DeepToneNets

Repository for DeepToneNets, a deep learning model for functional harmony recognition directly on audio.

Requirements

Make sure to install the required dependencies specified in the environment.yml file, or set up a suitable Python environment using conda or pip.

Installation

Clone this repository:

git clone https://github.com/ThiagoPoppe/deeptone-nets.git
cd deeptone-nets

Create and activate the conda environment (anaconda required):
```
conda env create -f environment.yml
conda activate dtns
```

Defining data constants

Under source/data/constants.py you can change AUDIOS_DATAPATH and PROCESSED_HOOKTHEORY_DATAPATH values to reflect the datapath in which your audios are stored (needs to be in H5PY format) and the processed HookTheory JSON file.

Model pre-trained weights

The model weights with results reported in the paper can be found at pretrained/ folder.

# Loading example using PyTorch
model = HarmonyBassChromaNetworks()
model.load_state_dict(torch.load('pretrained/harmonybass_chroma_networks_weights.pth'))

Dataset Creation

In order to train your own models, you need to create a dataset for each specific task with the create_harmonybass_dataset.py and create_functional_harmony_dataset.py scripts. Next, we give a usage example for each one and their command line arguments.

HarmonyBass Chroma Network (HBCN)

# Usage example
python create_harmonybass_dataset.py --artist <ARTIST> [--no-split]

Parameter	Type	Requirement	Description
`--artist`	`str`	Required	The artist to create the dataset for (e.g. `top100`, `classical`, `the-beatles`)
`--no-split`	`flag`	Optional	Flag to avoid splitting the dataset into train/valid/test.

DeepToneNets (DTNs)

# Usage example
python create_functional_harmony_dataset.py --artist <ARTIST> [--no-split]

Parameter	Type	Requirement	Description
`--artist`	`str`	Required	The artist to create the dataset for (e.g. `top100`, `classical`, `the-beatles`)
`--hbcn-checkpoint`	`str`	Required	Path to the HBCN checkpoint file with the weights
`--no-split`	`flag`	Optional	Flag to avoid splitting the dataset into train/valid/test.

Training Models

HarmonyBass Chroma Network (HBCN)

You can re-train the HarmonyBass Chroma Network (HBCN) using the train_hbcn.py script. This script only requires the dataset name as an argument.

# Usage example
python train_hbcn.py --dataset <DATASET_NAME>

Parameter	Type	Requirement	Description
`--dataset`	`str`	Required	Name of the dataset to train the model on (must exist within datasets folder)

DeepToneNets (DTNs)

You can also re-train the DeepToneNets (DTNS) model using the train_dtns.py script. This script allows you to specify the dataset and variations of the DTNs for training.

# Usage example
python train_dtns.py --dataset <DATASET_NAME> [--use-harmony-only-chroma] [--no-keysim-bias]

Parameter	Type	Requirement	Description
`--dataset`	`str`	Required	Name of the dataset to train the model on
`--use-harmony-only-chroma`	`flag`	Optional	Flag to use only harmony-related chroma features
`--no-keysim-bias`	`flag`	Optional	Disable key similarity bias during training