Awesome

Deep Bilevel Learning [Project Page]

This repository contains demo code of our ECCV2018 paper. It contains code for bilevel training of an Inception network on the CIFAR-10 dataset with noisy labels.

Requirements

The code is based on Python 2.7 and tensorflow 1.12.

How to use it

1. Setup

• Set the paths to the data and log directories in constants.py.
• Run init_cifar10.py to download and convert the CIFAR-10 dataset. This also creates several noisy label files.

2. Bilevel training and evaluation

• To train the Inception network with bilevel training on CIFAR-10 run experiments_Inception_CIFAR10_bilevel_noise.py.
• The bilevel algorithm is implemented in the build_model method of the BilevelTrainer.