Pytorch implementation of Mahendran & Vedaldi, CVPR, 2015, with automated hyperparameter optimization

This repository contains a Pytorch-based implementation of Mahendran & Vedaldi, CVPR, 2015. The code here is desigined for inverting a VGG16-like network trained on CIFAR10 dataset, but you can modify slightely to work for other network architectures and datasets.

Additional features

In addition to the inversion procedures described in the original paper, this repository contains some additional features:

• "Change-of-variable" technique, as introduced in Carlini & Wagner, 2017. Instead of optimizing the raw input pixel values x, this technique optimizes another variable w, where x_i = 0.5 * (tanh(w_i) + 1), such that 0 <= x_i <= 1 is guaranteed.
• Automated hyperparameter tuning powered by Optuna [Akiba+, KDD, 2019]

Tested environment

• Python 3.6.6
• Pytorch 1.0.1

Usage

1. run ff_training.py to train the feed-forward network on CIFAR10 (input --> target).
2. run get_rpr.py to obtain the internal representation at the specified layer (input --> hidden).
3. run invert_MV.py to invert the internal representation back to the input space (hidden --> input).

Incorporate change-of-variable technique: python invert_MV.py --variable_change
Perform hyperparameter tuning: python invert_MV.py --tune_hyperparams

References

Understanding Deep Image Representations by Inverting Them https://arxiv.org/abs/1412.0035

Towards Evaluating the Robustness of Neural Networks https://arxiv.org/abs/1608.04644

Optuna: A Next-generation Hyperparameter Optimization Framework https://arxiv.org/abs/1907.10902