Implementation:Facebookresearch Audiocraft Export encodec

Metadata

Knowledge Sources	Facebookresearch Audiocraft
Domains	Audio_Compression Model_Export Checkpoint_Management
Last Updated	2026-02-13 00:00 GMT

Overview

Concrete implementation of the EnCodec model export utility within Audiocraft. The export_encodec() function extracts the best model weights and configuration from a full training checkpoint and saves them in a compact format suitable for inference and distribution.

Description

The function loads a full training checkpoint from disk, extracts the best_state['model'] weights dictionary and the xp.cfg configuration, serializes the config to a YAML string, tags the package with the current library version and an exported flag, creates the output directory if needed, and saves the result using torch.save().

The exported file is significantly smaller than the training checkpoint because it omits optimizer state, EMA buffers, scheduler state, and other training metadata. The resulting file can be loaded via the standard Audiocraft pretrained model API.

Usage

Import and call directly after training completes:

from audiocraft.utils.export import export_encodec

Typically invoked from a script or notebook after a Dora training run finishes.

Code Reference

Source Location

• Repository: facebookresearch/audiocraft
• File: audiocraft/utils/export.py (lines 20--33)

Signature

def export_encodec(
    checkpoint_path: Union[Path, str],
    out_file: Union[Path, str],
) -> Union[Path, str]:
    """Export only the best state from the given EnCodec checkpoint.

    This should be used if you trained your own EnCodec model.

    Args:
        checkpoint_path: Path to the full training checkpoint file.
        out_file: Destination path for the exported lightweight checkpoint.

    Returns:
        The out_file path.
    """
    pkg = torch.load(checkpoint_path, 'cpu')
    new_pkg = {
        'best_state': pkg['best_state']['model'],
        'xp.cfg': OmegaConf.to_yaml(pkg['xp.cfg']),
        'version': __version__,
        'exported': True,
    }
    Path(out_file).parent.mkdir(exist_ok=True, parents=True)
    torch.save(new_pkg, out_file)
    return out_file

Import

from audiocraft.utils.export import export_encodec

Dependencies

• torch -- for torch.load() and torch.save() checkpoint I/O
• omegaconf -- for OmegaConf.to_yaml() serialization of the Hydra config
• audiocraft.__version__ -- library version string embedded in the export

I/O Contract

Inputs

Input Contract

Name	Type	Description
checkpoint_path	Union[Path, str]	Path to the full training checkpoint produced by CompressionSolver (via Dora experiment management). The checkpoint must contain keys best_state (with nested model dict) and xp.cfg (an OmegaConf DictConfig).
out_file	Union[Path, str]	Destination file path for the exported checkpoint. Parent directories are created automatically if they do not exist.

Outputs

Output Contract

Name	Type	Description
Return value	Union[Path, str]	The out_file path, passed through for convenience.
Exported file	File on disk	A torch.save() file containing: best_state (model state dict), xp.cfg (YAML string), version (str), exported (True).

Exported Checkpoint Structure

Exported Checkpoint Keys

Key	Type	Description
best_state	dict	The model's state_dict() containing all encoder, decoder, and quantizer weights. Flattened from the training checkpoint's best_state['model'].
xp.cfg	str	The full Hydra configuration serialized as a YAML string via OmegaConf.to_yaml(). Contains model architecture parameters, training hyperparameters, dataset configuration, and all other settings.
version	str	The audiocraft.__version__ string at export time, enabling version compatibility tracking.
exported	bool	Always True. Used by the loading code to distinguish exported checkpoints from training checkpoints.

Usage Examples

Example 1: Exporting a Trained EnCodec Model

Exporting a training checkpoint to a lightweight inference checkpoint.

from audiocraft.utils.export import export_encodec

# Export from a Dora training checkpoint
export_encodec(
    checkpoint_path='/checkpoints/xps/my_encodec_run/checkpoint.th',
    out_file='/exports/my_encodec_model.th',
)

Example 2: Loading the Exported Model

After export, the model can be loaded via the pretrained model API.

from audiocraft import models

# Load the exported checkpoint
model = models.CompressionModel.get_pretrained('/exports/my_encodec_model.th')
model.eval()

# Use for inference
import torch
audio = torch.randn(1, 1, 32000)  # 1 second at 32kHz
with torch.no_grad():
    compressed = model.encode(audio)
    reconstructed = model.decode(compressed)

Example 3: Exporting for Use as MusicGen Tokenizer

The exported compression model can be packaged alongside a MusicGen language model.

from audiocraft.utils.export import export_encodec, export_pretrained_compression_model

# First export the EnCodec model
export_encodec(
    checkpoint_path='/checkpoints/xps/my_encodec_run/checkpoint.th',
    out_file='/exports/my_encodec.th',
)

# Then package it for use with a MusicGen model
export_pretrained_compression_model(
    pretrained_encodec='/exports/my_encodec.th',
    out_file='/exports/musicgen_compression.th',
)

Related Pages

• Principle:Facebookresearch_Audiocraft_Compression_Model_Export