RE' Learning and Evaluating Representations for Deep One-Class Classification

This repository contains code to reproduce results from Table 2 and Table 7 of the paper: "Learning and Evaluating Representations for Deep One-Class Classification" as part of the COMP6248 UoS Reproducability Challenge.

Reproduction

Each folder contains scripts to generate the respective method representations and subsequently perform one class-classification with linear and RBF kernel OC-SVMs.

• ResNet18-50_Baseline_Model: reproduction of experiments on ResNet18 (random weights) and an ImageNet pre-trained ResNet50 on f-MNIST, CIFAR10 and CIFAR100.
• Denoising_Model: reproduction of experiments with a denoising autoencoder on fMNIST and CIFAR10.
• Rotation_Prediction_Model: reproduction of experiments with a rotation prediction ResNet18 network on fMNIST.
• SimCLR: reproduction of experiments with the SimCLR network on fMNIST and CIFAR10.
• Table_Means_Verification: verification of the row means of all tables in the paper.

Requirements

- Python=3.7
- PyTorch=1.8
- torchvision=0.9
- scikit-learn=0.22

Team members:

• Niko Chazaridis (@chazarnik)
• Marios Christodoulou (@mchris7)
• Ian Simpson (@statsonthecloud)

---