Exploring Automatic Music Generation using Transformer encoder-based Language Models

You can find the full paper of the thesis here

DISCLAIMER: The codebase relies on the MusicBERT open-source model.

Installation

git clone https://github.com/aspil/bsc-thesis.git
cd bsc-thesis
./setup.sh

1. Environment

Python:        3.8
fairseq:       git+https://github.com/pytorch/fairseq@336942734c85791a90baa373c212d27e7c722662#egg=fairseq

The original MusicBERT checkpoints seem to work only with the version above.

2. Dataset

2.1 Preparing dataset

• The dataset used in the paper can be found in data/raw directory, but you can use any dataset.
• Run the dataset processing script. (preprocess.py)
  python -u src/processing/preprocess.py
• The script should prompt you to input the path of the midi zip and the path for the preprocessed output.

  Dataset zip path: /data/raw/GiantMIDI-Baroque.zip
  OctupleMIDI output path: intermediate
  SUCCESS: test_midi.mid
  

• Binarize the raw text format dataset. (this script will read lmd_full_data_raw folder and output lmd_full_data_bin) bash binarize_pretrain.sh GiantMIDI-Baroque
  bash scripts/binarize_pretrain.sh data/baroque

3. Train / Fine-tune

Both training and fine-tuning are done using the Masked Language Modelling task. Therefore, for both scenarios use the following command:

bash scripts/train_mask.sh baroque base

In the current set, the data consists of Baroque musical pieces. The original pre-trained checkpoints can be downloaded from here. Create a checkpoints directory and place them there.

4. Generate

Use the following command to generate a musical piece:

python generate.py [-h] --save_path SAVE_PATH [--data DATA] [--checkpoint CHECKPOINT] [--config CONFIG] [--sampling_method {seq,gibbs,fill_mask}] [--prompt PROMPT] [--n_tokens N_TOKENS] [--topk TOPK] [--topp [0.0-1.0]] [--temperature [0.0-1.0]] [--max_steps MAX_STEPS]

The config parameter must be a YAML file, and if given any, all other arguments are ignored.