noisifier

noisifier is a simple, yet effective python library.

noisifier allows you to add noise to the labels of your dataset. You can use the noisified dataset in your experiments, also to train your machine learning model robustly against label noise.

Your dataset can be single label or multi-label; just create the right type of noisifier and keep adding noise.

Prerequisites

python 3.8.5

numpy 1.19.2

Installation

pip install noisifier

Link to the project

https://pypi.org/project/noisifier/

How to Use

Convert your Tensorflow or PyTorch tensors into numpy arrays with tensor.numpy(), or load a numpy array. The output will be numpy array. Convert to tensorflow tensor with tf.convert_to_tensor(y) and to PyTorch tensor with torch.from_numpy(y).

Caveats

Provide the input in one-hot encoded form. If you use keras, you can do that by

y = keras.utils.to_categorical(y, number_of_classes)

Use a float between 0. and 1. for all the noise rates.

Tests

To run all the tests:

python -m unittest discover test

To run partial test:

python -m unittest test.test_multi_label

Licence

noisifier is released under the MIT licence.

Version

0.4.6