aRI0U/recvis19_a3

Object recognition and computer vision 2019/2020

Assignment 3: Image classification

Requirements

Install PyTorch from http://pytorch.org
Run the following command to install additional dependencies

pip install -r requirements.txt

(Optional) Install CUDA to enable GPU acceleration

Dataset

We will be using a dataset containing 200 different classes of birds adapted from the CUB-200-2011 dataset.

On Linux/Mac, the dataset can be downloaded by executing

./download_dataset.sh

Training and validating your model

Run the script main.py to train your model.

Noisy Student

To increase the accuracy of the predicted labels, you can use a method called Noisy Student. When running main.py, add option --student N to add the less uncertain unlabeled images to the training set.

Evaluating your model on the test set

As the model trains, model checkpoints are saved to files such as model_x.pth to the current working directory. You can take one of the checkpoints and run:

python evaluate.py --data [data_dir] --model [model_file]

That generates a file kaggle.csv that you can upload to the private kaggle competition website.

Results

Table below details the effect of improvements on the overall accuracy:

Improvement	Accuracy (%)
None	61.9
Data augmentation	69.0
DA + LR scheduler	77.4
DA + scheduler + Noisy Student	78.1

References

@article{NoisyStudent, 
arxivId = {1911.04252},
author = {Xie, Qizhe and Hovy, Eduard and Luong, Minh-Thang and Le, Quoc V.},
month = {nov},
title = {{Self-training with Noisy Student improves ImageNet classification}},
url = {http://arxiv.org/abs/1911.04252},
year = {2019}
}

@article{ResNext101,
arxivId = {1905.00546},
author = {Yalniz, I. Zeki and J{\'{e}}gou, Herv{\'{e}} and Chen, Kan and Paluri, Manohar and Mahajan, Dhruv},
month = {may},
title = {{Billion-scale semi-supervised learning for image classification}},
url = {http://arxiv.org/abs/1905.00546},
year = {2019}
}

Acknowledgments

Adapted from Rob Fergus and Soumith Chintala https://github.com/soumith/traffic-sign-detection-homework.
Adaptation done by Gul Varol: https://github.com/gulvarol

aRI0U / recvis19_a3