aakropotkin / artist-lyric-gen

A research project in conditional lyric generation conditioned on artists

artist-lyric-gen

A research project in conditional lyric generation conditioned on artists

Prerequisites

This project uses conda to manage packages and requirements - however pip can also be used.

• conda

• pip

See setup.sh to get started and install requirements for the project

Dataset

Important files relating to the dataset (train splits, artist list, token lists, etc) can be found in src/dataset

The dataset can be recreated from following the notebook src/dataset/dataset.ipynb

BPE

We forked Sennrich's BPE implementation here and BPE can be run using the notebook

Training

For CVAE, we have a training notebook which demonstrates the commands needed to train the model and generate verses

For BART, we have a training repo artist-lyric-gen-bart

Evaluation

For BART, generation and perplexity (things needing the BART model) can be done following our notebook src/Generation + Perplexity ... .ipynb

All other metrics using the generated verses can be done following src/artist_gen_eval.ipynb