Expose API for model building / publishing automation
When offering ML Prediction as Service to our customers it can be crucial to automate model creation, as different models need to be generated for different customers. It'd probably be enough to enable importing R scripts via an API or WebService call to solve this.
Please see http://aka.ms/amlps
Rui Quintino commented
Hi there, I think azuremlps isn't a solution, at least yet , for this request needs. As far as I know we can't create an experiment from scratch, enabling easy generation of models, ex: importing R scripts, only export/reimport.
So this issue shouldn't be closed as completed. Still applies. Or needs clarification on how can we do this with azuremlps?
"Please note that the exported JSON file only contains references to the exact instance and version of the assets (modules, trained models, datasets, etc.). The assets themselves are NOT serialized into the JSON file. As a consequence, when you import it back into the Workspace, make sure the exact same instance and version of those assets do exist in the Workspace, otherwise you will not be able to create a valid Experiment. Also, please make sure you use absolute path when referring to the json file."
If you could expose all UI operations as API that would awesome! I think the new Azure Portal uses this approach. What can be done in the UI can be done in API because UI is built on top of API.
Thanks - excellent progress guys. To further pare down this idea - for an initial version:
Just the ability to replace (not even insert/delete), via API, an existing R script "module" within an experiment would handle about 50% of use cases, at least for me.
We have released the Retraining APIs enabling programmatic retraining. See this article for more details: http://azure.microsoft.com/en-us/documentation/articles/machine-learning-retrain-models-programmatically/
Neeraj (MSFT) commented
We will support API for retraining the models and update web services with retrained model. We are working on providing the Retraining API for AzureML as a top priority feature in an upcoming release
Rui Quintino commented
3 votes for this one! top priority, would open a completely new range of possibilities. The exposed api should be able to fully build & train experiments from scratch, and also expose model score results for external evaluation & reporting (even allow for continuous integration?). Seems already clear that there's some very typical ml studio workflow patterns that would greatly benefit from model automation (read, split, train/param sweep, score, evaluate...). & this would certainly spike "billable" usage of azure ml :) Also related, being able to download the model as file, and use it as a template for the api.
Bess Walker commented
I'd rather have automated model building and publishing without R scripts being necessarily involved, but any automation would be handier than none.