Skip to content

Latest commit

 

History

History
80 lines (61 loc) · 3.8 KB

README.md

File metadata and controls

80 lines (61 loc) · 3.8 KB
Raw

Billion Songs, AI-powered song lyrics generator

Generate a song

NOTE: this repository has git submodules. So clone with --recurse-submodules. Learn about them here.

See the blog post Writing billion songs with C# and Deep Learning for a detailed explanation how it works.

This project mainly serves as a demonstration of Gradient, our TensorFlow binding for C# and other .NET languages.

Other Gradient samples.

What is it, and how does it work?

This is a deep learning-powered song lyrics generator, based on GPT-2, wrapped as a ASP.NET Core website.

It generates songs word by word (or rather token by token), using the statistical relationships learned by a deep learning model, called GPT-2. The actual generator code is in GradientTextGenerator class.

Text generation is pretty slow even with a powerful GPU, so we have a bunch of caches in /Web to provide a better user experience. There is also PregeneratedSongProvider, which continuously creates new texts in the background to ensure clicking "Make Random" button gives an instant result.

Detailed explanation in a blog post

Prerequisites

  1. Download and install Python and TensorFlow as described in Gradient documentation
  2. Install Python package, called regex (python -m pip install regex --user)
  3. Install the latest .NET Core SDK

Run instructions

  1. Clone the repository and enter the Web folder
  2. .NET Core 3+ only: ensure you have Entity Framework tool installed: dotnet tool install --global dotnet-ef
  3. After cloning the repository, enter the Web folder and run dotnet ef database update. That should create songs.db file in the same directory.
  4. Edit appsettings.json (see appsettings.Development.json for an example):
    • add "DB": "sqlite"
    • modify DefaultConnection to "DefaultConnection": "Data Source=songs.db"
    • ensure that Generator is not dummy, if you want lyrics to actually be generated
  5. Run dotnet run web. This should print some logs. Wait for Now listening on: http://, then open that URL in the browser. It will take up to 4 minutes to generate the first song.

NOTE: if you see "Can't choose between the following Python environments, as they are equally matching", set PYTHON_CONDA_ENV_NAME to the name of Conda environment where you installed TensorFlow and regex modules.

Train instructions

NOTE: training requires a lot of RAM (>16GB), and will be slow on non-GPU

  1. Download the original 117M GPT-2 model by running one of download_model.* scripts in External/Gradient-Samples/GPT-2 from the same directory.
  2. Download and extract any lyrics dataset (I used Every song you have heard (almost)!), and unpack it if needed.
  3. From the command line in the same directory (GPT-2), run dotnet run train --include *.csv --column Lyrics path/to/lyrics/folder --run Lyrics (change the column parameter to the name of the lyrics column in you dataset)

NOTE: dev instance was trained with train -i "*.csv" --column=Lyrics Downloads\every-song-you-have-heard-almost -r Lyrics --checkpoint=fresh --save-every=100 -n 3. If training from IDE, set working directory to GPT-2 (which should contain models subfolder downloaded previously).

  1. Interrupt training process, when samples start looking good.
  2. Try the trained model by running dotnet run --run Lyrics