Implementation:Facebookresearch Audiocraft MultiScaleDiscriminator

Knowledge Sources	Facebookresearch_Audiocraft MelGAN
Domains	Audio_Synthesis, GAN
Last Updated	2026-02-14 01:00 GMT

Overview

Concrete tool for evaluating audio waveforms at multiple temporal resolutions using average pooling and 1D convolutions for adversarial training.

Description

MultiScaleDiscriminator implements the MSD architecture that evaluates audio at multiple temporal scales. It applies average pooling with increasing factors to create downsampled views of the audio, then runs 1D convolutional sub-discriminators on each scale. This captures both fine-grained and coarse temporal patterns.

Usage

Import this class when building adversarial training pipelines for audio compression or generation models that need a multi-scale temporal discriminator.

Code Reference

Source Location

• Repository: Facebookresearch_Audiocraft
• File: audiocraft/adversarial/discriminators/msd.py
• Lines: 1-126

Signature

class MultiScaleDiscriminator(MultiDiscriminator):
    def __init__(self, in_channels: int = 1, out_channels: int = 1,
                 n_layers: int = 4, downsample_factor: int = 2,
                 n_scales: int = 3):
        """
        Args:
            in_channels: Number of input audio channels.
            out_channels: Output channels per sub-discriminator.
            n_layers: Number of convolution layers per scale.
            downsample_factor: Pooling factor between scales.
            n_scales: Number of temporal scales.
        """

Import

from audiocraft.adversarial.discriminators.msd import MultiScaleDiscriminator

I/O Contract

Inputs

Name	Type	Required	Description
x	torch.Tensor	Yes	Audio waveform [B, C, T]

Outputs

Name	Type	Description
logits	list[torch.Tensor]	Discriminator logits from each scale
features	list[list[torch.Tensor]]	Intermediate features for feature matching loss

Usage Examples

from audiocraft.adversarial.discriminators.msd import MultiScaleDiscriminator
import torch

msd = MultiScaleDiscriminator(in_channels=1, n_scales=3)
wav = torch.randn(4, 1, 24000)
logits, features = msd(wav)

Related Pages

• Principle:Facebookresearch_Audiocraft_Multi_Scale_Discrimination