README.md

In dfalbel/densenet: Densely Connected Convolutional Networks in Keras

densenet

densenet implements the Densely Connected Convolutional Networks for R usning keras. This implementation is based on Somshubra python/keras implementation available here.

Installation

You can install densenet from github with:

# install.packages("devtools")
devtools::install_github("dfalbel/densenet")

Example

You can use densenet the same way you would use an application function from keras (eg. application_vgg16)

The following lines show how you would define DenseNet-40-12 to classify images for the cifar10 dataset.

library(keras)
library(densenet)

input_img <- layer_input(shape = c(32, 32, 3))
model <- application_densenet(include_top = TRUE, input_tensor = input_img, dropout_rate = 0.2)

opt <- optimizer_sgd(lr = 0.1, momentum = 0.9, nesterov = TRUE)

model %>% compile(
  optimizer = opt,
  loss = "categorical_crossentropy",
  metrics = "accuracy"
)

As much of the code to train the model is for preprocressing the images, we are ommiting a lot of code. You can see the full code in the packages vignette using vignette("cifar10-DenseNet-40-12"). The model takes ~125s per epoch on a high-end GPU (Nvidia GeForce 1080 Ti). Final accuracy on test set was 0.9351 versus 0.9300 reported on the paper.

dfalbel/densenet documentation built on May 31, 2019, 11:49 p.m.