micahjsmith
commented
5 years ago Here are some thoughts on implementation:
- For each method in
atm/methods/*.json, find a corresponding primitive inHDI-Project/MLPrimitives/mlblocks_primitives, such as RandomForestClassifier https://github.com/HDI-Project/MLPrimitives/blob/master/mlblocks_primitives/sklearn.ensemble.RandomForestClassifier.json - Replace
atm/methods/dir withatm.constants.primitivesor similar, a list of strings identifying MLPrimitives - Use MLBlocks to make a pipeline: https://github.com/HDI-Project/ATM/blob/master/atm/model.py#L95
Ultimately, we want people to be able to include/upload their own primitives for use with MLBlocks.
Currently, ATM defines learning methods (classification only) in json files within the repo. The MLBlocks/MLPrimitives projects provide similar functionality, but include many more methods, including neural networks, gradient boosted machines, and regression models (#72). We should explore using these projects as the backend for learning methods.