50.039 Theory and Application of Deep Learning: Project
Git clone this project. To predict the music genre of your audio file, just use
$ python predict_audio.py --music_path [music-path] # defaults to use our final model
The outputs will be the classification labels for each predetermined chunk of the music file that is aggregated into its normalised value counts. For example,
Predictions for test.mp3:
classical 0.989474
jazz 0.010526
This means that the model is 98.9% percent sure that it is of the classical genre.
- Python 3.9 and above
- torch
- sklearn
- pandas
- numpy
- librosa
- tqdm
- seaborn
- scikit-plot
- matplotlib
- ffmpeg (for mp3 input - you might need to install it as a standalone app from their website or your preferred package manager if pip installing still yields audio backend error)
Alternatively, you can install from the requirements file from our virtual environment:
$ pip install -r requirements.txt
$ python [script-name] -h # for help
mfcc_extractor.pywill extract MFCC features from the dataset.process.pywill process and split the data from the output ofmfcc_extractor.pyinto train and test sets if needed.cnn_2d_parallels.pywill train the model based on the outputs ofprocess.py.predict_audio.pywill run the model chosen to output a prediction for a music file.
The config used to create the final model is as follows in cnn_2d_parallels.yaml:
n_epochs: 15
batch_size: 32
optimiser_cfg:
lr: 0.001
We initially trained the model using a 20% test split. These are the results. The actual model is trained on the full dataset.
From Models/2022-04-24_21-38-09_CNN_2D_Split_True.pt:
precision recall f1-score support
blues 0.89 0.73 0.80 208
classical 0.88 0.97 0.92 202
country 0.64 0.75 0.69 192
disco 0.90 0.61 0.72 201
hiphop 0.83 0.78 0.80 209
jazz 0.95 0.76 0.85 186
metal 0.81 0.93 0.87 211
pop 0.72 0.88 0.79 204
reggae 0.88 0.73 0.80 212
rock 0.52 0.70 0.60 175
accuracy 0.79 2000
macro avg 0.80 0.78 0.78 2000
weighted avg 0.81 0.79 0.79 2000
