USADTM

This repo is the official implementation of the paper of USADTM:Unsupervised Semantic Aggregation and Deformable Template Matching for Semi-Supervised Learning. The code is developed based on C^3 Framework.

Quick Start

Environment:

GTX-1080ti gpu
ubuntu-18.04
Python 3.7.6

Install

Clone this repo: git clone https://github.com/taohan10200/USADTM.git
Install other dependencies:pip install -r requirements.txt

Data preparation

Manual way

You need to manually download the CIFAR10, CIFAR100,

STL10, SVHN(train), SVHN(test)

The directory tree of dataset shold be like this:

$USADTM/dataset
├── CIFAR10
|    ├── cifar-10-batches-py
|    └── cifar-10-python.tar.gz
├── CIFAR100
|    ├── cifar-100-python
|    └── cifar-100-python.tar.gz
├── STL10
|    ├── stl10_binary.tar.gz
|    └── stl10_binary
├── SVHN
|    ├── train_32x32.mat
|    └── test_32x32.mat

Automatic way
- Instead of creating the dataset by hand, the datasets and the tree will be downloaded and created automatically when excuting the command: python train.py.

Train the model

To train the model, you should set the correct parameters in config.py. In particular, the following parameters need to be confirmed:

__C.dataset = 'CIFAR10'  # 'CIFAR10' ,'CIFAR100', 'STL10', 'SVHN'
__C.n_classes = 10
__C.n_labeled = 250     # option: CIFAR10[40, 250,  4000]  CIFAR100[400,2500,10000] STL10, SVHN[40, 250,  1000]
__C.batch_size = 64     # train batch size of labeled samples

where __C.n_labeled is the number of labeled sample during training.

Training with the command:

     $ python train.py

Monitoring training progress with the command:

     $ tensorboard --logdir=exp --port=1111

Results

After training, the metrics curves might be observed like this :

Citation

If you find this project is useful for your research, please cite:

@inproceedings{han2020unsupervised,
  title={Unsupervised Semantic Aggregation and Deformable Template Matching for Semi-Supervised Learning},
  author={Han, Tao and Gao, Junyu and Yuan, Yuan and Wang, Qi},
  journal={Advances in Neural Information Processing Systems},
  volume={33},
  year={2020}
}