Reimplementation of DenseNet on Image Recognition

Densely Connected Convolutional Networks(DenseNet), won the Best Paper Award on CVPR 2017.

This is an (re-)implementation of DenseNet in TensorFlow for image recognition tasks. The (re-)implementation is based on official Torch DenseNet with Tensorflow Slim.

In the paper, DenseNet-264 seems to be a typo, since there is no way the number of layers to be an even number. See DenseNet-121, 169, 201, if you add up, which is 1 + 6 x 2 + 1 + 12 x 2 + 1 + 64 x 2 + 1 + 48 x 2 + 1 + 1 = 265.

• DenseNet-B, DenseNet-C, DenseNet-BC
• DenseNet-121, DenseNet-169, DenseNet-201, DenseNet-265
• Training on CIFAR
• Training on SVHN
• Training on ImageNet

python 3.5
tensorflow 1.4
CUDA  8.0
cuDNN 6.0

python 3.5
tensorflow 1.2
CUDA  8.0
cuDNN 5.1

pip3 install -r requirements.txt