Takanori Asanomi, Shinnosuke Matsuo, Daiki Suehiro, Ryoma Bise
Learning from label proportions (LLP) is a promising weakly supervised learning problem. In LLP, a set of instances (bag) has label proportions, but no instance-level labels are given. LLP aims to train an instance-level classifier by using the label proportions of the bag. In this paper, we propose a bag-level data augmentation method for LLP called MixBag, based on the key observation from our preliminary experiments; that the instance-level classification accuracy improves as the number of labeled bags increases even though the total number of instances is fixed. We also propose a confidence interval loss designed based on statistical theory to use the augmented bags effectively.
- python >= 3.9
- cuda && cudnn
We strongly recommend using a virtual environment like Anaconda or Docker. The following is how to build the virtual environment for this code using anaconda.
# pytorch install
$ pip install torch==1.9.1+cu111 torchvision==0.10.1+cu111 torchaudio==0.9.1 -f https://download.pytorch.org/whl/torch_stable.html
$ pip install -r requirements.py
You can create dataset by running following code. Dataset will be saved in ./data directory.
$ python create_dataset.py
The Data structure of ./data directory is written as belows.
./data
└── cifar10 # dataset name
│ ├── 0 # 5-fold
│ │ ├── train_bags.npy # train data (512, 10, 32, 32, 3) = (the number of bags, bag size, height, width, channel)
│ │ ├── train_labels.npy # train class label of each data (512, 10) = (the number of bags, bag size)
│ │ ├── train_lps.npy # train label proportions (512, 10) = (the number of bags, class label proportions)
│ │ ├── val_bags.npy # val data (10, 64, 32, 32, 3) = (the number of bags, bag size, height, width, channel)
│ │ ├── val_labels.npy # val class label of each data (10, 64) = (the number of bags, bag size)
│ │ └── val_lps.npy # val label proportions(10, 10) = (the number of bags, class label proportions)
│ │
│ ├── :
│ │
│ ├── 4 # 5-fold
│ │ ├── train_bags.npy # train data (512, 10, 32, 32, 3) = (the number of bags, bag size, height, width, channel)
│ │ ├── train_labels.npy # train class label of each data (512, 10) = (the number of bags, bag size)
│ │ ├── train_lps.npy # train label proportions (512, 10) = (the number of bags, class label proportions)
│ │ ├── val_bags.npy # val data (10, 64, 32, 32, 3) = (the number of bags, bag size, height, width, channel)
│ │ ├── val_labels.npy # val class label of each data (10, 64) = (the number of bags, bag size)
│ │ └── val_lps.npy # val label proportions(10, 10) = (the number of bags, class label proportions)
│ │
│ ├── test_data.npy # (10000, 32, 32, 3) = (the number of data, height, width, channel)
│ │
│ └── test_label.npy # (10000,) = (the number of data labels)
│
├── svhn # dataset name
│ ├── 0 # 5-fold
: :
After creating your python environment and Dataset which can be made by following above command, you can run Mixbag code.
If you want to train network, please run following command.
5 fold cross-validation is implemented and Test is automatically done in our code.
$ python run.py
If you want to train network in all 8 dataset, please run following command.
The training and test will be automatically done in 8 dataset.
(However, This process takes a lot of time, so you should be carefull.)
sh run_all_dataset.shYou can set up any parameters at arguments.py
If you find MixBag useful in your work, please cite our paper:
@inproceedings{asanomi2023mixbag,
title={MixBag: Bag-Level Data Augmentation for Learning from Label Proportions},
author={Asanomi, Takanori and Matsuo, Shinnosuke and Suehiro, Daiki and Bise, Ryoma},
booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision},
pages={16570--16579},
year={2023}
}
👤 Takanori Asanomi
- Github: @takanoriasanomi
- Contact: [email protected]
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent