Tensorflow Implementation of DenseNets

Two types of Densely Connected Convolutional Networks (DenseNets) are available:

• DenseNet - without bottleneck layers

• DenseNet-BC - with bottleneck layers

Each model can be tested on such datasets:

• Cifar10

• Cifar10+ (with data augmentation)

• Cifar100

• Cifar100+ (with data augmentation)

• ImageNet

Example run:

python train_densenet_cifar.py

There are also many other implementations - they may be useful also.

Citation:

@article{Huang2016Densely,
       author = {Huang, Gao and Liu, Zhuang and Weinberger, Kilian Q.},
       title = {Densely Connected Convolutional Networks},
       journal = {arXiv preprint arXiv:1608.06993},
       year = {2016}
}

Dependencies

• Model was tested with Python 2.7 with and without CUDA.

• Model should work as expected with TensorFlow >= 1.4.

Repo supported with requirements file - so the easiest way to install all just run pip install -r requirements.txt.