music-classification-using-knn-cnn

Report and Presentation

For detailed project report and write up, please check this file.

For a abstract overview of our project, please refer to our slides.

Dataset

The Nsynth Dataset we used can be obtained here. We processed a subset of this dataset for use in our training. Due to the large size of the file (> 1GB), we could not put it up here. But we've included few samples of files in eg/tr and eg/ts.

CNN Usage

Prepare two directory containing your training set and validation set, respectively. Your dataset which should be a set of *.wav files.

You might want to change these variables at the top of featExtractionCNN.py to reflect your dataset.

# Unique index for every class/label
classes = {"string": 0, "keyboard": 1, "vocal": 2, "guitar": 3, "brass": 4}
train_dir = "nsynth-train" # Directory to your training set 
valid_dir = "nsynth-valid" # Directory to your test set
datafile = "./dataCNN.npz" # What to save the numpy arrays as

Now run the following to produce a .npz file that contains the processed data.

python3 featExtractionCNN.py

With this file, you can proceed to test your dataset with a set of pretrained weights.

python3 --test eg/weightsCNN.hdf5 [data.npz]

To train a new model, run

python3 --train [data.npz]

KNN Usage

Generate a .npz from the dataset, as described above. Then, run the kNN.

python knn_categorizer.py

Dependencies

The CNN model works with python 2 or 3. The KNN model works with python 2

• Keras 2.0.6 (For CNN only)
• Tensorflow 1.4.0 or 1.2.1 (For CNN only)
• Librosa 0.5.1
• Sklearn 0.18.1
• Numpy 1.13.3
• hdf5