revnet-public

Code based on the paper and code The Reversible Residual Network: Backpropagation without Storing Activations. [arxiv]

Installation

Customize paths first in setup.sh (data folder, model save folder, etc.).

git clone git://github.com/renmengye/revnet-public.gitcd revnet-public# Change paths in setup.sh# It also provides options to download CIFAR and ImageNet data. (ImageNet# experiments require dataset in tfrecord format)../setup.sh

CIFAR-10/100

./run_cifar_train.py --dataset [DATASET] --model [MODEL]

Available values for DATASET are cifar-10 and cifar-100. Available values for MODEL are resnet-32/110/164 and revnet-38/110/164.

ImageNet

# Run synchronous SGD training on 4 GPUs.
./run_imagenet_train.py --model [MODEL]

# Evaluate a trained model. Launch this on a separate GPU. 
./run_imagenet_eval.py --id [EXPERIMENT ID]

Available values for MODEL are resnet-50/101 and revnet-54/104.

Provided Model Configs

See resnet/configs/cifar_configs.py and resnet/configs/imagenet_configs.py

Pretrained RevNet Weights

You can use our pretrained model weights for the use of other applications.

RevNet-104: 23.10% error rate on ImageNet validation set (top-1 single crop).

wget http://www.cs.toronto.edu/~mren/revnet/pretrained/revnet-104.tar.gz

Future Releases

• tf.while_loop implementation of RevNets, which achieves further memory savings.

Citation

If you use our code, please consider cite the following: Aidan N. Gomez, Mengye Ren, Raquel Urtasun, Roger B. Grosse. The Reversible Residual Network: Backpropagation without Storing Actications. CoRR, abs/1707.04585, 2017.

@article{gomez17revnet,
  author   = {Aidan N. Gomez and Mengye Ren and Raquel Urtasun and Roger B. Grosse},
  title    = {The Reversible Residual Network: Backpropagation without Storing Activations}
  journal  = {CoRR},
  volume   = {abs/1707.04585},
  year     = {2017},
  url      = {https://arxiv.org/abs/1707.04585},
  }