Skip to content

Latest commit

 

History

History

training

README.md

Source code for preparing the training data

Note that the source code is not well-organized. But we think simply releasing the code can still be useful.

Step 1 - Collect qualified piano-rolls

python tracks_parsing.py

Given a list of qualified MSD ids (Rock_C_id here), this file merge the data into five tracks: Bass, Drums, Guitar, Piano and Strings. Instruments out of the five categories are considered as part of the strings.

  • Multiple tracks in LMD => {Drum, Piano, Guitar, Bass, Others}
  • 2844 selected in 6646

The collected piano-rolls are placed in tracks/ folder with names "[instrument_name]/[msd_id].npz", e.g. Drum/TRAAEEH128E0795DFE.npz.

Step 2 - Prepare files for structure analysis

python "song_analysis.py"

This python file does the following.

  • Convert piano-rolls to .mat format

    • Piano_Roll.mat: matrix for running segmentation in MATLAB. The shape is (128, T).

  • Create activation labels

    • act_instr/[msd_id].npy: 2D numpy array of shape (num_bar, 5). Indicate whether the corresponding instrument is activated (non-empty) in each bar.
    • act_all/[msd_id].npy: list of bool. Indicate whether the number of activated instruments is larger than a threshold (3 here).

The results are placed in the trakcs/ folder.

Step 3 - Segment and label the collected piano-rolls

matlab main_seg.m
matlab main_lab.m

These files perform the segmentation and labeling, respectively. The algorithm we employed is called Structural Feature [1, 2]. Although it is designed for raw audios, we found it works well on symbolic data (i.e. piano-rolls) as well. For python user, you can use MSAF to perform structure analysis as well.

The code is modified from Wayne's previous project — Music-Structure-Analysis-in-Matlab [3].

The analysis results are placed in the structure/ folder.

Step 4 - Collect the training data

python gen_data_bar.py

This file collect the training data for the five tracks and generate an additional act_instr track which indicates the instrument activations. The training data is filtered by act_all and collected following some rules on top of the analysis results obtained in Step 3.

References

  1. Joan Serrá, Meinard Müller, Peter Grosche and Josep Ll. Arcos, "Unsupervised Detection of Music Boundaries by Time Series Structure Features, in AAAI Conference on Artificial Intelligence (AAAI), 2012

  2. Joan Serrá, Meinard Müller, Peter Grosche and Josep Ll. Arcos, "Unsupervised Music Structure Annotation by Time Series Structure Features and Segment Similarity", in IEEE Transactions on Multimedia, vol. 16, no. 5, pp. 1229-1240, 2014.

  3. https://github.com/wayne391/Music-Structure-Analysis-in-Matlab