Deep Representation One-class Classification (DROC).

This is not an officially supported Google product.

Tensorflow 2 implementation of the paper: Learning and Evaluating Representations for Deep One-class Classification published at ICLR 2021 as a conference paper by Kihyuk Sohn, Chun-Liang Li, Jinsung Yoon, Minho Jin, and Tomas Pfister.

This directory contains a two-stage framework for deep one-class classification example, which includes the self-supervised deep representation learning from one-class data, and a classifier using generative or discriminative models.

Install

The requirements.txt includes all the dependencies for this project, and an example of install and run the project is given in run.sh.

$sh deep_representation_one_class/run.sh

Download datasets

script/prepare_data.sh includes an instruction how to prepare data for CatVsDog and CelebA datasets. For CatVsDog dataset, the data needs to be downloaded manually. Please uncomment line 2 to set DATA_DIR to download datasets before starting it.

Run

The options for the experiments are specified thru the command line arguments. The detailed explanation can be found in train_and_eval_loop.py. Scripts for running experiments can be found

• Rotation prediction: script/run_rotation.sh

• Contrastive learning: script/run_contrastive.sh

• Contrastive learning with distribution augmentation: script/run_contrastive_da.sh

Evaluation

After running train_and_eval_loop.py, the evaluation results can be found in $MODEL_DIR/stats/summary.json, where MODEL_DIR is specified as model_dir of train_and_eval_loop.py.

Contacts

kihyuks@google.com, chunliang@google.com, jinsungyoon@google.com, minhojin@google.com, tpfister@google.com