Adversarial Motorial Prototype Framework for Open Set Recognition

Official PyTorch implementation of "Adversarial Reciprocal Points Learning for Open Set Recognition". It needs to say, these codes are modified by http://github.com/iCGY96/ARPL.

1. Requirements

Environments

These codes are supposed to be run with a Linux system. If you use Windows system to run them, it may encounter some errors.

Currently, requires following packages

• python 3.6+
• torch 1.4+
• torchvision 0.5+
• CUDA 10.1+
• scikit-learn 0.22+

Datasets

For Tiny-ImageNet, please download the following datasets to ./data/tiny_imagenet and unzip it.

• tiny_imagenet

2. Training

Open Set Recognition

To train open set recognition models in paper, run this command:

python osr.py --dataset <DATASET> --loss <LOSS>

Option

--loss can be one of ARPLoss/RPLoss/GCPLoss/Softmax/AMPFLoss.

--dataset is one of mnist/svhn/cifar10/cifar100/tiny_imagenet.

To run ARPL+CS, add --cs after this command.

To run MPF, just use --loss AMPFLoss.

To run AMPF, use --loss AMPFLoss --cs.

To run AMPF++, use --loss AMPFLoss --cs++.

3. Results

We visualize the deep feature of Softmax/GCPL/RPL/ARPL/MPF as below.

/MPFLoss_visualize_open_set.png width="800">

Before getting the figure above, you need to train the LeNet++ network, whose architecture is in "./models/model.py".

Citation

• If you find our work or the code useful, please consider cite our paper using:

  @misc{xia2021adversarial,
    title={Adversarial Motorial Prototype Framework for Open Set Recognition}, 
    author={Ziheng Xia and Penghui Wang and Ganggang Dong and Hongwei Liu},
    year={2021},
    eprint={2108.04225},
    archivePrefix={arXiv},
    primaryClass={cs.CV}
  }