Implementation:Facebookresearch Audiocraft MagnetSolver

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

Overview

Concrete tool for training MAGNeT masked audio generation models using span-based masking and per-codebook cross-entropy loss.

Description

MagnetSolver extends MusicGenSolver to implement the MAGNeT training procedure. Each training step randomly selects a codebook level, applies span-based masking with a cosine schedule, and computes cross-entropy loss only on the masked positions. AudioMagnetSolver further specializes this for environmental sound generation.

Usage

Use this solver when training MAGNeT models for non-autoregressive text-to-music or text-to-sound generation.

Code Reference

Source Location

• Repository: Facebookresearch_Audiocraft
• File: audiocraft/solvers/magnet.py
• Lines: 1-276

Signature

class MagnetSolver(musicgen.MusicGenSolver):
    def __init__(self, cfg: DictConfig): ...
    def run_step(self, idx: int, batch, metrics: dict) -> dict: ...
    def _get_mask(self, mask_probs, B, T, device) -> torch.Tensor: ...
    def _compute_cross_entropy_magnet(self, logits, targets, mask, stage) -> torch.Tensor: ...

class AudioMagnetSolver(MagnetSolver):
    DATASET_TYPE = builders.DatasetType.SOUND

Import

from audiocraft.solvers.magnet import MagnetSolver, AudioMagnetSolver

I/O Contract

Inputs

Name	Type	Required	Description
batch	tuple	Yes	(audio_tokens, segment_attributes) tuple
cfg	DictConfig	Yes	Hydra config with masking params

Outputs

Name	Type	Description
metrics	dict	Training metrics including ce, ppl, lr

Related Pages

• Principle:Facebookresearch_Audiocraft_Masked_Parallel_Token_Generation