Max-MIG

Max-MIG copied to clipboard

Introduction

Implementation for ICLR2019 paper MAX-MIG: AN INFORMATION THEORETIC APPROACH FOR JOINT LEARNING FROM CROWDS

paper link: https://openreview.net/forum?id=BJg9DoR9t7

arxiv : https://arxiv.org/abs/1905.13436

[Slide]

Synthesized Crowd-sourcing dataset

• To run experiments of Dogs vs. Cats dataset in Dogs vs. Cats directory:

python3 main.py --case case_num --expertise ex --path path_to_dataset --device device_num

case_num: number of experimental case( see our paper) 
		1: Independent mistakes
		2: Naive majority
		3: Correlated mistakes
expertise: the expertise of senior expertise
		0: Low expertise
		1: High expertise
		
path_to_dataset: path to the dataset

device_num : GPU number
		

• To run experiments of CIFAR-10 dataset in Cifar10 directory:

python3 main.py --case case_num --expertise ex --path path_to_dataset --device device_num

case_num: number of experimental case( see our paper) 
		1: Independent mistakes
		2: Naive majority
		3: Correlated mistakes
expertise: the expertise of senior expertise
		0: Low expertise
		1: High expertise
		
path_to_dataset: path to the dataset

device_num : GPU number
		

• To run experiments of LUNA dataset in LUNA16 directory:

python3 main.py --case case_num --expertise ex --path path_to_dataset --device device_num

case_num: number of experimental case( see our paper) 
		1: Independent mistakes
		2: Naive majority
		3: Correlated mistakes
expertise: the expertise of senior expertise
		0: Low expertise
		1: High expertise
		
path_to_dataset: path to the dataset

device_num : GPU number
		

Real world crowd-sourcing dataset:

• To run experiments of Labelme dataset in labelme directory:

python3 cotraining_labelme.py  --device device_num

device_num : GPU number
		

The Labelme dataset can be downloaded at http://fprodrigues.com//deep_LabelMe.tar.gz . Please place prepared document in the same folder with your code. :smile:

To cite our paper:

@article{cao2018max,

  title={Max-MIG: an Information Theoretic Approach for Joint Learning from Crowds},

  author={Cao, Peng and Xu, Yilun and Kong, Yuqing and Wang, Yizhou},

  year={2018}

}