search

aws / sagemaker-mxnet-training-toolkit

Toolkit for running MXNet training scripts on SageMaker. Dockerfiles used for building SageMaker MXNet Containers are at https://github.com/aws/deep-learning-containers.
Apache License 2.0
60 stars 55 forks source link

DEFAULT_FILENAMES hardcoded #30

Closed dmadeka closed 6 years ago

dmadeka commented 6 years ago

Im not sure why the default file names are hardcoded, and not a member of ModuleTransformer- wouldn't it make sense to have them be parameters that are set to those values by default?

jesterhazy commented 6 years ago

Thanks for using SageMaker!

I assume you mean the DEFAULT_MODEL_FILENAMES defined in train.py here.

These filenames are used by the default_save method in this module, and the default_model_fn in the serving code here.

If there were parameterized, the onus would be on users to make sure they provide the same filename parameters when they create hosting endpoints. We'd also need to create a way for users to pass these parameters to their training and hosting jobs.

If you want to use different filenames, you certainly can. Just provide your own save(mode, model_dir) function in your training script. If you plan to create hosting endpoints, provide a compatible model_fn(model_dir) function in your hosting script.

Closing this ticket. Please reopen if you still have questions about this.

dmadeka commented 6 years ago

Im not sure I understand. If you read these from a CONFIG file - it would solve all these problems and not force users to rename every model as model-symbol.json.

dmadeka commented 6 years ago

Also @jesterhazy this isn't a "ticket", this is a GitHub issue. You cant "reopen" it unless your an admin. If you "reopen" it - we can continue the discussion in the open.