Authors:
- Oleh Pomazan [email protected] 31566448
- Iryna Tyshchenko [email protected] 45983811
- Vitalii Tsiapa [email protected] 37864347
Github repo https://github.com/DanmerZ/ml_assignment
The whole solution in a single Jupyter notebook deployed on Kaggle https://www.kaggle.com/code/danmerz/urbansound8kclassification
- report.ipynb: main solution with explanations
- knn.py: implementation of k-nearest neighbors classification in Numpy
- features.py: audio features extraction code using librosa library
- model.py: implementation of VGG11 CNN for spectrogram classification
- dataset.py: PyTorch datasets and dataloaders
- train.py: training VGG1 using spectrograms
In order to train VGG11 model download Urban8kSound dataset into input
directory:
input
├── audio
├── FREESOUNDCREDITS.txt
├── metadata
└── UrbanSound8K_README.txt
Pre-trained weights for VGG11 model and extracted_features.csv/.pcl
are in data
directory.