This is a repository to investigate how to generate music based in Hyperbolic Embeddings
Euclidean Space is a generalization of the space we live in, which has zero curvature. Lately, some investigations have show that, negative curvature space, a.k.a. Hyperbolic Space, has some properties that can be exploited for hierarchical data in deep learning.
On this repository we explore and show where Hyperbolic layers can be usefull to improve some deep learning task. Here are two folders:
- TorchExperiments: This folder contains the experiments and models done with PyTorch
- HyperbolicTImbrenet: This folders contain the timbrenet model of Agustin Macaya, and our experiments applying hyperobolic layers to the model.
PhD's advisors: Denis Parra, Mircea Petrache, Rodrigo Cadiz
TODO lista:
- Adapt the wrapped hyperbolic layers to the timbrenet model
- Train the model with the hyperbolic layers
- Try to incorporate the hyperbolic cnn incorporating Hyperbolic CNN and this Hyperbolic VAE