densenet_1d

This repository contains a Keras implementation of the DenseNet paper (Huang et al, "Densely Connected Convolutional Networks", CVPR 2017). This implementation will focus on use-cases where the inputs are 1D sequences.

Setup

To install densenet, simply clone this repository, and run

python setup.py install

Usage

The classifiers directory contains classifiers implemented as subclasses of keras.models.Model classes. This means that once a densenet.classifier is instantiated, it contains all of the usual methods of keras.models.Model, such as fit, predict, evaluate, summary, etc.

Here is an instantiation of the model that matches the original Huang et al. paper, except using a one-dimensional input rather than a two-dimensional input:

from densenet.classifiers.one_d import DenseNet121
model = DenseNet121(input_shape=(224, 13))
print(model.summary())

Upon running those lines, you should see an extensive summary indicating the layers in the model.

Note that the DenseNet implementations are highly customizable. For example, say you want to replace the default width-3 convolutions with width-5 ones. Simply instantiate your model as

from densenet.classifiers.one_d import DenseNet121
model = DenseNet121(input_shape=(224, 13), conv_kernel_width=5)
print(model.summary())

References

• Original paper: Huang et al. "Densely Connected Convolutional Networks", CVPR 2017.
• Another great implementation of DenseNets in Keras, although not one that uses 1D sequences. https://github.com/tdeboissiere/DeepLearningImplementations/blob/master/DenseNet/densenet.py
• An implementation of Resnet in Keras whose directory structure was loosely used for this implementation: https://github.com/broadinstitute/keras-resnet