LTF-Label-Transformation-Framework

The codes of paper "LTF: A Label Transformation Framework for Correcting Target Shift" on ICML2020.

Requirements

• PyTorch
• Python 3.6.5

Training

To train the LTF models, you need to download the pre-trained generative model in the https://drive.google.com/file/d/1PaeHIvjF8VFz2_kfiUrbl49pOiuyH4bh/view?usp=sharing, and unzip it as result folder.

To run the experiment for the fashion-mnist dataset:

$ python main.py --dataset='f-m'  --num_class=10 --c_epochs=20

dataset:

• f-m: fashion mnist
• mnist: mnist
• cifar10

num_class: The number of classes is 10

c_epochs: The training epochs for the classifier.

tweak: target shift setting

• 0: Random Dirichlet Shift In this shift, we randomly gener-ate a label distributionPTYby employing the Dirichletdistribution with different values of the concentration parameter α.
• 1: Tweak-One Shift To evaluate the performance on the largelabel probability quantification. In our experiments,the ratio of one class is set to[0.5,0.6,0.7,0.8,0.9],respectively, while ratios of other classes are uniform,
• 2: Minority-Class Shift To evaluate the performance on thesmall label probability quantification. In our experi-ments,[20%,30%,40%,50%]classes are set to 0.001,respectively, while ratios of other classes are uniform.

The results are shown as the output file, e.g. outputf-m010.txt

To run the experiment for the synthetic regression dataset:

$ python label_shift_regression.py --id=1

id: target shift setting

• 1: Left Gaussian The Gaussian with the mean -0.707.
• 2: Right Gaussian The Gaussian with the mean 0.707.
• 3: Mix Gaussian The Mix Gaussian of 1 and 2.
• 4: Random The target distribution is generated by a random network.

Reference

@inproceedings{guo2020ltf,
  title={LTF: A Label Transformation Framework for Correcting Label Shift},
  author={Guo, Jiaxian and Gong, Mingming and Liu, Tongliang and Zhang, Kun and Tao, Dacheng},
  booktitle={International Conference on Machine Learning},
  pages={3843--3853},
  year={2020},
  organization={PMLR}
}