mboehm7 commented on pull request #1125: URL: https://github.com/apache/systemds/pull/1125#issuecomment-747535307
In general, that's a good starting point. We had another use case of importing sk-learn pipelines in mind, but adding the sklearn-onnx-dml model converter is also an interesting exploratory project which we can put into staging (so the dependencies do not affect our runtime). For sk-learn pipelines we would have taken some example pipelines (see make_pipeline) composed of primitives like scaling, encoding, cross validation, model training and converted that into a DML script that calls the respective existing builtin functions for scale, encoding, etc. ONNX on the other hand is primarily used as exchange format for neural network specifications and models. So it's not a direct fit for encoding such ML pipelines. However, focusing on the exchange of a trained model and its use, many linear models can be encoded via fully connected layers and similar operations. So from my perspective, it's fine to pursue this project. I moved the existing onnx converter from our python API to staging because an incompatible API change. Fixing the onnx importer in the process would be nice, and should not be that much overhead. ---------------------------------------------------------------- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. For queries about this service, please contact Infrastructure at: [email protected]
