Project for DD2424 Deep Learning in Data Science

Authors: Linus Östlund and Marco Amoretti

Project description

Something on the dataset (cite their work)
Something on the architecture
Something on Molly

How to use

Create Virtual Environment

Using either conda or pip, create a virtual environment. Activate it and then call:

pip install -r requirements.txt

Generate data

To generate training data from dataset.pkl use:

python3 src/data_processing/preprocessing.py
python3 src/data_processing/create_data.py

This will convert the pickled dataset to midi files, the midi files to abc notation and then perform a cleaning step that will:

remove files with low-occurence tokens (less than 100 occurrences in the whole dataset);
reformat all training data in one file (one line per piece, with space-separated tokens).

Train the model (LSTM)

To train a model call:

python3 src/train.py

The command line arguments provide costumization for model size and layers, number of epochs, patience for early stopping and so on. A checkpoint for each epoch will be saved in ./ckpts/.

The flag -ld {CKPT} loads a previous checkpoint and continues training from there. The size of the loaded model has to match the provided command line arguments (e.g. if the checkpoint has 3 layers add -l 3 and similar).

Inference the model

To generate from a checkpoint use:

python3 src/train.py -i -ld {CKPT} -m {'greedy', 'topk', 'topp'} -t {TEMPERATURE}

The model can be sampled with greedy, Top-P and Top-K sampling. Generation starts from an empty sequence starting with <sos> and continues until either the maximum sequence length is reached or a <eos> token is generated.

Output is printed to the console. To listen to the output:

create an ABC-file with the following fields:

X:0
M:4/4
L:1/8
Q:120
K:C

copy all tokens between <sos> and <eos> and paste them after K:C;
invoke abc2midi <abc_file> -o <midi_file>;
listen in your program of choice or with timidity <midi_file>