Cadet

Cadence Detection in Symbolic Classical Music using Graph Neural Networks (ISMIR2022).

Introduction

This repository contains the training and the models from the paper Cadence Detection in Symbolic Classical Music using Graph Neural Networks submitted at ISMIR 2022. pdf

Requirements and Installation

Installation

If you are using conda please install the environment:

conda env create -f environment.yml
conda activate cadet
cd path/to/cadet
pip install .

If this fails, follow the steps bellow.

Pre-installation requirements:

conda create -n cadet python=3.8 pip
conda activate cadet

Proceed by visiting the following websites and installing the appropriate version of the following packages.

Pytorch >=1.8.1 link;
DGL >= 0.7 link.

After the installation of Pytorch and DGL with the platform of your choice (pip and conda supported on this repo), you can install the rest of the requirements using the following commands:

To install other requirements using pip:

pip install -r requirements.txt

To install other requirement using conda:

conda env update -f environment.yml

You might also need to install the repo as a package if you are running from the terminal. It's suggested to use the experimental pip install to keep up with new versions or edits you might want to make:

pip install . -e

Getting Started

To run the pre-trained models you will need a wandb account. You can find the project results and download the models at https://wandb.ai/melkisedeath/Cadence Detection.

Train a model

cd cadet/train
python train_lightning.py --dataset wtc --cad-type pac

Using the above command you will train PAC detection on the Bach fugues of the 1st Welle tempered clavier book.

If you wish not to log your run with WANBD on the cloud then run:

WANDB_MODE=offline python script.py --args

Load and train pre-trained model

cd cadet/train
python train_lightning --dataset wtc --cad-type pac --wandb_id --load_from_checkpoints

Using the above command you will load a pretrained model for PAC detection, pretrained on String Quartets and fine tune it on the Bach fugues of the 1st Well tempered clavier book.

Reproduce results from pre-trained model

Load pretrained model from Bach Fugues PAC prediction and skip training only to reproduce results on the test set.

cd cadet/train
python train_lightning --dataset wtc --cad-type pac --wandb_id  --skip-training --load_from_checkpoints

Create Graph from Score

To create the graphs from the score you will need to provide a directory of scores.

cd cadet/utils
python create_homo_graph_dataset.py --data_dir

External Repositories

An external repository is provided with the dataset which can also be found as a git sub-module

You can browse the latest version of the repository here.

Feature Extraction

The feature extraction includes three categories of featrures:

The General Note features
The graph topology features
The Cadence Relevant features

The topology features are produced by taking the $N$ first eigenvectors of the Laplacian of the Adjacency matrix. The script that produces this values can be find in cad/utils/pos_enc.py. The complete list of the other features is presented below.

List of Features

The features are computed per note/rest in the score. Chord refers to the set of notes that have the same onset value as the current note.

