Hierarchical-Note-Segmentation

Realization for note segmentation by using hierarchical objective function with utils of Resnet-18 or attention-RNN

Requirements

Before starting all the tasks bellow, make sure the environment is well set:

• python 3.5
• pytorch 0.4.1
• torchvision 0.2.1
• mir-eval 0.4

Demo

For full and real-world demo, please refer to https://github.com/Itachi6912110/WAV2MIDI
For supplementary and results preview, please refer to https://www.dropbox.com/sh/my2zxv9g6if6xvt/AACL-7w2Tc8ezjUHVtdyrDG2a?dl=0

Loading Data

bash/dataset.sh

This will load the data you need for training and evaluation.
Training data are from TONAS dataset, and testing data are from ISMIR2014 evaluation dataset.
The features used for training is extracted by the tool at https://github.com/leo-so/VocalMelodyExtPatchCNN
Note that the features for training may be too large, and may lead to download abort.

Training

For training, you can change hyper-parameters in the scripts train_sdt6_resnet_top.sh or train_sdt6_top.sh

• For training resnet-18

bash script/train_sdt6_resnet_top.sh

• For training with Data Augmentation and semi-supervised learning

bash script/train_sdt6_resnet_semi_top.sh

• For training with Data Augmentation

bash script/train_sdt6_resnet_aug_top.sh

• For training with Data Augmentation and 5-class

bash script/train_5class_resnet_semi_top.sh

• For training rnn-attn-19

bash script/train_sdt6_top.sh

Evaluation

For evaluation, you can change hyper-parameters in the scripts eval_resnet_fmeasure.sh or eval_sdt6_fmeasure.sh

• For evaluating on resnet-18, with or without SSL

bash script/eval_resnet_fmeasure.sh

• For evaluating on resnet-18, with or without SSL and on 10 epochs

bash script/eval_resnet_fmeasure.sh 1

• For evaluating on resnet-18, with or without SSL and on 5-class

bash script/eval_5class_resnet_fmeasure.sh 1

• For evaluating on rnn-attn-19

bash script/eval_sdt6_fmeasure.sh

Visualization

You can visualize your results by the following command

python3 src/plot_P.py <file#>