Skip to content

Latest commit

 

History

History
128 lines (88 loc) · 4.86 KB

IMAE_Code_Illustration.md

File metadata and controls

128 lines (88 loc) · 4.86 KB
Raw

IMAE for Noise-Robust Learning: Mean Absolute Error Does Not Treat Examples Equally and Gradient Magnitude's Variance Matters

Following work: Derivative Manipulation for General Example Weighting

For source codes, the usage is conditioned on academic use only and kindness to cite our work: Derivative Manipulation and IMAE.
As a young researcher, your interest and kind citation (star) will definitely mean a lot for me and my collaborators.
For any specific discussion or potential future collaboration, please feel free to contact me. 

@article{wang2019imae,
  title={ {IMAE} for Noise-Robust Learning: Mean Absolute Error Does Not Treat Examples Equally and Gradient Magnitude's Variance Matters},
  author={Wang, Xinshao and Hua, Yang and Kodirov, Elyor and Robertson, Neil M},
  journal={arXiv preprint arXiv:1903.12141},
  year={2019}
}

Downloading Link

https://www.dropbox.com/sh/pmrnh4kdrxbisbj/AACtxvIOc7AoO9xLeySfHvgLa?dl=0

Command for Visualising the Repository/Directory Tree Structure

cd directory_name
tree

Dependencies

The core functions are implemented in the caffe framework. We use matlab interfaces matcaffe for data preparation.

Setup

  • Install dependencies on Ubuntu 16.04

    sudo apt-get install libprotobuf-dev libleveldb-dev libsnappy-dev libopencv-dev libhdf5-serial-dev protobuf-compiler
sudo apt-get install --no-install-recommends libboost-all-dev
sudo apt-get install libopenblas-dev
sudo apt-get install python-dev
sudo apt-get install libgflags-dev libgoogle-glog-dev liblmdb-dev

  • Install MATLAB 2017b

    Download and Run the install binary file

    ./install

  • Compile Caffe and matlab interface

    Note you may need to change some paths in Makefile.config according your system environment and MATLAB path

    cd ../CaffeMex_UnifiedWeight_V01
make -j8  && make matcaffe

cd ../CaffeMex_MAE_V00
make -j8  && make matcaffe

Usage

Examples for reproducing our results on CIFAR-100 are given.

  • Data preparation for CIFAR-100

    • Prepare testing data:
    cd CIFAR100_Data_Toolkit
matlab -nodisplay -nosplash -nodesktop -r "run('test_data_preparation.m');exit;" | tail -n +11
    • Prepare training data (symmetric noise rate: 0.0, 0.2, 0.4, 0.6):
    matlab -nodisplay -nosplash -nodesktop -r "run('train_data_preparationV2_noise_0_0.m');exit;" | tail -n +11
matlab -nodisplay -nosplash -nodesktop -r "run('train_data_preparationV2_noise_0_2.m');exit;" | tail -n +11
matlab -nodisplay -nosplash -nodesktop -r "run('train_data_preparationV2_noise_0_4.m');exit;" | tail -n +11
matlab -nodisplay -nosplash -nodesktop -r "run('train_data_preparationV2_noise_0_6.m');exit;" | tail -n +11
    • Copy data
    cd .. 
echo CIFAR100_ResNet44_V03_lambda1/pre_pro_process | xargs -n 1 cp CIFAR100_Data_Toolkit/TestImageDataCell.mat
echo CIFAR100_ResNet44_V03_lambda1/pre_pro_process | xargs -n 1 cp CIFAR100_Data_Toolkit/TrainImageDataCell0.0.mat
echo CIFAR100_ResNet44_V03_lambda1/pre_pro_process | xargs -n 1 cp CIFAR100_Data_Toolkit/TrainImageDataCell0.2.mat
echo CIFAR100_ResNet44_V03_lambda1/pre_pro_process | xargs -n 1 cp CIFAR100_Data_Toolkit/TrainImageDataCell0.4.mat
echo CIFAR100_ResNet44_V03_lambda1/pre_pro_process | xargs -n 1 cp CIFAR100_Data_Toolkit/TrainImageDataCell0.6.mat

  • Train & Test

    Run the training and testing scripts in the training folder of a specific setting defined by its corresponding prototxt folder.

    For example,

    cd CIFAR100_ResNet44_V03_lambda1/train_Res44_USW_Beta02_lambda1_0.6
matlab -nodisplay -nosplash -nodesktop -r "run('train.m');exit;" | tail -n +11
matlab -nodisplay -nosplash -nodesktop -r "run('test.m');exit;" | tail -n +11

Our trained results

  • Our trained results are stored in corresponding folders. For example, in Folder CIFAR100_ResNet44_V03_lambda1/train_Res44_USW_Beta02_lambda1_0.6, there are:

    • accuracy.txt
    • accuracy_curve.png

  • Without changing the random seed (123), you are supposed to obtain exactly the same results.

Acknowledgements

Our implementation benefits from: