Implementation:Facebookresearch Audiocraft WatermarkSolver

Knowledge Sources	Facebookresearch_Audiocraft AudioSeal
Domains	Audio_Watermarking, Training
Last Updated	2026-02-14 01:00 GMT

Overview

Concrete tool for training AudioSeal watermarking models that embed and detect imperceptible watermarks in audio provided by the AudioCraft library.

Description

WatermarkSolver extends StandardSolver to implement the complete training loop for audio watermarking models. It trains a generator (watermark embedder) and detector jointly, using adversarial losses for audio quality, detection losses for watermark presence/absence, and multi-bit message decoding losses. The solver includes crop/shuffle/pad data augmentation for localization training, evaluation with ViSQOL/SI-SNR/PESQ quality metrics, and augmentation robustness testing.

Usage

Use this solver when training AudioSeal watermarking models. It manages the full training pipeline including balanced loss optimization, audio augmentation robustness, and localization evaluation.

Code Reference

Source Location

• Repository: Facebookresearch_Audiocraft
• File: audiocraft/solvers/watermark.py
• Lines: 1-716

Signature

class WatermarkSolver(StandardSolver):
    def __init__(self, cfg: DictConfig):
        """Initialize watermark solver with losses, augmentations, and model."""

    def run_step(self, idx: int, batch: torch.Tensor, metrics: dict):
        """Perform one training/valid step: embed watermark, augment, detect, compute losses."""

    def evaluate(self) -> dict:
        """Run evaluation with quality metrics, augmentation robustness, and localization."""

    @staticmethod
    def model_from_checkpoint(checkpoint_path, device="cpu") -> WMModel:
        """Load a trained watermark model from checkpoint."""

Import

from audiocraft.solvers.watermark import WatermarkSolver

I/O Contract

Inputs

Name	Type	Required	Description
batch	torch.Tensor	Yes	Audio batch [B, C, T]
cfg	DictConfig	Yes	Hydra configuration with loss weights, augmentation params

Outputs

Name	Type	Description
metrics	dict	Training metrics including g_loss, d_loss, pesq, ratio1, ratio2
WMModel	WMModel	Trained watermark model (from checkpoint loading)

Usage Examples

Training via Dora

# WatermarkSolver is launched via Dora experiment grids
# Example: dora run solver=watermark/default

# Loading a trained model from checkpoint:
from audiocraft.solvers.watermark import WatermarkSolver

model = WatermarkSolver.model_from_checkpoint(
    'path/to/checkpoint.th',
    device='cuda'
)

Related Pages

• Principle:Facebookresearch_Audiocraft_Audio_Watermark_Training