Exploring XAI for the Arts: Explaining Latent Space in Generative Music

Augmentation of Ashis Pati and Alexander Lerch's "Latent Space Regularization for Explicit Control of Musical Attributes" (2019) and "Attribute-based Regularization of Latent Spaces for Variational Auto-Encoders" (2020).

Implementation source: https://github.com/ashispati/AttributeModelling.

Web Demo Links:

https://xai-lsr-ui.vercel.app/

https://xai-no-lsr-ui.vercel.app/

Python and Cuda Versions:

python/3.7.7
cuda/10.2-cudnn8.0.5

For the python packages:

pip install -r requirements.txt

Dataset:

Download from: https://github.com/ashispati/AttributeModelling

Unzip the downloaded file and put the datasets and folk_raw_data folders under data.

Generated Pianoroll and Audio Files with LSR for Demonstration:

They are named with their corresponding musical metric levels (10 discrete levels for each of the 4 metrics). For example, midi_3_4_5_3.mid means, this file has 3/10 rhythmic complexity, 4/10 note range, 5/10 note density and 3/10 average interval jump.