Group Name: YOLO9000
Authors: Ahad Rauf, Chris Sun, Michael Lavva, Kei Watanabe
Link to the Paper: [URL]
For the final project of CS182, we created a robust computer vision classifier that performs well in a dataset that contains perturbations. To achieve this, we used various data augmentation and other deep learning model techniques, such as model ensembling, denoising, and adversarial training. These methods helped improve the robustness against both naturally perturbed and adversarial datasets. In addition, we implemented an explainable AI component to understand how the model makes its classification decisions.
This project requires PyTorch, OpenCV, NumPy, MatPlotLib, PIL, etc.
The dataset for this project is from https://tiny-imagenet.herokuapp.com/. To download it, go to the data folder in the master branch and run get_data.sh.
-
Create a conda environment and install the packages: conda create -n yolo9000-testing python=3.6 pip conda activate yolo9000-testing pip install -r requirements.txt
-
Create a directory called data at the top directory and place data set such that relative path becomes ./data/tiny-imagenet-200/train/
-
Create a directory called models at the top directory and download all models from the google drive folder and place them inside models directory : https://drive.google.com/drive/folders/1LVNFqUJAmhSGgYT4cVlzuImfZKi2ctyI?usp=sharing
-
Run
python test_submission.py 'eval.csv'where you need to replace 'eval.csv' with the path to the eval.csv
There are several branches that include the work in the project:
- master: Contains denoising script.
- dev-ahad-adversarial: Generates adversarial examples.
- dev-ahad-finite-horizons
- dev-ahad-master
- dev-ahad-xai: Code for Explainable AI.
- dev-kei-resnet-vgg: Model & Snapshot Ensembling code
- resnet: Code for ResNet model.
- test-submission: Test submission code.