TensorFlowHubwithKeras.R

##実行環境
#macOS Catalina 10.15.5 
#R version 3.6.3 (2020-02-29) -- "Holding the Windsock"
#Copyright (C) 2020 The R Foundation for Statistical Computing
#Platform: x86_64-apple-darwin15.6.0 (64-bit)
#RStudio Version 1.2.5042

#rm(list=ls())
#sudo rm -rf /Library/Frameworks/R.framework /Applications/R.app #/usr/local/bin/R /usr/local/bin/Rscript
#sudo rm -rf /Users/[YourHostName]/Library/R
#library()

#Install & Load Packages
#install.packages(c("assertthat", "memoise", "pkgload", "sessioninfo", "fs", "devtools"))
#install.packages("Rcpp") # "Select: No"
#devtools::install_github("rstudio/keras")
library(keras)

#不要
#install_keras()

##Choose to use Python (/usr/local/bin/python)
reticulate::use_python("/usr/local/bin/python", required =T)

reticulate::py_config()
#python:         /usr/local/bin/python
#libpython:      /usr/local/opt/python/Frameworks/Python.framework/Versions/3.7/lib/python3.7/config-3.7m-darwin/libpython3.7.dylib
#pythonhome:     /usr/local/opt/python/Frameworks/Python.framework/Versions/3.7:/usr/local/opt/python/Frameworks/Python.framework/Versions/3.7
#version:        3.7.7 (default, Mar 10 2020, 15:43:33)  [Clang 11.0.0 (clang-1100.0.33.17)]
#numpy:          /Users/skume/Library/Python/3.7/lib/python/site-packages/numpy
#numpy_version:  1.16.3
#tensorflow:     /usr/local/lib/python3.7/site-packages/tensorflow
#NOTE: Python version was forced by use_python function

library(tensorflow)

#Call `pkgbuild::check_build_tools(debug = TRUE)` to diagnose the problem.の場合
#options(buildtools.check = function(action) TRUE )

#devtools::install_github("rstudio/reticulate")
library(reticulate)
#devtools::install_github("rstudio/tfhub")
library(tfhub)
#devtools::install_github("rstudio/tfds")
library(tfds)
#devtools::install_github("rstudio/tfdatasets")
library(tfdatasets)

##Install Python module
#pip install "tensorflow>=2.0.0"
#pip install --upgrade tensorflow-hub
#pip install tensorflow-datasets
#pip install pillow

#TensorFlow Hub with Keras
library(keras)
library(tfhub)
#install.packages("magick")
library(magick)
#install.packages("abind")
library(abind)
#install.packages("pins")
library(pins)

##An ImageNet classifier

#Download the classifier
classifier_url <- "https://tfhub.dev/google/tf2-preview/mobilenet_v2/classification/2" 
mobilenet_layer <- layer_hub(handle = classifier_url)

#Create the Keras model
input <- layer_input(shape = c(224, 224, 3))
output <- input %>% mobilenet_layer()

model <- keras_model(input, output)
model

#Download the test image
img <- magick::image_read('https://storage.googleapis.com/download.tensorflow.org/example_images/grace_hopper.jpg') 
img

img <- img %>%
  magick::image_resize(geometry = "224x224x3!") %>% 
  magick::image_data() %>% 
  as.numeric() %>% 
  abind::abind(along = 0) # expand to batch dimension

result <- predict(model, img)
mobilenet_decode_predictions(result[,-1, drop = FALSE])
#class_name class_description    score
#1  n03763968  military_uniform 9.355025
#2  n03787032       mortarboard 5.400678
#3  n02817516          bearskin 5.297816
#4  n04350905              suit 5.200008
#5  n09835506        ballplayer 4.792094


##Simple transfer learning
## Downloads Dataset
## 花の５クラス分類: daisy, dandelion, roses, sunflowers, tulips
data_root <- pins::pin("https://storage.googleapis.com/download.tensorflow.org/example_images/flower_photos.tgz", "flower_photos")
data_root <- fs::path_dir(fs::path_dir(data_root[100])) 


image_generator <- image_data_generator(rescale = 1/255, validation_split = 0.2)
training_data <- flow_images_from_directory(
  directory = data_root, 
  generator = image_generator,
  target_size = c(224, 224), 
  subset = "training"
)

validation_data <- flow_images_from_directory(
  directory = data_root, 
  generator = image_generator,
  target_size = c(224, 224), 
  subset = "validation"
)

#Download the model
feature_extractor_url <- "https://tfhub.dev/google/tf2-preview/mobilenet_v2/feature_vector/2"
feature_extractor_layer <- layer_hub(handle = feature_extractor_url)

#Create the Model
input <- layer_input(shape = c(224, 224, 3))
output <- input %>% 
  feature_extractor_layer() %>% 
  layer_dense(units = training_data$num_classes, activation = "softmax")

model <- keras_model(input, output)
summary(model)

#Compile
model %>%
  compile(
    loss = "categorical_crossentropy",
    optimizer = "adam",
    metrics = "acc"
  )

#Train the model
model %>%
  fit_generator(
    training_data, 
    epochs = 10,
    steps_per_epoch = 20,
    validation_data = validation_data,
    verbose = 2
  )

