This reposity is used for keras implementation of Music Auto-tagging based on CRNN.
The idea of CRNN for music classification comes from "Convolutional recurrent neural networks for music classification"
In that paper, there are 4 conv layers and 2 GRU layers in its network architecture. The whole structures shown by below figure:
This code refers to music-auto_tagging-keras .
The dataset that we used in this reposity is MagnaTagATune.
Please visiting the-magnatagatune-dataset for downloading it. This code needs Audio data and Tag annotations.
Furthermore, more information about MagnaTagATune will be found in music-auto_tagging-keras.
.
├── checkpoints # saving weight checkpoints
├── constants.py # const parameters
├── data
│ ├── mp3 # audios with mp3 format
│ └── npys # mel-spectrogram features
├── models.py # crnn model created by keras
├── preprocess.py # preprocess
├── README.md
├── run.py # train & test
├── requirements.txt # environment requirements
└── src # README source
└── imgPython version: 3.6.8
Install all requirements: pip install -r requirements.txt
Check and modify const parameters in constants.py
Run preprocess: python preprocess.py
Run train: python run.py train
Run test: python run.py test
ROC-AUC=0.692121