Learning Rate

Controls how quickly the model adapts during training.

The learning rate determines the step size when updating model weights. It lives under Advanced Options because the default is the validated publication value and rarely needs changing.

The control

• A numeric field accepting values from 1e-6 to 0.01

• Default: 0.001 (used by every preset and by both the N2V and StructN2V branches)

Too high

• Training may diverge (loss increases or oscillates)

• Model fails to converge

Too low

• Very slow training

• May get stuck in poor solutions

Recommendations

• Keep the default of 0.001 for most data

• Lower it (for example to 1e-4) if the loss curve is unstable

• Values are typed directly, so any number in the 1e-6 to 0.01 range is accepted