Implementation:Facebookresearch Audiocraft Export pretrained compression model

Overview

Export_pretrained_compression_model packages a compression model (EnCodec or DAC) for distribution alongside a language model. It handles both pretrained model references and custom-trained model files.

Implements

Principle:Facebookresearch_Audiocraft_Pretrained_Compression_Export

API Signature

def export_pretrained_compression_model(
    pretrained_encodec: str,
    out_file: Union[Path, str]
) -> None

Source: audiocraft/utils/export.py, lines 36-58

Import:

from audiocraft.utils.export import export_pretrained_compression_model

Parameters

Parameter	Type	Default	Description
`pretrained_encodec`	`str`	required	Either a pretrained model identifier (e.g., `'facebook/encodec_32khz'`) or a filesystem path to a custom-trained, already-exported compression model file
`out_file`	`Union[Path, str]`	required	Destination path for the compression model export (typically `compression_state_dict.bin`)

Return Value

Returns None. The export file is written to out_file as a side effect.

Dependencies

torch -- for loading and saving checkpoint files
audiocraft.__version__ -- for embedding the version tag

How It Works

from audiocraft.utils.export import export_pretrained_compression_model

# Export using a standard pretrained model reference
export_pretrained_compression_model(
    pretrained_encodec='facebook/encodec_32khz',
    out_file='/path/to/exports/my_model/compression_state_dict.bin'
)

# Export using a custom-trained compression model
export_pretrained_compression_model(
    pretrained_encodec='/path/to/custom_encodec/checkpoint_exported.th',
    out_file='/path/to/exports/my_model/compression_state_dict.bin'
)

The implementation:

def export_pretrained_compression_model(pretrained_encodec, out_file):
    if Path(pretrained_encodec).exists():
        pkg = torch.load(pretrained_encodec)
        assert 'best_state' in pkg
        assert 'xp.cfg' in pkg
        assert 'version' in pkg
        assert 'exported' in pkg
    else:
        pkg = {
            'pretrained': pretrained_encodec,
            'exported': True,
            'version': __version__,
        }
    Path(out_file).parent.mkdir(exist_ok=True, parents=True)
    torch.save(pkg, out_file)

Output Format

The output depends on the input type:

For pretrained model references:

Key	Type	Description
`pretrained`	`str`	The model identifier string (e.g., `'facebook/encodec_32khz'`)
`exported`	`bool`	Always `True`
`version`	`str`	AudioCraft version string

For custom-trained models:

Key	Type	Description
`best_state`	`dict`	Full model state dictionary
`xp.cfg`	`str`	YAML-serialized model configuration
`version`	`str`	AudioCraft version string
`exported`	`bool`	Always `True`

Loading Behavior

When the exported compression model is loaded by load_compression_model() in audiocraft/models/loaders.py, the presence of the pretrained key triggers a call to CompressionModel.get_pretrained(), which fetches the actual model weights from HuggingFace Hub. This deferred loading strategy keeps the export file small while ensuring the correct model is used at inference time.

Related Implementations

Page Connections

Double-click a node to navigate. Hold to expand connections.

Principle

Implementation

Heuristic

Environment