My Expedition to Machine-Generated Music Using Tensorflow Magenta Part 2: Linkin Park Continued

1 minute read

Repository

https://github.com/tensorflow/magenta

Introduction

If you have not read the first post of the series, please read it before reading this post.

I have tried using two NVIDIA Tesla V100 16GB SXM2 GPUs and the MusicVAE model (https://magenta.tensorflow.org/music-vae) to generate music based on MIDI files of songs from the albums of Linkin Park.

The first trained model showed that a computer could generate harmonic (or should I say normal) music, but the sample size was too small to generate original music. (Quiz: what was the most predominant Hybrid Theory/Meteora song on the first samples?)

Because of this, I have gone further and expanded the sample size of Linkin Park songs to 42 songs from all seven Linkin Park albums.

Post Body

The methodology did not differ much from the first post, so please consult the first post for how you can do this yourself.

This time, it also took me around a week on one Tesla V100 16GB SXM2 GPU using --hparams=batch_size=32,learning_rate=0.0005, same as the first run. Ten samples were generated this time.

This time, the music samples were quite original while it did remind me of what Linkin Park would have composed.

The accuracy of the model did not increase fast as the first model, and it increased gradually until the 200k steps finished.

Apart from that, the result was quite impressing.

image.png image.png image.png image.png image.png image.png image.png image.png image.png image.png

The results are available Here.

image.png I can feel them while listening to these samples. Don’t you? (Image courtesy of Linkin Park)

If you want to generate new music using this trained model, Download the Checkpoint from Here and use Tensorflow Magenta to generate using This Command. You can also interpolate with other Midi files to add a taste of Linkin Park to other music.

Next time, I will be back with some different artists, or combinations of different artists using the MusicVAE model of Tensorflow Magenta.

I will also perform a detailed analysis of models trained from now on using Tensorboard.

Resources

MusicVAE model (https://magenta.tensorflow.org/music-vae) in Tensorflow Magenta (https://github.com/tensorflow/magenta)
FreeMidi.org (https://freemidi.org/)
FLAC Synthesizer by Musescore (https://musescore.com/ or https://github.com/musescore/MuseScore)

This is the second post of a series of blog posts on Machine-Generated Music. The collection of posts will expand based on new selections on genres and music artists or bands.

First Post

Comments