Combining Reduction and Dense Blocks for Music Genre Classification

This is the official repo of the paper entitled "Combining Reduction and Dense Blocks for Music Genre Classification" :https://link.springer.com/chapter/10.1007/978-3-030-92310-5_87

This repo contains a music genre classifier built using Reduction Block B of Inception-v4 and Dense Blocks. Using the same code we were able to achieve an accuracy of 97.51% and 74.39% over the GTZAN dataset and the small-subset of the FMA dataset respectively. Kindly note that we used some block of code from the following repository since our CNN architecture is inspired by the BBNN network. https://github.com/CaifengLiu/music-genre-classification

Main requirements:

1. Python 3.8.5
2. Tensorflow-gpu 2.2.0
3. Keras 2.3.0

To install the requirements before configuring and starting the training:

1. cd to the directory where requirements.txt is located.
2. activate your virtual environment.
3. run the following command to install all the requirements

pip install -r requirements.txt

Training

The training consist of converting music/audio tracks to spectrograms, slicing the spectrograms into normalized slices in width and height, and finally training the sices through the proposed network.

Parameters

In the train.py, the parameters were configured in order to train the small-subset of the FMA dataset. We enumerate below the paramters that should be changed in order to run the code on a custom dataset.

1. Change the dataset_path variable based on the path to your dataset
2. The get_all_music and slice_spectrograms are responsable to generate and slice the greyscale spectrograms. If you wish to save the slices in other directories their corresponding parameters should be changed.
3. Change the first parameter of the create_dataset_from_slices function based on the number of slices per genre in the use case.
4. Change the num_classes value based on the number of classes in the use case.

Hyperparameters

Concerning the hyperparameters, they are configured already for performing 10-Fold Cross Validation over the dataset. You can change the Fold value as well as following hyperparameters based on the use case:

1. k_fold : Number of folds

2. num_classes: Number of classes in the dataset

3. train_size: The training set ratio

4. val_size: The validation set ratio

5. test_size: The testing set ratio

6. epochs: The epochs size

7. batch_size: The Batch size

8. lr: The initial learning rate value

At this stage, you are ready to initiate the training by running the train.py using the python command. The results should be saved in the results/ directory based on the configuration of the following parameters:

1. file_name0: Path to the resulting models after each fold
2. path : Path to the log files
3. csv_name0: Path to the accuracy results at each fold

Citation

@InProceedings{10.1007/978-3-030-92310-5_87,
author="El Achkar, Charbel
and Couturier, Rapha{\"e}l
and At{\'e}chian, Talar
and Makhoul, Abdallah",
editor="Mantoro, Teddy
and Lee, Minho
and Ayu, Media Anugerah
and Wong, Kok Wai
and Hidayanto, Achmad Nizar",
title="Combining Reduction and Dense Blocks for Music Genre Classification",
booktitle="Neural Information Processing",
year="2021",
publisher="Springer International Publishing",
address="Cham",
pages="752--760",
abstract="Embedding music genre classifiers in music recommendation systems offers a satisfying user experience. It predicts music tracks depending on the user's taste in music. In this paper, we propose a preprocessing approach for generating STFT spectrograms and upgrades to a CNN-based music classifier named Bottom-up Broadcast Neural Network (BBNN). These upgrades concern the expansion of the number of inception and dense blocks, as well as the enhancement of the inception block through reduction block implementation. The proposed approach is able to outperform state-of-the-art music genre classifiers in terms of accuracy scores. It achieves an accuracy of 97.51{\%} and 74.39{\%} over the GTZAN and the FMA dataset respectively. Code is available at https://github.com/elachkarcharbel/music-genre-classifier.",
isbn="978-3-030-92310-5"
}

Contributing

Pull requests are welcome. For major changes, please open an issue first to discuss what you would like to change.

Please make sure to update tests as appropriate.

License

MIT