#Evaluate the model
Result <- model %>% 
  predict_generator(training_data, steps =10)
head(Result)

#Epoch 1/10
#20/20 - 7s - loss: 0.3130 - acc: 0.8898
#20/20 - 18s - loss: 0.3130 - acc: 0.8898 - val_loss: 0.3598 - val_acc: 0.8769
#Epoch 2/10
#20/20 - 8s - loss: 0.2853 - acc: 0.9062
#20/20 - 17s - loss: 0.2853 - acc: 0.9062 - val_loss: 0.3845 - val_acc: 0.8700
#Epoch 3/10
#20/20 - 8s - loss: 0.2705 - acc: 0.9109
#20/20 - 18s - loss: 0.2705 - acc: 0.9109 - val_loss: 0.3561 - val_acc: 0.8741
#Epoch 4/10
#20/20 - 8s - loss: 0.2429 - acc: 0.9359
#20/20 - 17s - loss: 0.2429 - acc: 0.9359 - val_loss: 0.3597 - val_acc: 0.8659
#Epoch 5/10
#20/20 - 8s - loss: 0.2451 - acc: 0.9312
#20/20 - 17s - loss: 0.2451 - acc: 0.9312 - val_loss: 0.3368 - val_acc: 0.8810
#Epoch 6/10
#20/20 - 8s - loss: 0.2273 - acc: 0.9297
#20/20 - 17s - loss: 0.2273 - acc: 0.9297 - val_loss: 0.3329 - val_acc: 0.8824
#Epoch 7/10
#20/20 - 8s - loss: 0.2114 - acc: 0.9312
#20/20 - 17s - loss: 0.2114 - acc: 0.9312 - val_loss: 0.3519 - val_acc: 0.8646
#Epoch 8/10
#20/20 - 8s - loss: 0.2063 - acc: 0.9339
#20/20 - 17s - loss: 0.2063 - acc: 0.9339 - val_loss: 0.3405 - val_acc: 0.8687
#Epoch 9/10
#20/20 - 8s - loss: 0.1731 - acc: 0.9594
#20/20 - 18s - loss: 0.1731 - acc: 0.9594 - val_loss: 0.3377 - val_acc: 0.8700
#Epoch 10/10
#20/20 - 9s - loss: 0.1617 - acc: 0.9516
#20/20 - 19s - loss: 0.1617 - acc: 0.9516 - val_loss: 0.3413 - val_acc: 0.8728

#Transfer learning with tfhub
#This tutorial classifies movie reviews as positive or negative using the text of the review. 
#This is an example of binary — or two-class — classification, an important and widely applicable kind of machine learning problem.
#We’ll use the IMDB dataset that contains the text of 50,000 movie reviews from the Internet Movie Database. These are split into 25,000 reviews for training and 25,000 reviews for testing. 
#The training and testing sets are balanced, 
#meaning they contain an equal number of positive and negative reviews.
#https://tensorflow.rstudio.com/tutorials/beginners/basic-ml/tutorial_basic_text_classification_with_tfhub/

#Download the IMDB dataset
imdb <- tfds_load(
  "imdb_reviews:1.0.0", 
  split = list("train[:60%]", "train[-40%:]", "test"), 
  as_supervised = TRUE
)
summary(imdb)

#To create a dataset you can use:
first <- imdb[[1]] %>% 
  dataset_batch(1) %>%
  reticulate::as_iterator() %>% 
  reticulate::iter_next()
str(first)

#Build the model
embedding_layer <- layer_hub(handle = "https://tfhub.dev/google/tf2-preview/gnews-swivel-20dim/1")
embedding_layer(first[[1]])

#For this example we will use a pre-trained text embedding model from TensorFlow Hub called google/tf2-preview/gnews-swivel-20dim/1.

##build the  model
model <- keras_model_sequential() %>% 
  layer_hub(
    handle = "https://tfhub.dev/google/tf2-preview/gnews-swivel-20dim/1",
    input_shape = list(),
    dtype = tf$string,
    trainable = TRUE
  ) %>% 
  layer_dense(units = 16, activation = "relu") %>% 
  layer_dense(units = 1, activation = "sigmoid")

summary(model)
#Model: "sequential"
#__________________________________________________________________________
#Layer (type)                     Output Shape                 Param #     
#==========================================================================
#  keras_layer_1 (KerasLayer)       (None, 20)                   400020      
#__________________________________________________________________________
#dense (Dense)                    (None, 16)                   336         
#__________________________________________________________________________
#dense_1 (Dense)                  (None, 1)                    17          
#==========================================================================
#Total params: 400,373
#Trainable params: 400,373
#Non-trainable params: 0
#__________________________________________________________________________

#Optimizer & Loss function:
model %>%
  compile(
    optimizer = "adam",
    loss = "binary_crossentropy",
    metrics = "accuracy")

#Train the model
model %>%
  fit(
    imdb[[1]] %>% dataset_shuffle(10000) %>% dataset_batch(512),
    epochs = 20,
    validation_data = imdb[[2]] %>% dataset_batch(512),
    verbose = 2
  )

#Evaluate the model
model %>% 
  evaluate(imdb[[3]] %>% dataset_batch(512), verbose = 0)
#loss  accuracy 
#0.3220522 0.8619600