Implementation:Facebookresearch Audiocraft SampleManager

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

Overview

Concrete tool for storing, retrieving, and comparing generated audio samples across training epochs provided by the AudioCraft library.

Description

SampleManager provides an API for managing audio sample outputs during training. It stores generated audio samples alongside ground truth references, organizes them by epoch, and supports XP-based path resolution via Dora. The manager handles distributed training by only writing on rank zero, generates unique sample IDs, and maintains a mapping from reference to sample IDs for comparison across epochs.

Usage

Import this class in solver code when you need to save generated audio samples during the generate() stage of training. It is used by all AudioCraft solvers (MusicGenSolver, CompressionSolver, DiffusionSolver, WatermarkSolver) to store evaluation samples.

Code Reference

Source Location

• Repository: Facebookresearch_Audiocraft
• File: audiocraft/utils/samples/manager.py
• Lines: 1-386

Signature

class SampleManager:
    def __init__(self, xp: dora.XP, map_reference_to_sample_id: bool = False):
        """
        Args:
            xp: Dora experiment object for path resolution.
            map_reference_to_sample_id: If True, map reference filenames to sample IDs.
        """

    def add_samples(self, samples: torch.Tensor, epoch: int,
                    ground_truth_wavs: tp.Optional[torch.Tensor] = None,
                    conditioning: tp.Optional[tp.List] = None) -> None:
        """Save generated samples and optional ground truth for the given epoch."""

    def get_samples(self, epoch: int) -> tp.List[Sample]:
        """Retrieve stored samples for the given epoch."""

Import

from audiocraft.utils.samples.manager import SampleManager

I/O Contract

Inputs

Name	Type	Required	Description
samples	torch.Tensor	Yes	Generated audio samples [B, C, T]
epoch	int	Yes	Current training epoch
ground_truth_wavs	torch.Tensor	No	Reference audio for comparison
conditioning	list	No	Conditioning attributes for each sample

Outputs

Name	Type	Description
Sample files	WAV files	Saved to XP directory organized by epoch
Sample objects	List[Sample]	Retrieved sample metadata and paths

Usage Examples

from audiocraft.utils.samples.manager import SampleManager

# Used inside a solver's generate() method:
sample_manager = SampleManager(self.xp, map_reference_to_sample_id=True)
sample_manager.add_samples(
    generated_wav.cpu(),
    epoch=self.epoch,
    ground_truth_wavs=reference.cpu(),
)

Related Pages

• Principle:Facebookresearch_Audiocraft_Sample_Management