README.md

---
license: mit
language:
- en
tags:
- music
- autoencoder
- variational autoencoder
- music generation
---

# Pivaenist

Pivaenist is a random piano music generator with a VAE architecture.

By the use of the aforementioned autoencoder, it allows the user to encode piano music pieces and to generate new ones.

## Model Details

### Model Description

<figure>
<img src="https://huggingface.co/TomRB22/pivaenist/resolve/main/.images/architecture.png" style="width:100%; display:block; margin:auto">
<figcaption align = "center"><b>Pivaenist's architecture.</b></figcaption>
</figure>

- **Developed by:** TomRB22
- **Model type:** Variational autoencoder
- **License:** MIT

### Sources

**Code:** Some of the code of this repository includes modifications (not the entire code, due to the differences in the architecture) or implementations from the following sites:
1. [TensorFlow. (n.d.). Generate music with an RNN | TensorFlow Core](https://www.tensorflow.org/tutorials/audio/music_generation) - Tensorflow tutorial where pretty-midi is used
2. [Han, X. (2020, September 1). VAE with TensorFlow: 6 Ways](https://towardsdatascience.com/vae-with-tensorflow-6-ways-9c689cb76829) - VAE explanation and code
3. [Li, C. (2019, April 15). Less pain, more gain: A simple method for VAE training with less of that KL-vanishing agony. Microsoft Research.](https://www.microsoft.com/en-us/research/blog/less-pain-more-gain-a-simple-method-for-vae-training-with-less-of-that-kl-vanishing-agony/)  - Microsoft article on the KL training schedule which was applied in this model

There might be acknowledgments missing. If you find some other resemblance to a site's code, please notify me and I will make sure of including it.


### Using pivaenist in colab

If you preferred directly using or testing the model without the need to install it, you can use the following colab notebook and follow its instructions. Moreover, this serves as an example of use.
[colab link]

## Installation

To install the model, you will need to **change your working directory to the desired installation location** and execute the following commands:

**_Windows_**
```console
git clone https://huggingface.co/TomRB22/pivaenist
sudo apt install -y fluidsynth
pip install -r ./pivaenist/requirements.txt
```

**_Mac_**
```console
git clone https://huggingface.co/TomRB22/pivaenist
brew install fluidsynth
pip install -r ./pivaenist/requirements.txt
```

The first one will clone the repository. Then, fluidsynth, a real-time MIDI synthesizer, is also set up in order to be used by the pretty-midi library. With the last line, you will make sure to have all dependencies on your system.

[More Information Needed]

## Training Details

Pivaenist was trained on the [MAESTRO v2.0.0 dataset](https://magenta.tensorflow.org/datasets/maestro), which contains 1282 midi files [check it in colab]. Their preprocessing involves splitting each note in pitch, duration and step, which compose a column of a 3xN matrix (which we call song map), where N is the number of notes and a row represents sequentially the different pitches, durations and steps. The VAE's objective is to reconstruct these matrices, making it then possible to generate random maps by sampling from the distribution, and then convert them to a MIDI file.

<figure>
    <img src="https://huggingface.co/TomRB22/pivaenist/resolve/main/.images/map_example.png" style="width:30%; display:block; margin:auto">
    <figcaption align = "center"><b>A horizontally cropped example of a song map.</b></figcaption>
</figure>

# Documentation

## **_Audio_**

### midi_to_notes

```python
def midi_to_notes(midi_file: str) -> pd.DataFrame:
```
Convert midi file to "song map" (dataframe where each note is broken
into its components)

Parameters:
* midi_file (str): Path to the midi file.

Returns:
* pd.Dataframe: 3xN matrix where each column is a note, composed of pitch, duration and step.

### display_audio

```python
def display_audio(pm: pretty_midi.PrettyMIDI, seconds=-1) -> display.Audio:
```
Display a song in PrettyMIDI format as a display.Audio object. This method is especially useful in a Jupyter notebook.

Parameters
* pm (pretty_midi.PrettyMIDI): PrettyMIDI object containing a song.
* seconds (int): Time fraction of the song to be displayed. When set to -1, the full length is taken.

Returns
* display.Audio: Song as an object allowing for display.