Closed exalate-issue-sync[bot] closed 1 year ago
Jan Sterba commented: was merged into onnx https://github.com/onnx/onnxmltools/pull/358 will be nex onnxmltools release
maybe we might want to add a chapter to Productionizing H2O doc chapter? [~accountid:557058:04659f86-fbfe-4d01-90c9-146c34df6ee6] WDYT?
Michal Kurka commented: [~accountid:5bd82be4216ccf49622e5137] sounds like good idea to me, cc: [~accountid:557058:6e44bc1a-dd50-499b-a331-2e049f28773b]
JIRA Issue Migration Info
Jira Issue: PUBDEV-7057 Assignee: Jan Sterba Reporter: Michal Kurka State: Closed Fix Version: 3.28.0.2 Attachments: N/A Development PRs: Available
Linked PRs from JIRA
ONNX is a standardized format used to represent ML models: https://onnx.ai/
Popular ML frameworks provide support for ONNX via convertors: https://github.com/onnx/onnxmltools/tree/master/onnxmltools/convert (XGBoost, Spark).
The goal of this task is to create a convertor from H2O GBM MOJO to ONNX representation. Inspiration could be taken from XGBoost implementation of the convertor: https://github.com/onnx/onnxmltools/tree/master/onnxmltools/convert/xgboost
Not all distributions / objective functions needs to be supported on Day 1 (XGBoost doesn't support Tweedie for example). Main challenge will be most likely to handle categorical encoding. For initial POC it is also okay to do just 1-hot encoded models.