Auto-train, auto-tune & auto-serve the best TF-DF model directly from CSV files

[rishiraj]

As my GSoC contributions are almost over, as a part of my additional work, I'm working on developing a layer above TF-DF that:

• auto train TF-DF directly from CSV files
• auto tune TF-DF using Keras Tuner
• auto serve best TF-DF model using FastAPI

This will help in making TF-DF a favorite choice for dealing with tabular data in the Kaggle community where most training data are in CSV format.

[janpfeifer]

That sounds awesome @rishiraj !! This tutorial may be a starting point.

Some optional suggestions/ideas:

• For small datasets, add support for n-fold cross-validation. This will make the best use of the data available. Also for small datasets it's so fast to train, that it doesn't cost much.
• For large datasets, two suggestions, we often make here:
  • Do most of the hyperparameter tuning on a smaller (sub-sampled) dataset for speed. It may not be optimal, but often due to resources constraints it's more feasible. Once the best parameters are found, train on the whole data.
  • Parallelize the hyperparameter tuning in various machines -- requires knowhow of whatever cloud solution one uses to parallelize. The complication here is not ML, but dealing with starting jobs and collecing results.

cheers!