Learning Rate Range Test
Diagnostic method that exponentially increases the learning rate while observing loss – finds the optimal LR range in a single training run.
The LR finder exponentially increases the learning rate over one epoch and finds the optimal range at the steepest loss decrease – saves hours of hyperparameter tuning.
Explanation
The LR increases from very small to very large. The optimal range is where the loss decreases most steeply (before the divergence point). Takes only 1 epoch.
Marketing Relevance
Eliminates guessing for learning rate selection. Essential for one-cycle policy and saves hours of trial-and-error experiments.
Common Pitfalls
Result depends on batch size and model initialization. Gives a range, not an exact value. Not applicable for all scheduler combinations.
Origin & History
Leslie Smith (2015) introduced the LR range test in "Cyclical Learning Rates for Training Neural Networks." It became an integral part of the super-convergence methodology and the Fastai library.
Comparisons & Differences
Learning Rate Range Test vs. Grid Search (für LR)
Grid search trains completely for each LR (expensive); LR finder finds the range in one epoch (cheap and fast).
Learning Rate Range Test vs. Bayesian Optimization
Bayesian optimization tunes all hyperparameters simultaneously; LR finder is a quick one-time test for learning rate only.