Pytorch implementation of i-RevNets.
i-RevNets define a family of fully invertible deep networks, built from a succession of homeomorphic layers.
Reference: Jörn-Henrik Jacobsen, Arnold Smeulders, Edouard Oyallon. i-RevNet: Deep Invertible Networks. International Conference on Learning Representations (ICLR), 2018. (https://iclr.cc/)
The i-RevNet and its dual. The inverse can be obtained from the forward model with minimal adaption and is an i-RevNet as well. Read the paper for theoretical background and detailed analysis of the trained models.
Requirements: Python 3, Numpy, Pytorch, Torchvision
Download the ImageNet dataset and move validation images to labeled subfolders. To do this, you can use the following script: https://raw.githubusercontent.com/soumith/imagenetloader.torch/master/valprep.sh
We provide an Imagenet pre-trained model: Download
Save it to this folder.
Train small i-RevNet on Cifar-10, takes about 5 hours and yields an accuracy of ~94.5%
$ python CIFAR_main.py --nBlocks 18 18 18 --nStrides 1 2 2 --nChannels 16 64 256
Train bijective i-RevNet on Imagenet, takes 7-10 days and yields top-1 accuracy of ~74%
$ python ILSVRC_main.py --data /path/to/ILSVRC2012/ --nBlocks 6 16 72 6 --nStrides 2 2 2 2 --nChannels 24 96 384 1536 --init_ds 2
Evaluate pre-trained model on Imagenet validation set, yields 74.018% top-1 accuracy
$ bash scripts/evaluate_ilsvrc-2012.sh
Invert output of last layer on Imagenet validation set and save example images
$ bash scripts/invert_ilsvrc-2012.sh
i-RevNets perform on par with baseline RevNet and ResNet.
| Model: | ResNet | RevNet | i-RevNet (a) | i-RevNet (b) |
|---|---|---|---|---|
| Val Top-1 Error: | 24.7 | 25.2 | 24.7 | 26.0 |
Reconstructions from ILSVRC-2012 validation set. Top row original image, bottom row reconstruction from final representation.
Contributions are very welcome.
@inproceedings{
jacobsen2018irevnet,
title={i-RevNet: Deep Invertible Networks},
author={Jörn-Henrik Jacobsen and Arnold W.M. Smeulders and Edouard Oyallon},
booktitle={International Conference on Learning Representations},
year={2018},
url={https://openreview.net/forum?id=HJsjkMb0Z},
}