AdamAnomalyGuardMode

Enums · AiDotNet.Enums

Policy for the PyTorch GradScaler-style anomaly guard on the Adam optimizer’s tape-based Step. The guard scans every gradient element for NaN/Inf and skips the entire step when one is found, preventing permanent poisoning of the m/v moment accumulators.

Fields

Field Summary
AdamAnomalyGuardMode Always Always scan gradients before each step.
AdamAnomalyGuardMode Auto Default: scan gradients before each step.
AdamAnomalyGuardMode Never Skip the anomaly scan entirely.
int value__