About

This is a musical instrument classifier that works with two different network architectures:

• Convolutional Neural Network
• Transformer

The report over this project can be found here.

Data

The dataset used is the Medley Solos DB. The instruments within the recordings are:

• Clarinet
• Distorted Electric Guitar
• Female Singer
• Flute
• Piano
• Tenor Saxophone
• Trumpet
• Violin

Training

During training, a random slice of two seconds are used to have a more diverse training set. Once training is complete, a file should be saved under ./trained_models/.

To train, run:

CNN:

python3 train.py --model cnn --epochs 15 --lr 1e-4

Transformer:

python3 train.py --model transformer --epochs 15 --lr 1e-4

where --lr is the learning rate (1e-4 was found to be the best) and --epochs is the number of epochs.

Once training is completed, the loss over training epochs should be displayed. Also, a confusion matrix from the test set will be displayed along with the accuracy report printed to console.