Genre Classifier is a multi-use package that allows you to:
1- create your own music dataset to be trained by a neural network
2- convert any copyright free mp3 to .wav file to either add to your dataset or evaluate with a model
3- train or use an existing model to attempt to guess the genre of your music
Install libraries via requirements.txt
Genre Classifier uses the Marsyas dataset to train it's model, you can download the file (1.2GB) here:
https://huggingface.co/datasets/marsyas/gtzan
*Note Jazz.00054.wav is corrupted in the dataset and should be removed
As a work in progress, Genre Classifier also supports the FMA-small dataset for training, you can download the file (7.2GB) here:
https://github.com/mdeff/fma - Credit also given to this group for preparing each version of dataset
Download the FMA-small package and fma_metadata.zip to utilize the package
Run the convert_mp3_to_wav.py file and select your .mp3 file
The file will be in the /test_output folder listed as converted_output.wav
You can use the file to be evaluated by the model, or incorportate it in test data under genre folder
Input the respective paths of your dataset and output beginning at line 10 of the prepare_dataset.py file
Take the Marsyas, FMA, or personal dataset and run the prepare_dataset.py file
The dataset will be converted to a h5 of mfcc values usable by the neural network model
Using a h5 file of mfcc values obtained from the prepare_dataset.py file, run the create_nn_model.py file to create your own model
Please note this model replaces the current model (.ckpt) file in the package
Use: tensorboard --logdir tb_logs/
to evaluate the accuracy of your model
Run the predict_song_genre.py file to predict the genre of a song of your choice
You will be prompted to submit a .wav file
Alternatively, run the app.py file and open a browser with the given locally hosted link to perform the same task
TODO: environment with relevant packages --done
TODO: implement GTZAN dataset and create genre folders with associated mfccs --done
TODO: implement FMA dataset and create h5 files with associated mfccs --done
TODO: 80-20 train test split --done
TODO: pytorch neural network --done
TODO: minimize loss/error --8/24 accuracy rate 70% suggestions include:
Further data augmentation: convolutions, segmentation, increase files/genre --done
TODO: implement a short front end to experience model results --done
TODO: Implement data augmentation for the FMA dataset
TODO: Increase model robustness, move to a CNN or LTSM netowrk
TODO: Improve browser capabilities to upload a given dataset
TODO: Increase training accuracy rate to 95% --8/25 accuracy rate 80%
Credits and study references to @ValerioVelardoTheSoundofAI
https://github.com/musikalkemist