Layer Adaptive Out-of-Distribution detection for CS640

Preliminary results of LA-OOD on in-distribution CIFAR10 and OOD Tiny ImageNet on VGG16 backbone.

Team members

GM Harshvardhan

• Email: [email protected]

Harshil Gandhi

• Email: [email protected]

Kathakoli Sengupta

• Email: [email protected]

Overview

In this project, we will be using Wang et. al (2022)'s LA-OOD https://github.com/haoliangwang86/LA-OOD (paper: https://arxiv.org/pdf/2203.00192.pdf) for out-of-distribution detection on satellite images on So2Sat (https://github.com/zhu-xlab/So2Sat-LCZ42).

Implementation

The implementation pipeline is as follows:

1. Train the backbone models

python train_backbone_model.py --model vgg16 --dataset so2sat

2. Download the OOD datasets

Download the following dataset:

• (for preliminary) Tiny ImageNet: https://image-net.org/index.php
• (for final results) Satellite imagery dataset: https://www.kaggle.com/datasets/mahmoudreda55/satellite-image-classification

save the unzipped files in ./data folder

3. Generate the dataset

Generate the InD and OOD datasets:

python generate_datasets.py

4. Save the intermedia outputs

python save_inter_outputs.py --model vgg16 --ind so2sat

5. Train OOD detectors

python train_ood_detectors.py --model vgg16 --ind so2sat

The above step trains all the One-Class SVMs (OCSVMs) to detect whether an image is out-of-distribution or not.

6. Test: OOD detection

python detect_oods.py --model vgg16 --ind so2sat