An open API service indexing awesome lists of open source software.

https://github.com/kremerj/relabeling

Code repository for the robust active label correction paper.
https://github.com/kremerj/relabeling

active-learning aistats-2018 convolutional-neural-networks label-noise logistic-regression

Last synced: 8 months ago
JSON representation

Code repository for the robust active label correction paper.

Awesome Lists containing this project

README

          

# Robust Active Label Correction

This is the code repository complementing the paper

Jan Kremer, Fei Sha, and Christian Igel. [Robust Active Label Correction](http://proceedings.mlr.press/v84/kremer18a.html). *PMLR: Volume 84 (AISTATS)*, 2018

```
@inproceedings{Kremer18,
author = {J. Kremer and F. Sha and C. Igel},
title = {Robust Active Label Correction},
booktitle = {Proceedings of the 21st International Conference on Artificial Intelligence and Statistics},
series = {Proceedings of Machine Learning Research},
year = 2018,
volume = 84,
publisher = {PMLR}
}
```

Please cite us if you use any of the code provided here. All experiments from the paper can be reproduced from this repository. We use Python 3 and tensorflow 1.4. You can create and activate the conda environment by running

```
conda env create --file environment/relabeling.yml
source activate relabeling
```

or if you have GPU support

```
conda env create --file environment/relabeling-gpu.yml
source activate relabeling-gpu
```

To get the necessary data and the pretrained [model weights](http://www.cs.toronto.edu/~guerzhoy/tf_alexnet/) for the CNN experiment, run

```
sh scripts/fetch_model.sh
```

and get the data (images and annotions) from http://bit.ly/2Duy6nK and should be unpacked into ```data/baidu```.

The necessary Cython code can be compiled by calling

```
sh scripts/compile.sh
```

The logistic regression experiments can be reproduced by calling

```
sh scripts/relabeling.sh
```

The results can be found in ```output/experiment/std```.

The CNN experiments can be reproduced by calling

```
sh scripts/relabeling_deep.sh
```

The results can be found in ```output/experiment/deep```.

All plots can be generated by calling

```
sh scripts/generate_plots.sh
```

The figures can be found in ```output/experiment/std/figures``` for the logistic regression experiments and in ```output/experiment/deep/figures``` for the CNN experiment.

A single experiment can be run by calling

```
python relabeling.py
```

The command-line help should guide you regarding available options.