Pytorch Sound is a modeling toolkit that allows engineers to train custom models for sound related tasks. It focuses on removing repetitive patterns that builds deep learning pipelines to boost speed of related experiments.

• Register models and call it other side.
  • It is inspired by https://github.com/pytorch/fairseq

import torch.nn as nn
from pytorch_sound.models import register_model, register_model_architecture


@register_model('my_model')
class Model(nn.Module):
...


@register_model_architecture('my_model', 'my_model_base')
def my_model_base():
    return {'hidden_dim': 256}

from pytorch_sound.models import build_model


# build model
model_name = 'my_model_base'
model = build_model(model_name)

• Several dataset sources (preprocess, meta, general sound dataset)

LibriTTS, Maestro, VCTK and VoiceBank are prepared at now.

Freely suggest me a dataset or PR is welcome!

• Abstract Training Process
  • Build forward function (from data to loss, meta)
  • Provide various logging type
    • Tensorboard, Console
    • scalar, plot, image, audio

import torch
from pytorch_sound.trainer import Trainer, LogType


class MyTrainer(Trainer):

    def forward(self, input: torch.tensor, target: torch.tensor, is_logging: bool):
        # forward model
        out = self.model(input)

        # calc your own loss
        loss = calc_loss(out, target)

        # build meta for logging
        meta = {
            'loss': (loss.item(), LogType.SCALAR),
            'out': (out[0], LogType.PLOT)
        }
        return loss, meta

• English handler sources are brought from https://github.com/keithito/tacotron

  • Add types

• General sound settings and sources

• ffmpeg v4

$ sudo add-apt-repository ppa:jonathonf/ffmpeg-4
$ sudo apt update
$ sudo apt install ffmpeg
$ ffmpeg -version

• install package

$ pip install -e .

1. Download data files

• In the LibriTTS case, checkout READMD

2. Run commands (If you want to change sound settings, Change settings.py)

$ python pytorch_sound/scripts/preprocess.py [libri_tts / vctk / voice_bank] in_dir out_dir

3. Checkout preprocessed data, meta files.

• Maestro dataset is not required running preprocess code at now.

• Source (Speech) Separation with audioset : https://github.com/AppleHolic/source_separation

• Python > 3.6
• pytorch 1.0
• ubuntu 16.04

1. Data and its meta file
2. Data Preprocess
3. General functions and modules in sound tasks
4. Abstract training process

• Preprocess docs in README.md
• Add test codes and CI
• Document website.

• This repository is under BSD-2 clause license. Check out the LICENSE file.