Implementation:Facebookresearch Audiocraft WMLoss

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

Overview

Concrete tool for computing watermark detection and multi-bit message decoding losses for AudioSeal training.

Description

This module provides WMDetectionLoss for binary watermark presence/absence classification using NLL loss on detector outputs, and WMMbLoss for multi-bit message decoding using either BCE or MSE loss. Both losses support masked regions for localization-aware training.

Usage

Import these losses when training AudioSeal watermarking models. They are used by WatermarkSolver during the training loop.

Code Reference

Source Location

• Repository: Facebookresearch_Audiocraft
• File: audiocraft/losses/wmloss.py
• Lines: 1-104

Signature

class WMDetectionLoss(nn.Module):
    def __init__(self, p_weight: float = 1.0, n_weight: float = 1.0):
        """Binary watermark detection loss."""
    def forward(self, positive, negative, mask, message) -> torch.Tensor: ...

class WMMbLoss(nn.Module):
    def __init__(self, temperature: float, loss_type: Literal["bce", "mse"] = "bce"):
        """Multi-bit message decoding loss."""
    def forward(self, positive, negative, mask, message) -> torch.Tensor: ...

Import

from audiocraft.losses.wmloss import WMDetectionLoss, WMMbLoss

I/O Contract

Inputs

Name	Type	Required	Description
positive	torch.Tensor	Yes	Detector output for watermarked audio [B, 2+nbits, T]
negative	torch.Tensor	Yes	Detector output for clean audio [B, 2+nbits, T]
mask	torch.Tensor	Yes	Watermark region mask [B, 1, T]
message	torch.Tensor	No	Original message bits [B, nbits] (for WMMbLoss)

Outputs

Name	Type	Description
loss	torch.Tensor	Scalar loss value

Usage Examples

from audiocraft.losses.wmloss import WMDetectionLoss, WMMbLoss

det_loss = WMDetectionLoss(p_weight=1.0, n_weight=1.0)
mb_loss = WMMbLoss(temperature=1.0, loss_type="bce")

loss = det_loss(positive, negative, mask, None) + mb_loss(positive, negative, mask, message)

Related Pages

• Principle:Facebookresearch_Audiocraft_Audio_Watermark_Training