CNN for music genres
A CNN for music genre classification and TSC in general
CNN for music genres
This post presents a CNN for music genre classification. Over the last weeks, I got many positive reactions for my implementations of a CNN and LSTM for time-series classification. With this post, we stretch the TSC domain to long signals. Music has a typical sample frequency of 44.1 kHz. For basic classification, you need at least 1 or 2 seconds of data. That implies a signal length of roughly 50.000 samples. This post implements a CNN with accuracies around 90%.
Downsampling architecture
The implementation downsamples in two stages. At first, extract_music.m loads the data and applies a hard-coded downsample with factor 30. The resulting sample frequency is 1470Hz. The second downsampling occurs after the first conv-layer in CNN_music_main.py. An immediate downsampling with factor 90 would discard useful information. The first conv-layer in the CNN graph can extract useful information. The strided max-pooling condenses information with a factor 3. With a combined convolution and max-pooling, we allow for valuable information to be passed to the second conv-layer
Your own use
The goal is to show that even long sparse signals, like music, also allow for time-series classification. In this project, the MATLAB code in extract_music.m extract chunks of signal from the music in a specific directory. You can cut down the project at every layer for your own use
- At the base level, you can import your own music.
- In the MATLAB code, there's many knobs to play with
- The chunk-length
- The down-sample factor
- For better performance, you can synchronize the chunks with the beat of the music
- You can discard the chunking all together and feed the music at original sample frequency to the CNN.
- At last, the CNN allows for many tuning, like the architecture, learning rate, number of neurons, regularization, dropout and so on.
Output
With the implementation set in the .py file, you can expect these outputs:
The evolution of the loss-function and the accuracies.
The graph in TensorBoard
As always, I am curious to any comments and questions. Reach me at romijndersrob@gmail.com