Counting 2,784 Big Data & Machine Learning Frameworks, Toolsets, and Examples...
Suggestion? Feedback? Tweet @stkim1


i-RevNet: Deep Invertible Networks

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. (


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.

Pytorch i-RevNet Usage

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:

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 --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 --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/

Invert output of last layer on Imagenet validation set and save example images

$ bash scripts/

Imagenet ILSVRC-2012 Results

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.


title={i-RevNet: Deep Invertible Networks},
author={Jörn-Henrik Jacobsen and Arnold W.M. Smeulders and Edouard Oyallon},
booktitle={International Conference on Learning Representations},