Implementation:Facebookresearch Audiocraft BaseQuantizer

Knowledge Sources	Facebookresearch_Audiocraft
Domains	Quantization, Audio_Compression
Last Updated	2026-02-14 01:00 GMT

Overview

Concrete tool for defining the abstract quantizer interface and data structures used by all vector quantization implementations in AudioCraft.

Description

This module defines QuantizedResult (a dataclass holding quantized tensor, codes, bandwidth, and penalty), BaseQuantizer (the abstract base class with encode/decode/forward interface), and DummyQuantizer (a pass-through implementation for testing). All concrete quantizers (RVQ, etc.) implement this interface.

Usage

Import these base classes when building or extending quantization modules. DummyQuantizer is useful for testing compression models without actual quantization.

Code Reference

Source Location

• Repository: Facebookresearch_Audiocraft
• File: audiocraft/quantization/base.py
• Lines: 1-99

Signature

@dataclass
class QuantizedResult:
    x: torch.Tensor
    codes: torch.Tensor
    bandwidth: torch.Tensor
    penalty: tp.Optional[torch.Tensor] = None

class BaseQuantizer(nn.Module):
    @abstractmethod
    def forward(self, x: torch.Tensor, frame_rate: int) -> QuantizedResult: ...
    @abstractmethod
    def encode(self, x: torch.Tensor) -> torch.Tensor: ...
    @abstractmethod
    def decode(self, codes: torch.Tensor) -> torch.Tensor: ...

class DummyQuantizer(BaseQuantizer): ...

Import

from audiocraft.quantization.base import BaseQuantizer, QuantizedResult, DummyQuantizer

I/O Contract

Inputs

Name	Type	Required	Description
x	torch.Tensor	Yes	Continuous latent tensor [B, D, T]
frame_rate	int	Yes	Frame rate for bandwidth computation

Outputs

Name	Type	Description
result	QuantizedResult	Quantized output with codes and bandwidth

Related Pages

• Principle:Facebookresearch_Audiocraft_Encoder_Decoder_Architecture