trainer

Pure Rust trainer that fits NAM-compatible models from WAV pair(s) and exports .nam.

A2-focused capabilities

• --a2-mode: enables A2 default trainer presets (longer run + windowing defaults).
• --a2-strict: enforces strict A2 checks (48kHz, min length, Proteus-like detection).
• --detect-input-version: strong + weak version/hash detection (v1/v1.1/v2/v3/proteus patterns).
• --run-calibration --calibration-mode none|rms|rms-dc-block --calibration-target-rms <f>.
• A2 latency analysis is computed and exported in metadata (recommended, manual, warning flags).
• A2 data checks (clipping/DC/SNR heuristics) are computed and exported in metadata.
• Strict thresholds: --a2-min-snr-db, --a2-max-dc-offset.

Architectures

• Linear
• ConvNet
• LSTM
• WaveNet

Training controls

• --loss mse|esr
• --lr-schedule constant|cosine
• --warmup-steps, --min-learning-rate
• --weight-decay, --momentum, --grad-clip
• --window-size, --window-hop
• checkpointing/resume with optimizer state:
  • --checkpoint-path, --checkpoint-every, --resume-checkpoint

Config-driven mode

Use --trainer-config trainer.json to override core trainer values.

Notes

• WAV files must be mono and pair-matched in sample rate and length.
• Vulkan acceleration is available for constrained fast paths (Linear + limited ConvNet case).
• Full bit-exact equivalence still requires iterative validation/tuning.