Adding support for Baseline Model #58
Conversation
| if model_type is ModelType.custom or model_type is ModelType.rasa: | ||
| return "@artifacts([PickleArtifact('function'), PickleArtifact('kwargs')])" | ||
| elif model_type is ModelType.sklearn: | ||
| return f"@artifacts([PickleArtifact('model'), PickleArtifact('function'), PickleArtifact('kwargs')])" |
There was a problem hiding this comment.
Did you test this for regular sklearn models? The PickeArtifact is more general, so it should work for sklearn models. However, I think it's safer to use BentoML's SklearnModelArtifact for regular sklearn models. You probably made this choice because of the Too many open files error, so I would suggest using PickleArtifact only for the models from the quick baseline and not for all sklearn models we deploy
There was a problem hiding this comment.
Talked to @gbayomi about this: "PickleArtifact, according to Bo (from BentoML), is the same as SklearnArtifact, but more reliable, that’s why I made the change"
examples/tabular-classification/sklearn/churn-classifier/churn-classifier-sklearn.ipynb Outdated Show resolved Hide resolved
…e are issues with the dataset
gustavocidornelas force-pushed the gabe/baseline-model branch from 6759f87 to e597790 Compare 
openlayer/projects.py Outdated
| *args, | ||
| **kwargs, | ||
| ) -> Model: | ||
| def add_model(self, *args, **kwargs,) -> Model: |
There was a problem hiding this comment.
this auto-formatting violates black. Can you run black on the openlayer directory and push?
| requests | ||
| tqdm | ||
| marshmallow | ||
| scikit-learn==0.24.1 |
There was a problem hiding this comment.
this is going to cause problems / warnings if the user is uploading their own scikit-learn models with a different package version
| categorical_feature_names=categorical_feature_names, | ||
| requirements_txt_file="auto-requirements.txt", | ||
| col_names=col_names, | ||
| preprocessor=AutoMunge(), |
There was a problem hiding this comment.
anything that isn't exposed via the api should just be the default value for the arg.
Then you can remove from Automunge import AutoMunge from this file
| from sklearn.ensemble import VotingClassifier | ||
| from sklearn.preprocessing import LabelEncoder | ||
| | ||
| # ------------------------------- MONKEY PATCH ------------------------------- # |
There was a problem hiding this comment.
The code for creating a baseline in this manner is now available for anyone copy / use without actually using Openlayer. Are you sure you don't want to run this in the backend?
There was a problem hiding this comment.
That's true. Personally, I would prefer running on the backend. cc: @gbayomi
I'm adding support for baseline models.
Baselines come from auto-sklearn (https://automl.github.io/auto-sklearn/master/) and data processing comes from Automunge (https://www.automunge.com/).
The baseline.py basically extracts pure sklearn models from auto-sklearn. Then, we use Automunge to create processor object and build a prediction function.
@gustavocidornelas will help me merge this PR so I can go back to building the sales pipeline (:
The logic is relatively simple though!