Description
This is a quick implementation of the DenseNet model described in the paper "Densely Connected Convolutional Networks" by Huang et al. (arXiv)
It has only been tested on the Cifar-10 dataset without data augmentation, but it should work fine on any dataset.
Getting started
The basic model is defined in DenseNet.py.
The scipt cifar10_densenet_classification.py provides an example on how to create and use the model on Cifar-10 classification.
Finally, utils.py contains a few helper functions.
Prerequisites
- Keras (>= 2) (only tested with Tensorflow backend)
- numpy (>= 1.13)
Results
Below are the results of running both DenseNet and DenseNet-BC models on Cifar-10 dataset with the same hyperparameters and optimization techniques as in the original paper.
DenseNet (L=40, k=12)
DenseNet-BC (L=100, k=12)
TODO
- Add data augmentation techniques
- Try different architectures
- Try other optimizers, eg Adam
- Try out transfer learning on ImageNet