Ludwig on Ray: distributed AutoML setup and tuning

Travis Addair explains how Ludwig uses a declarative configuration to build deep learning models from typed inputs and outputs. He describes an open source framework that combines text, categorical and other data in one model, with options to adjust learning rates, optimizers and model components rather than rely entirely on defaults.

Uploaded on November 17, 2021, the talk covers Ludwig v0.4. The supplied metadata does not give a recording date. Its setup examples and capabilities describe that release, rather than current installation guidance.

The main explanation shows how to move a configuration from training on a local laptop to execution on a Ray cluster. It presents architecture and configuration examples rather than a live completed training run. Addair contrasts the original pandas-based workflow, which required data to fit in memory, with Dask preprocessing and evaluation plus Horovod distributed training. He says the same model configuration can carry over to the cluster. The example uses AWS EC2 instances, a CPU head node and workers with one T4 GPU each.

The talk also explains Ray Tune configuration for parallel hyperparameter search, including search constraints, an optimization objective and custom search algorithms. Addair presents an experimental auto-train API that would infer a configuration from a prediction target, explicitly warning that its APIs could change. The closing section outlines Predibase's proposed SQL-style approach to prediction, which he describes as still in stealth.