Learning Rate

Controls how quickly the model adapts during training.

The learning rate determines the step size when updating model weights. This is one of the most important hyperparameters.

Too high

• Training may diverge (loss increases or oscillates)

• Model fails to converge

Too low

• Very slow training

• May get stuck in poor solutions

Available options

• 1e-5: Very conservative, stable

• 5e-5: Good for fine-tuning

• 1e-4: Standard default, works well for most cases

• 2e-4: Faster training, slightly less stable

Recommendations

• Start with 1e-4 (the default)

• Reduce to 1e-5 if training is unstable

• Stage 2 of autoStructN2V often benefits from a lower rate