IRMAS TORCH

Convolutional Neural Networks applied to the IRMAS.

This project applies known architectures to the task of predominant musical instrument identification in polyphonic recordings.

The networks are implemented in PyTorch, and rely on TorchAudio and TorchVision.

Content:

The src folder contains all the utilities, such as transforms and datasets.
The train.py file is a commandline utility to start the training process.
The test.py file is a script to perform predictions on the the test set.
On the nbs folder, notebooks displaying the results can be seen.
The dataset, or a symlink to it, should be placed on the folder data.
The trained models and predictions results are written on the data folder by default, this can be changed on the config.py file.