This repo contains a simple implementation of a CNN for classifying music genres. It is implemented in Pytorch and Pytorch Lightning. The repo also comes with a VSCode .devcontainer configuration to allow for easy development using VSCode. It was implemented as a homework for an interview.
Either open Dev container using VSCode to have all requirements installed automatically or manually install packages using pip install -r requirements.txt.
You can download the dataset and perform initial pre-processing using make download-dataset. After this step run python scripts/prepare_data.py to pre-process the data to be used by the model. You can use the --augment-dataset flag to enable data augmentation. You can control the number of augmentations per wav file using --augments-per-sample.
All scripts are located in the scripts folder. The data will be downloaded to .data.
Below, you can see a small description of each file:
create_dataset_csv.py: This file contains the script used for generating a csv file of the paths of the wav files from the original dataset. It is ran automatically withmake download-dataset.data_exploration.ipynb: This notebook contains code for doing data exploration on the data.gtzan_dataset.py: This file contains the custom Pytorch DataSet class which is used by the model.models.py: This file contains the implementation of the GTZAN classifier model.prepare_data.py: This file contains the methods for converting wav files to mel-spectograms alongside a simple pipeline for processing the data to be used by the model.train.ipynb: This notebook contains the main training loop which used the other classes to train a model and evaluate it.vscode_audio.py: This file contains a workaround to enable playing wav files in VSCode notebooks. (not important)
- VSCode
- Docker
- VSCode plugins needed for remote development. For more information visit here.