Source code and supplementary information for IJCNN2021.

The supplementary information is available at: https://kinwaicheuk.github.io/IJCNN2021.github.io/

The source code is available under the codes folder.

The MAPS dataset can be downloaded via https://amubox.univ-amu.fr/index.php/s/iNG0xc5Td1Nv4rR. Download and unzip the dataset into the codes folder.

After unzipping all the files inside, the ENSTDkAm1 and ENSTDkAm2 folder should be combined as one single folder ENSTDkAm.

Run the following scripts to train different models:

• python train_original.py with <args>
• python train_fast_local_attent.py with <args>
• python train_simple.py with <args>

The following arguments are available:

• device: choose what device to use. Can be cpu, cuda:0 or any device that is available in your PC. Default cuda:0.

• LSTM: Train the model with or without the LSTM layer. Either True or False. Default True.

• onset_stack: Train the model with or without the onset stack. Either True or False. Default True.

• batch_size: Setting the batch size. Default 16.

The following arguments are for train_fast_local_attent.py only

• w_size: The attention window size. Default: 30.
• attention_mode: Choosing which feature to attend to. Either onset or activation or spec. Default onset.

The PyTorch dataset class MAPS() inside each script will process and prepare the dataset if you are running it for the first time.

The weights can be downloaded here

The bash files contain all the commanands to obtain the results reported in the paper. Run the following bash scripts to get all the results

• bash get_table1.sh
• bash get_figure1.sh
• bash get_table2.sh

The accuracy reports and the midi files will be saved in the results folder.