Densely Connected Convolutional Network (DenseNet) is a network architecture where each layer is directly connected to every other layer in a feed-forward fashion. It's quite similar to ResNet but in contrast DenseNet concatenates outputs instead of using summation. If you need a quick introduction about how DenseNet works, please read the original paper[1]. It's well written and easy to understand.

I implemented a DenseNet in Python using Keras and TensorFlow as backend. Because of this I can't guarantee that this implementation is working well with Theano or CNTK. In the next months I will update the code to TensorFlow 2.x. Besides I will try to optimize this architecture in my own way with some modifications. You can find several implementations on GitHub.

I used this notebook to evaluate the model on fashion-MNIST with following parameters:

Feel free to try it on your own with another parameters.

• TensorFlow 2.x
• Python 3.x

Feel free to use this implementation:

import densenet
model = densenet.DenseNet(input_shape=(28,28,1), nb_classes=10, depth=10, growth_rate=25,
                          dropout_rate=0.1, bottleneck=False, compression=0.5).build_model()
model.summary()

This will build the following model:

[1] Densely Connected Convolutional Networks
[2] DenseNet - Lua implementation

Christopher Masch