This project intends to implement a general purpose main cycle for the training, generation, validation and testing of generative models, but applied in this case to the symbolic musical generation task.

Models vs Modeloids:

• a modeloid has a fixed static architecture, a state and it is related with some specific datasets in a specific representation and it is a concretization of a model
• a model is a generalization of modeloids (usually inclues in its definition a way to create modeloids given some datasets)

In a file system a modeloid is a directory that has a mconfig.py and all modeloid's auxiliar files.

In this project I do not have a modeloid class but every class implementing a modeloid must include:

• id: a unique id string

• name: name do modeloide

• authors: the authors of the modeloid

• doc: a description of the modeloid

• model: modeloid specific parameter (tensorflow, learning rates, architecture ...)

• load_model: a function that is responsible for creating the python object that implements the modeloid with the right interface and in the state it should be

• train()

• loss()

• sample()

• save()

Just for testing all the rest of the components of the project

It is one RBM based model with two layers

In 2017 musegan used GAN's do generate symbolic music, https://salu133445.github.io/musegan/

GAN's have been used to generate images since it's introduction in 2014.

Some models such as CycleGAN (2017) and DiscoGAN (2017) have been proposed for style transfer tasks in visual field, where GAN's are improved with one cycle consistency or reconstruction loss.

To the best of our knowledge these models have not been applied to Symbolic Music Data.

In 2017, one usage of these models have been used in audio musical data represented as spectograms, http://gauthamzz.com/2017/09/23/AudioStyleTransfer/

This model presents an approach of symbolic music generation combining:

• DCGAN
• cycle consistency loss
• LSTMs

Aiming to improve the state of the art on symbolic music neural generative models in the following aspects:

• Music Dynamics
• Temporal relations modelling
• Instrumentation
• Introducing a mechanism of inspiration [Not style transfer!!]

Future Work

Report module allows to save all relevant data collected from the main cycle in the desired formats.

This project may be framed within the area of Computational Creativity. So I leave some topics that may:

• Randomness, overfitting and learning power
• The role of explainability on creativity process
• Defining what is dificult to define
• Meta operations
• Emerging creative numeric artists in network: what is the role of that society has in creativity evaluation in real cases and what should it be in CC cases