search

astromz / gce_ml

Easier ML model training for Google Compute Engine (GCE). It is similar to Cloud ML but is more cost effective and offers more flexibility
7 stars 2 forks source link

Hyperparameter Tuning #1

Open liamb315 opened 6 years ago

liamb315 commented 6 years ago

Thanks for creating this! Do you have any expectation to add in hyperparameter tuning support via the Cloud-ML tools? Currently, it looks like the VMs running the Cloud-ML hyperparameter tuning do not allow for customized boot images.

astromz commented 6 years ago

Thanks for your interest. That's a good question. Right now it's hard to incorporate the hyperparameter tuning functionality available in Cloud ML. The hyperparameter tuning in Cloud ML uses Gaussian Process bandits, which is much more sophisticated than simple grid search or random search. This package right now is just for simple model training. Grid-search or random search based model tuning can be specified in the trainer code.