search

google / nitroml

NitroML is a modular, portable, and scalable model-quality benchmarking framework for Machine Learning and Automated Machine Learning (AutoML) pipelines.
Apache License 2.0
41 stars 6 forks source link

[MetaLearning] Custom Tuner #39

Closed cweill closed 4 years ago

cweill commented 4 years ago

We want to find good configurations early on. So the custom tuner contains either one or two Keras Tuners:

  1. If Keras Tuner supports warm-starting (suggested trials): Use those first, then fallback to the original search space.
  2. Else create one Keras Tuner for the meta-learner-defined search space, and a second for the original search space.

This ensures that if our metalearner fails to find good configurations, it won't harm final performance if the user requests sufficient trials.