// the find
minimaxir/automl-gs
Provide an input CSV and a target field to predict, generate a model + code to run it.
automl-gs takes a CSV and a target column and runs a random hyperparameter search to produce a trained model plus the Python code that built it. The generated code is plain Python using your chosen framework (TensorFlow or XGBoost), so you own the output and aren't locked into the tool. Aimed at people who want a baseline model quickly without hand-crafting preprocessing pipelines.
The code-generation approach is the right call — you get readable, tweakable output instead of a black box artifact. Encoders are serialized as JSON rather than pickled objects, which sidesteps the usual deserialization brittleness. The Jinja template architecture makes it straightforward to understand exactly what code gets emitted for a given data type. Metrics and hyperparameters are externalized to YAML files, so the search space is auditable without reading Python.
Abandoned since October 2019 — TensorFlow 1.x/2.x API churn alone means the generated TensorFlow code almost certainly doesn't run against a current install without fixes. Only two frameworks shipped out of the four listed; CatBoost and LightGBM are still in the 'to be implemented' column after six years. Hyperparameter search is random, not Bayesian, so you're burning compute on uniform sampling when Optuna or similar has been table stakes for years. Windows support was a known bug at abandonment, not a future roadmap item.