Self-Supervised Feature Learning by Learning to Spot Artifacts [Project Page]

This repository contains demo code of our CVPR2018 paper. It contains code for unsupervised training on the unlabeled training set of STL-10 and code for supervised finetuning and evaluation on the labeled datasets.

Requirements

The code is based on Python 2.7 and tensorflow 1.12. See requirements.txt for all required packages.

How to use it

1. Setup

• Set the paths to the data and log directories in globals.py.
• Run init_datasets.py to download and convert the STL-10 dataset.

2. Unsupervised Training

• To pre-train the autoencoder run train_autoencoder_stl10.py
• To train the classifier and the repair network run train_stl10.py

3. Transfer & Evaluation

• To finetune the learnt representations run fine_tune_stl10.py
• To evaluate the finetuned classifier run test_classifier_stl10.py