Grid Search Tuning in H2O Hydrogen Torch
- Predictive AI
- H2O Hydrogen Torch
This lesson from the Hydrogen Torch Starter Course walks through using grid search to tune an image metric learning model. It builds on the prior two lessons by creating a new experiment from an existing one and enabling a custom grid search across embedding size, backbone architecture, and learning rate. The video shows how to set embedding sizes of 256 and 1024, compare tf_efficientnet_b5_ns and resnet18 backbones, and test learning rates of 0.0005 and 0.01. It also covers how to track, filter, and manage queued experiments from the View Projects page while the grid search runs.
Key takeaways
- How to enable a custom grid search in H2O Hydrogen Torch to test multiple hyperparameter values at once
- What hyperparameters this tutorial tunes: embedding size, backbone architecture, and learning rate, each with two candidate values
- How to filter queued experiments by hyperparameter value, such as isolating runs with a specific learning rate
- Why larger backbone architectures are tested here, based on the intuition that they may improve model performance
- The options available from the three dot menu on each experiment: clone, rename, stop, or delete
Start owning your AI