Neural Loop Combiner: Neural Network Models For Assessing The Compatibility of Loops

This repository contains the code for "Neural Loop Combiner: Neural Network Models For Assessing The Compatibility of Loops" Proceedings of the 21st International Society for Music Information Retrieval Conference (ISMIR), 2020. Bo-Yu Chen, Jordan B. L. Smith, Yi-Hsuan Yang

If you want to hear more audio example, please check the demo page here
If you want to play it in interactive way, please check Beats with You [code]

Prerequisites

• python=3.7.8
• torch=1.7.0
• librosa=0.7.2
• download loopextractor script from here

Installing Required Libraries

git clone https://github.com/jblsmith/loopextractor.git
pip install -r requirements.txt

Run MongoDB

• Install mongo locally follow the article
• Run mongo to connect to your database, just to make sure it's working. Once you see a mongo prompt, exit with Control-D
• Set Database configuration in ./neural_loop_combiner/config/settings.py

Data Preprocessing

• Set Directory and Directory and Others configuration in ./neural_loop_combiner/config/settings.py
• Use data_preprocess.py file to preprocess input datas (INT_DIR in ./neural_loop_combiner/config/settings.py)

Data preprocessing consists of two main stages:

1. Load Tracks - Load tracks from inputs directory to database which is used to decompose
2. Data Generation - Decompose tracks to individual loops and layout (arrangement)

python data_preprocess.py [--load=(0, 1)] [--extract=(0, 1)] [--gpu_num=0]

• --load: whehter execute the load_tracks step (1 -> execute, 0 -> skip)
• --extract: whehter execute the data_generation step (1 -> execute, 0 -> skip)
• --gpu_num: specify which gpu should used to execute the code

Note that the second stage takes a fairly long time - more than an day.

Create Dataset

• Set Datasets configuration in ./neural_loop_combiner/config/settings.py
  • Set NG_TYPES to decide which negative sampling should include in the datasets
• Use create_dataseet.py file to create the train/val/test datas

Create Dataset consists of two main stages:

1. Loops Tag - Tag loops type (harmonic, percussion, bass), only use in selected negative sampling
2. Dataset Creation - Run negative sampling based on ./neural_loop_combiner/config/settings.py and create positive/negative data

python create_dataset.py [--tag=(0, 1)]

• --tag: whether execute the loops tag step or directly import from database (1 -> execute, 0 -> import)

Note that the first stage takes a fairly long time - more than an day.

Model Training

• Set Models configuration in ./neural_loop_combiner/config/settings.py
• Use train.py file to train the models

python train.py [--gpu_num=0] [--lr=0.01] [--epochs=20] [--batch_size=128] [--log_interval=10] [--neg_type=(random, selected, shift, rearrange, reverse)] [--model_type=(cnn/snn)]

• --gpu_num: specify which gpu should used to execute the code
• --lr: learning rate used to train the model
• --epochs: how many epochs should used to train the model
• --batch_size: batch size used to train the model
• --log_interval: how often should log the message
• --neg_type: specify the model should train in which negative sampling method (random, selected, shift, rearrange, reverse)
• --model_type: specify which kind of model you want to train (snn/cnn)

Contact

Please feel free to contact Bo-Yu Chen if you have any questions.