Counting 1,417 Big Data & Machine Learning Frameworks, Toolsets, and Examples...
Suggestion? Feedback? Tweet @stkim1

Last Commit
Apr. 23, 2017
Aug. 24, 2016

Densely Connected Convolutional Networks (DenseNets)

This repository contains the code for the paper Densely Connected Convolutional Networks.

The code is based on fb.resnet.torch.

Also, see

  1. Our Caffe Implementation
  2. Our space-efficient Torch Implementation.
  3. Our (much more) space-efficient Caffe Implementation.
  4. PyTorch Implementation (with BC structure) by Andreas Veit.
  5. PyTorch Implementation (with BC structure) by Brandon Amos.
  6. MXNet Implementation by Nicatio.
  7. MXNet Implementation (supporting ImageNet) by Xiong Lin.
  8. Tensorflow Implementation by Yixuan Li.
  9. Tensorflow Implementation by Laurent Mazare.
  10. Tensorflow Implementation (with BC structure) by Illarion Khlestov.
  11. Lasagne Implementation by Jan Schlüter.
  12. Keras Implementation by tdeboissiere.
  13. Keras Implementation by Roberto de Moura Estevão Filho.
  14. Keras Implementation (with BC structure) by Somshubra Majumdar.
  15. Chainer Implementation by Toshinori Hanya.
  16. Chainer Implementation by Yasunori Kudo.
  17. Fully Convolutional DenseNets for segmentation by Simon Jegou.

Note we didn't label all implementations which support BC structures.

If you find this helps your research, please consider citing:

  title={Densely connected convolutional networks},
  author={Huang, Gao and Liu, Zhuang and Weinberger, Kilian Q and van der Maaten, Laurens},
  journal={arXiv preprint arXiv:1608.06993},

Table of Contents

  1. Introduction
  2. Results
  3. Usage
  4. Contact


DenseNet is a network architecture where each layer is directly connected to every other layer in a feed-forward fashion (within each dense block). For each layer, the feature maps of all preceding layers are treated as separate inputs whereas its own feature maps are passed on as inputs to all subsequent layers. This connectivity pattern yields state-of-the-art accuracies on CIFAR10/100 (with or without data augmentation) and SVHN.

Figure 1: A dense block with 5 layers and growth rate 4.

densenet Figure 2: A deep DenseNet with three dense blocks.

Results on CIFAR

The table below shows the results of DenseNets on CIFAR datasets. The "+" mark at the end denotes standard data augmentation (crop after zero-padding, and horizontal flip). For a DenseNet model, L denotes its depth and k denotes its growth rate. On CIFAR-10 and CIFAR-100 (without augmentation), Dropout with 0.2 drop rate is adopted.

Method Parameters CIFAR-10 CIFAR-10+ CIFAR-100 CIFAR-100+
DenseNet (L=40, k=12) 1.0M 7.00 5.24 27.55 24.42
DenseNet (L=100, k=12) 7.0M 5.77 4.10 23.79 20.20
DenseNet (L=100, k=24) 27.2M 5.83 3.74 23.42 19.25
DenseNet-BC (L=100, k=12) 0.8M 5.92 4.51 24.15 22.27
DenseNet-BC (L=250, k=24) 15.3M 5.19 3.62 19.64 17.60
DenseNet-BC (L=190, k=40) 25.6M - 3.46 - 17.18

ImageNet and Pretrained Models (Torch and Caffe)

The Torch models are trained under the same setting as in fb.resnet.torch. The error rates shown are 224x224 1-crop test errors.

Network Top-1 error Torch Model
DenseNet-121 (k=32) 25.0 Download (64.5MB)
DenseNet-169 (k=32) 23.6 Download (114.4MB)
DenseNet-201 (k=32) 22.5 Download (161.8MB)
DenseNet-161 (k=48) 22.2 Download (230.8MB)

For ImageNet pretrained Caffe models, please see from shicai. Also, we would like to thank szq0214 for help on Caffe models.


For training on CIFAR dataset,

  1. Install Torch ResNet ( following the instructions there. To reduce memory consumption, we recommend to install the optnet package.
  2. Add the file densenet.lua to the folder models/.
  3. Change the learning rate schedule in the file train.lua: inside function learningRate(), change line 171/173 from decay = epoch >= 122 and 2 or epoch >= 81 and 1 or 0 to decay = epoch >= 225 and 2 or epoch >= 150 and 1 or 0
  4. Train a DenseNet-BC (L=100, k=12) on CIFAR-10+ using
th main.lua -netType densenet -depth 100 -dataset cifar10 -batchSize 64 -nEpochs 300 -optnet true

The file densenet-imagenet.lua is for training ImageNet models presented in the paper. The usage is very similar. Please refer to fb.resnet.torch for data preparation.


On CIFAR, by default, the growth rate k is set to 12, bottleneck transformation is used, compression rate at transiton layers is 0.5, dropout is disabled. On ImageNet, the default model is densenet-121. To experiment with other settings, please change densenet.lua accordingly (see the comments in the code).



  1. Add Imagenet results and pretrained models.
  2. Add DenseNet-BC structures.


  1. Add the code for imagenet training.


liuzhuangthu at
gh349 at
Any discussions, suggestions and questions are welcome!