MPI_SAGAN - MIDI Piano Image Generative Adversarial Network using Self-Attention

---

This project is the generating music using self attention gan.

Out result is here.

Before training the model, we convert midi file to grayscale midi image.

piano roll midi file

gray scale midi image

training the gan without thresh hold / training the gan applying 0.3 thresh / training the gan applying 0.5 thresh

Models arthitecture

---

Data source

---

GTZAN Dataset - Music Genre Classification
Link : https://www.kaggle.com/andradaolteanu/gtzan-dataset-music-genre-classification
Classical Music MIDI
Link : https://www.kaggle.com/soumikrakshit/classical-music-midi
Magenta Datasets
Link : https://magenta.tensorflow.org/datasets/
Music MIDI Collection
Link : https://www.kaggle.com/programgeek01/anime-music-midi

GAN Result

---

The DC gan model which trained 450 epoch generate fake music

youtube link: https://youtu.be/HZKP8TjhIk8

The DC gan model which trained 1000 epoch generate fake music

youtube link: https://youtu.be/apQXf4pcT2I

Self attention GAN Result

---

The self attention gan model which trained 150 epoch generate fake music

youtube link: https://youtu.be/QxlSpPadJnE

The self attention gan model which trained 190 epoch generate fake music

youtube link: https://youtu.be/5XvYgP_0rw0

The self attention gan model which trained 420 epoch generate fake music

youtube link: https://youtu.be/nfh7_8glCIM

The self attention gan model which trained 430 epoch generate fake music

youtube link: https://youtu.be/dNytCDL6Ss4

---