noisifier

While writing my Bachelor’s Thesis, which was about alleviating the negative effects of noisy labels in remote sensing image classification, I needed a flexible tool that can easily insert different kinds of label noise in the training set. Because especially for multi-label classification, there are different types of noise one can add actually. You can only remove the positive classes; you can only remove the negative classes or you can do both at the same time. Then you can choose the percentage of your noise bla bla. So I could not find anything that I like. Well I could have probably found things if I searched enough. I saw the opportunity to write some code for adding noise and then packaging it. Just to learn how to publish a Python package and the bullshit you have to go through. That is why I decided to create a python library called noisifier that can add label noise to image labels.

So if you are interested in the package, visit the Github repository here: noisifier. You can check the pypi package here: noisifier.

In our research department, we have also created a website for collecting our work relating to noisy labels. Check it out at noisy-labels-in-rs.org for more information.