TrainingDiagnosticLevel

Enums · AiDotNet.Configuration

Verbosity level for training-pipeline diagnostic output (gradient norms, optimizer step traces, tape replay events, etc.).

For Beginners

Think of this like log levels in any logging framework: Silent is OFF, Minimal is “just headline events”, Verbose is “everything per-batch”, PerStep is “per-parameter granularity — expensive but exhaustive”.

How It Works

Companion to GpuDiagnosticLevel. AiDotNet’s training pipeline (TrainWithTape, fused-optimizer fast path, gradient-tape backward) can fail in subtle ways — wrong-direction gradient flow, dropped layer gradients, optimizer skipping parameters. This enum gives consumers fine-grained control over training-side diagnostic output so production code can keep it silent and regression tests can opt in to detailed traces.

Fields

Field Summary
TrainingDiagnosticLevel Minimal Headline events only: loss value per Train call, optimizer step failure, gradient explosion/vanish warnings.
TrainingDiagnosticLevel PerStep Per-parameter granularity: every parameter tensor’s gradient L2 norm after backward, fused-optimizer fast-path hit/miss, scheduler ticks.
TrainingDiagnosticLevel Silent No training-pipeline diagnostic output.
TrainingDiagnosticLevel Verbose Per-batch trace: tape forward/backward boundaries, loss, aggregate gradient norm, optimizer step result.
int value__