kylelo / music-genre-cnn-classifier

A two-class music genre classfier based on CNN (Convolutional Neural Network)

Music Genre CNN Classifier

A two-class music genre classifier based on CNN (Convolutional Neuron Network).

What does this project do?

1. Extract Mel-spectrogram as the training set from two groups of songs with different genre

2. Train the CNN model (on your PC or Google TPU)

3. Evaluate the CNN model using the validation set

4. Visualize each kernel (For learning purpose)

NOTE that this project takes the classification of prograssive rock and non-prograssive rock musics as an example. Users can feel free to change them to other music genres.

|
|- code
|   |
|   |-- generatePatterns.py
|   |-- trainModel.py
|   |-- evaluateModel.py
|   |-- toolbox
|         |-- featureExtractTool.py (helper functions)
|         |-- evaluateModelTool.py  (helper functions)
|         |-- trainModelTPU.py   (optional)
|         |-- generateTestset.py (optional)
|         |-- visualizeKernel.py (optional)
| 
|- training songs (create manually)
|    |
|    |-- class1
|    |-- class2
|
|- validation songs (create manually)
     |
     |-- class1
     |-- class2

Prerequisites

1. Python3

Linux

sudo apt-get update
sudo apt-get install python3.6

MacOS

brew install python3

2. Feature extraction tool

pip3 install librosa

3. Model training tools

pip3 install keras tensorflow

Extract Mel-spectograms from Songs as Training set

Skip this step by downloading default training sets (Class1: Prop songs and Class2: Non-prog songs)

1. Put two groups of songs (.mp3) into ./training song/class1 and ./training song/class2

2. Run

python3 generatePatterns.py

3. After extraction, user will get training sets Class1Patterns.pkl and Class2Patterns.pkl

Train CNN Model

Skip this step by using the default trained model to try validation set

1. Run

python3 trainModel.py

2. After training, user will get cnnModel.h5 and Scalers.sav

Scalers are used for training set normalization (make values in each channel between -1 and +1). It is necessary to use the same scalers to normalize validation songs.

Train CNN Model using TPU (Optional)

1. Create a notebook .ipynb on Colab

2. Copy the code in trainModelTPU.py to notebook

3. Upload class1Patterns.pkl and class2Patterns.pkl to ./Colab Notebooks on your Google drive

4. Run and get cnnModel.h5 on the Google drive

Validate CNN Model

1. Put two groups of songs (.mp3) into ./validation songs/class1 and ./validation songs/class2

Songs for validation should be different from the ones for training

2. Run

python3 evaluateModel.py

3. See results in console

Visualize CNN Kernels (Optional)

This feature is for learning purpose. This part of the code is referenced from an amazing post Visualization of Filters with Keras (Yumi's Blog)

python3 visualizeKernel.py

Feel free to contact me if you find a bug or have any question :D