gianscarpe / chorals_hmm

HMM and FHMM for Bach Chorals dataset

Bach Chorales

Introduction

We implemented a fully functional framework to test temporal probabilistic model for prediction and sample of music. We compared HMM and FHMM, using various library and trying with different hyper-parameters.

Dependencies

Packages needed to play MIDI song:

• timidity: sudo apt install timidity
• fluid-soundfont-gm: sudo apt install fluid-soundfont-gm

To play a song run timidity -Os filename.mid

Packages needed to show scores:

• musescore: sudo apt install musescore

You can install instantiate a virtual environemnt of the project using conda env create --file=environment.yml

Usage

Parser

You can parse a new dataset using python parse.py. The script generates a new parsed dataset and the corresponding vocabulary.

Arguments:

--path: path to the dataset directory
--to-states: parse dataset to a list of state integers

HHM

You can train a FHMM model using python main_fhmm.py HMM arguments:

-K: dimension of states alphabet
-N: number of iterations

Dataset arguments:

--dataset-dir: path to dataset directory
--trainset-name: name of the dataset of training
--trainset-size: ['all', int_value] split train/test
--testset-name: name of the dataset of test

arguments for training process:

-F: framework: pom (pomegranate) or hmml (hmmlearn)
--tol: Set the convergence tolerance
-v: training verbose
-s: save the model
 

arguments for testing process:

--generate: generate a song using the current model 
--skip-training: skip the training process 
  (model-path is mandatory)
--model-path: path to a model to load

FHHM

You can train a FHMM model using python main_fhmm.py FHMM arguments:

-M: numeber of chains
-K: dimension of states alphabet
-N: number of iterations

Dataset arguments:

--dataset-dir: path to dataset directory
--trainset-name: name of the dataset of training
--trainset-size: ['all', int_value] split train/test
--testset-name: name of the dataset of test

arguments for training process:

-v: training verbose
-s: save the model
 

arguments for testing process:

--generate: generate a song using the current model 
--skip-training: skip the training process 
  (model-path is mandatory)
--model-path: path to a model to load

GUI

You can launch a GUI of the project using your favourite WSGI HTTP Server.

Gunicorn example:

gunicorn app:app.py

Report

You can find complete report of the project at relazione/Scarpellini_Belotti_Samotti_relazione.pdf (Italian Only)

LICENSE

Learning porpous only - no guarantee or assistance is provided. Please respect the license and cite us.

You can find us at:

@gianscarpe: gianluca[at]scarpellini.cloud
@belerico: f.belotti8[at]campus.unimib.it or belo.fede[at]outlook.com