solutions. For example, instead of requiring users to provide model_fn (which currently takes only one argument model_dir and basically if a model needs more argument to initialize it'd make the users frustrated), we can have a decorator like
model_fn
model_dir
@sagemaker.model_fn def foo(*args, **kwargs): # one arg should be named model_dir for example ...
(Same goes for transform_fn, input_fn, output_fn). Then the decorator while expansion, looks for model_dir and everything else follows.
transform_fn
input_fn
output_fn
The current SageMaker module wrapping process makes debugging very hard for both training and deployment. Now for inference, decorator is the simplest kind of solutions. For example, instead of requiring users to provide
model_fn
(which currently takes only one argumentmodel_dir
and basically if a model needs more argument to initialize it'd make the users frustrated), we can have a decorator like(Same goes for
transform_fn
,input_fn
,output_fn
). Then the decorator while expansion, looks formodel_dir
and everything else follows.