Function Name	Type	Description
General Note Features
onset_feature.score_position	float	normalized onset between 0 and 1
duration_feature.duration	float	duration of notes in formalized value
fermata_feature.fermata	binary	If note has fermata
grace_feature.n_grace	float	How many grace notes on this onset position
grace_feature.grace_pos	float	Which grace note in the sequence of grace notes
onset_feature.onset	float	normalized onset feature
polynomial_pitch_feature.pitch	float	normalized midi pitch between 0 and 1
grace_feature.grace_note	binary	is grace note
relative_score_position_feature.score_position	float
slur_feature.slur_incr	float
slur_feature.slur_decr	float
time_signature_feature.time_signature_num_1	float
time_signature_feature.time_signature_num_2	float
time_signature_feature.time_signature_num_3	float
time_signature_feature.time_signature_num_4	float
time_signature_feature.time_signature_num_5	float
time_signature_feature.time_signature_num_6	float
time_signature_feature.time_signature_num_7	float
time_signature_feature.time_signature_num_8	float
time_signature_feature.time_signature_num_9	float
time_signature_feature.time_signature_num_10	float
time_signature_feature.time_signature_num_11	float
time_signature_feature.time_signature_num_12	float
time_signature_feature.time_signature_num_other	float
time_signature_feature.time_signature_den_1	float
time_signature_feature.time_signature_den_2	float
time_signature_feature.time_signature_den_4	float
time_signature_feature.time_signature_den_8	float
time_signature_feature.time_signature_den_16	float
time_signature_feature.time_signature_den_other	float
vertical_neighbor_feature.n_total	float
vertical_neighbor_feature.n_above	float
vertical_neighbor_feature.n_below	float
vertical_neighbor_feature.highest_pitch	float
vertical_neighbor_feature.lowest_pitch	float
vertical_neighbor_feature.pitch_range	float
int_vec1	int	first value of interval vector computed for all notes on the onset of the current one.
int_vec2	int	second value of interval vector computed for all notes on the onset of the current one.
int_vec3	int	third value of interval vector computed for all notes on the onset of the current one.
int_vec4	int	fourth value of interval vector computed for all notes on the onset of the current one.
int_vec5	int	fifth value of interval vector computed for all notes on the onset of the current one.
int_vec6	int	sixth value of interval vector computed for all notes on the onset of the current one.
M/m	binary	is the interval vector equivalent to Major or Minor chord
sus4	binary	is the interval vector equivalent to a sus4 chord
M7	binary	is the interval vector equivalent to a dominant 7 chord
M7wo5	binary	is the interval vector equivalent to a dominant 7 chord without the 5th
Mmaj7	binary	is the interval vector equivalent to a major 7 chord
Mmaj7maj9	binary	is the interval vector equivalent to a major 7 chord with major 9
M9	binary	is the interval vector equivalent to a dominant chord with 9
M9wo5	binary	is the interval vector equivalent to a dominant chord with 9 without the 5th
m7	binary	is the interval vector equivalent to a minor 7 chord
m7wo5	binary	is the interval vector equivalent to a minor 7 chord without the 5th
m9	binary	is the interval vector equivalent to a minor chord with a major 9
m9wo5	binary	is the interval vector equivalent to a minor chord with a major 9 without the 5
m9wo7	binary	is the interval vector equivalent to a minor chord with a major 9 without the 7
mmaj7	binary	is the interval vector equivalent to a minor chord with a major 7
Maug	binary	is the interval vector equivalent to a augmented chord
Maug7	binary	is the interval vector equivalent to a augmented chord with 7
mdim	binary	is the interval vector equivalent to a diminshed chord
mdim7	binary	is the interval vector equivalent to a diminshed chord with 7
is_maj_triad	binary	is the set of notes present on the onset of the current note a major triad
is_pmaj_triad	binary	is the set of notes present on the onset of the current note a perfect major triad
is_min_triad	binary	is the set of notes present on the onset of the current note a minor triad
ped_note	binary	is the current note a pedal note
hv_7	binary	is the highest voice of the chord a 7 of the chord compared to the lowest pitch
hv_5	binary	is the highest voice of the chord a 5 of the chord compared to the lowest pitch
hv_3	binary	is the highest voice of the chord a 3 of the chord compared to the lowest pitch
hv_1	binary	is the highest voice of the chord an octave of the chord compared to the lowest pitch
chord_has_2m	binary	does the chord have a 2m
chord_has_2M	binary	does the chord have a 2M
Cadence Features
perfect_triad	binary	is the chord a perfect triad
perfect_major_triad	binary	is the chord a perfect major
is_sus4	binary	is the chord a sus4
in_perfect_triad_or_sus4	binary	is the chord a perfect triad or a sus4
highest_is_3	binary	is the highest voice a 3rd compared to the lowest on this onset
highest_is_1	binary	is the highest voice a 8th compared to the lowest on this onset
bass_compatible_with_I	binary	is the bass compatible with the Tonal of the scale
bass_compatible_with_I_scale	binary
one_comes_from_7	binary	does the one (compared to the lowest voice) comes from a leading tone (compared to previous onset)
one_comes_from_1	binary	was the one (compared to the lowest voice) present on the previous onset
one_comes_from_2	binary	does the one (compared to the lowest voice) comes from a second (compared to previous onset)
three_comes_from_4	binary	does the third (compared to the lowest voice) comes from a fourth (compared to previous onset)
five_comes_from_5	binary	was the fifth (compared to the lowest voice) present on the previous onset
strong_beat	binary	is the current onset a strong beat
sustained_note	binary	is the current note a pedal note (more or equal to a bar's duration)
rest_highest	binary	is there a rest on the highest voice
rest_lowest	binary	is there a rest on any middle voice
rest_middle	binary	is there a rest on the lowest voice
voice_ends	binary	is there a voice end after the particular onset
v7	binary	does the chord have a seventh
v7-3	binary	does the chord have a seventh and a third
has_7	binary	does the chord have a seventh
has_9	binary	does the chord have a ninth
bass_voice	binary	is the bass on this voice/note
bass_moves_chromatic	binary	does the bass (lowest pitch) move chromatically
bass_moves_octave	binary	does the bass (lowest pitch) move with an octave jump
bass_compatible_v-i	binary	does the bass (lowest pitch) move with a V to I
bass_compatible_i-v	binary	does the bass (lowest pitch) move with a I to V
bass_moves_2M	binary	does the bass (lowest pitch) does the mass move with a second major interval.

Crystal-LiuBojia / cadet