Performance of onset detection using neural networks

Supplementary code to my thesis. The purpose of this repository is to make it easy for others to reproduce the results that I've reported in my thesis.

Prerequisites

The Python packages Keras, tensorflow and madmom. These can all be installed using pip. ffmpeg is also needed. If it isn't already installed it can be installed using:

$ cd /tmp && wget https://ffmpeg.org/releases/ffmpeg-4.1.tar.bz2 \
    && tar xvjf ffmpeg-4.1.tar.bz2 && cd ffmpeg-4.1 \
    && ./configure && make
$ export PATH=/tmp/ffmpeg-4.1:$PATH

The Böck dataset

First the training dataset has to be downloaded. This is done using the download.py script:

$ python download.py /tmp/
* Downloading document with id 1ICEfaZ2r_cnqd3FLNC5F_UOEUalgV7cv to /tmp/onsets.zip.
* Extracting /tmp/onsets.zip

The script hardcodes the dataset location to this url. If it ever changes then the DOC_ID constant in the script needs to be updated.

Configuration

Paths to input, output and cache data has to be configured by modifying the CONFIGS constant in the config.py file. The right config is selected during runtime by matching on the system and hostname. This way the same config.py can be used on multiple systems without requiring any changes.

The data-dir field should be set to the directory containing the Böck dataset, cache-dir to a directory storing cache files in pickle format and model-dir to the directory in which built models should be stored.

The seed field contains the seed to the random number generators ensuring that exactly the same results a produced every time. digest contains the checksum of the cache file. It is important that the cache file does not change during training or evaluation.

Training

Training is done using the main.py script:

$ python main.py -t 0:8 -n rnn --epochs 20

The above command would train the eight folds using the recurrent neural network (rnn) architecture for 20 epochs each.

Evaluation

Evaluation is done using the main.py script:

$ python main.py -e 0:1 -n rnn
...
sum for 41 files
  #:   3368 TP:   2861 FP:   387 FN:   507
  Prec: 0.881 Rec: 0.849 F-score: 0.865

The above command evaluates the first fold (with index 0) of the rnn architecture.