MQGAN/README.md

UNDER CONSTRUCTION

This repo is messy and not quite done yet. It is mostly Julesslop

Isolated PreEncoder Model

This directory contains an isolated version of the PreEncoder model and its necessary dependencies. It is designed to be a self-contained Python package.

Structure

• preencoder.py: Contains the main PreEncoder class, ConvBlock2D, and sequence_mask.
• attentions.py: Contains ResidualBlock1D and its dependencies (CausalConv1da, APTx, TransposeLayerNorm, CBAM1D, etc.).
• quantizer.py: Contains the FSQ (Finite Scalar Quantizer) class and its helpers.
• discriminators.py: Contains MelSpectrogramPatchDiscriminator2D, MultiBinDiscriminator, and their helpers.
• losses.py: Contains LSGANLoss.
• feature_extractors.py: Contains ISTFTNetFE.
• stft.py: Contains TorchSTFT.
• requirements.txt: Lists the external Python package dependencies.
• __init__.py: Makes this directory usable as a Python package and exports PreEncoder, MelSpectrogramPatchDiscriminator2D, MultiBinDiscriminator, LSGANLoss, ISTFTNetFE, and TorchSTFT.

Usage

You should be able to import the classes as follows, assuming pre_encoder_isolated is in your Python path:

from pre_encoder_isolated import PreEncoder, MelSpectrogramPatchDiscriminator2D, MultiBinDiscriminator, LSGANLoss, ISTFTNetFE, TorchSTFT

# Example instantiation (replace with actual parameters)
# pre_encoder = PreEncoder(mel_channels=80, channels=[...], kernel_sizes=[...])
# disc_2d = MelSpectrogramPatchDiscriminator2D(mel_channels=80)
# disc_multi_bin = MultiBinDiscriminator(mel_channels=80)
# lsgan_loss = LSGANLoss()
# torch_stft = TorchSTFT(filter_length=1024, hop_length=256, win_length=1024)
# istft_fe = ISTFTNetFE(gen=None, stft=torch_stft) # Replace None with actual generator model