We are pleased to announce that the first version of tfhub is now on CRAN. tfhub is an R interface to TensorFlow Hub - a library for the publication, discovery, and consumption of reusable parts of machine learning models. A module is a self-contained piece of a TensorFlow graph, along with its weights and assets, that can be reused across different tasks in a process known as transfer learning.

The CRAN version of tfhub can be installed with:

install.packages("tfhub")

After installing the R package you need to install the TensorFlow Hub python package. You can do it by running:

tfhub::install_tfhub()

Getting started#

The essential function of tfhub is layer_hub which works just like a keras layer but allows you to load a complete pre-trained deep learning model.

For example you can:

library(tfhub)
layer_mobilenet <- layer_hub(
  handle = "https://tfhub.dev/google/tf2-preview/mobilenet_v2/classification/4"
)

This will download the MobileNet model pre-trained on the ImageNet dataset. tfhub models are cached locally and don’t need to be downloaded the next time you use the same model.

You can now use layer_mobilenet as a usual Keras layer. For example you can define a model:

library(keras)
input <- layer_input(shape = c(224, 224, 3))
output <- layer_mobilenet(input)
model <- keras_model(input, output)
summary(model)
Model: "model"
____________________________________________________________________
Layer (type)                  Output Shape               Param #    
====================================================================
input_2 (InputLayer)          [(None, 224, 224, 3)]      0          
____________________________________________________________________
keras_layer_1 (KerasLayer)    (None, 1001)               3540265    
====================================================================
Total params: 3,540,265
Trainable params: 0
Non-trainable params: 3,540,265
____________________________________________________________________

This model can now be used to predict Imagenet labels for an image. For example, let’s see the results for the famous Grace Hopper’s photo:

Grace Hopper 
img <- image_load("images/grace-hopper.jpg", target_size = c(224,224)) %>% 
  image_to_array()
img <- img/255
dim(img) <- c(1, dim(img))
pred <- predict(model, img)
imagenet_decode_predictions(pred[,-1,drop=FALSE])[[1]]
  class_name class_description    score
1  n03763968  military_uniform 9.760404
2  n02817516          bearskin 5.922512
3  n04350905              suit 5.729345
4  n03787032       mortarboard 5.400651
5  n03929855       pickelhaube 5.008665

TensorFlow Hub also offers many other pre-trained image, text and video models. All possible models can be found on the TensorFlow hub website .

TensorFlow Hub

You can find more examples of layer_hub usage in the following articles on the TensorFlow for R website:

Usage with Recipes and the Feature Spec API#

tfhub also offers recipes steps to make it easier to use pre-trained deep learning models in your machine learning workflow.

For example, you can define a recipe that uses a pre-trained text embedding model with:

rec <- recipe(obscene ~ comment_text, data = train) %>%
  step_pretrained_text_embedding(
    comment_text,
    handle = "https://tfhub.dev/google/tf2-preview/gnews-swivel-20dim-with-oov/1"
  ) %>%
  step_bin2factor(obscene)

You can see a complete running example here .

You can also use tfhub with the new Feature Spec API implemented in tfdatasets. You can see a complete example here .

We hope our readers have fun experimenting with Hub models and/or can put them to good use. If you run into any problems, let us know by creating an issue in the tfhub repository