Snare drum playing methods classification by using CNN (4-layer CNN and PANNs ResNet38)

Overview

I used 3 CNN models to classify a single note on the snare drum into 4 techniques(Strike, Rim, Cross Stick and Buzz). In the case of 4-layer CNN, the accuracy is 79.1%.

Requirement

Python 3.8.3

Pytorch 1.9.1

Data

Percussion Dataset

Model

4-layer CNN

PANNs ResNet38 (pretrained or not)

Run

Please clone this repo and download the Percusion Dataset and the ResNet38 pretrained model(ResNet38_mAP=0.434.pth).

The downloaded data looks like:

-PANNsResNet38_fineturing
-simpleCNN
-data
  └-MDLib2.2
     |-_MACOSX
     └-MDLib2.2
        |-Sorted
        | └-...
        |...
-model
  └-ResNet38_mAP=0.434.pth

If you want to use 4-layer CNN, run simpleCNN/train.py, if you want to use ResNet38, run PANNsResNet38_finetuning/train4snare.py with the python command.

You can choose whether or not to pre-train ResNet38 by commenting out line 143 of PANNsResNet38_finetuning/train4snare.py.

142:    PRETRAINED_CHECKPOINT_PATH = '../data/model/ResNet38_mAP=0.434.pth'
143:    model.load_from_pretrain(PRETRAINED_CHECKPOINT_PATH)  #If you want to train without pretraining, comment out this line
144:    model = torch.nn.DataParallel(model)

Result

model	epoch	accurancy
ResNet38 (pretrained)	30	44.0%
ResNet38 (not pretrained)	30	64.8%
4-layer CNN	10	79.1%

Discussion

For a single percussion instrument, which tends to be seen as monotonous, DNN could detect the differences between the 4 playing methods with about 80% accuracy.

The result is that the accuracy of the deep model is lower. This is different from what is commonly known. It may be that a simple CNN is better suited for this task, or it may be that my code is inadequate. If you notice anything, please message me.

Reference

• YouTube channel "Valerio Velardo - The Sound of AI" - Playlist "Pytorch for Audio + Music Processing"
• PANNs: Large-Scale Pretrained Audio Neural Networks for Audio Pattern Recognition