Avoiding using a pre-prepared / external model for classifying images using keras in R

Question

I have the following code. The dataset can be downloaded here or here . A dataset contains images classified as cator dog.

The purpose of this code is to prepare data on images of cats and dogs. So, given the picture, she can say it is a cat or a dog. This is motivated by this page . Below is the full code:

library(keras)


# Organize dataset --------------------------------------------------------
options(warn = -1)

# Ths input
original_dataset_dir <- "data/kaggle_cats_dogs/original/"


# Create new organized dataset directory ----------------------------------

base_dir <- "data/kaggle_cats_dogs_small/"
dir.create(base_dir)

train_dir <- file.path(base_dir, "train")
dir.create(train_dir)

validation_dir <- file.path(base_dir, "validation")
dir.create(validation_dir)

test_dir <- file.path(base_dir, "test")
dir.create(test_dir)

train_cats_dir <- file.path(train_dir, "cats")
dir.create(train_cats_dir)

train_dogs_dir <- file.path(train_dir, "dogs")
dir.create(train_dogs_dir)

validation_cats_dir <- file.path(validation_dir, "cats")
dir.create(validation_cats_dir)

validation_dogs_dir <- file.path(validation_dir, "dogs")
dir.create(validation_dogs_dir)

test_cats_dir <- file.path(test_dir, "cats")
dir.create(test_cats_dir)

test_dogs_dir <- file.path(test_dir, "dogs")
dir.create(test_dogs_dir)

# Copying files from original dataset to newly created directory
fnames <- paste0("cat.", 1:1000, ".jpg")
dum <- file.copy(file.path(original_dataset_dir, fnames), 
          file.path(train_cats_dir)) 


fnames <- paste0("cat.", 1001:1500, ".jpg")
dum <- file.copy(file.path(original_dataset_dir, fnames), 
          file.path(validation_cats_dir))

fnames <- paste0("cat.", 1501:2000, ".jpg")
dum <- file.copy(file.path(original_dataset_dir, fnames),
          file.path(test_cats_dir))

fnames <- paste0("dog.", 1:1000, ".jpg")
dum <- file.copy(file.path(original_dataset_dir, fnames),
          file.path(train_dogs_dir))

fnames <- paste0("dog.", 1001:1500, ".jpg")
dum <- file.copy(file.path(original_dataset_dir, fnames),
          file.path(validation_dogs_dir)) 

fnames <- paste0("dog.", 1501:2000, ".jpg")
dum <- file.copy(file.path(original_dataset_dir, fnames),
          file.path(test_dogs_dir))

options(warn = -1)

# Making model ------------------------------------------------------------


conv_base <- application_vgg16(
  weights = "imagenet",
  include_top = FALSE,
  input_shape = c(150, 150, 3)
)


model <- keras_model_sequential() %>% 
  conv_base %>%
  layer_flatten() %>% 
  layer_dense(units = 256, activation = "relu") %>% 
  layer_dense(units = 1, activation = "sigmoid")

summary(model)

length(model$trainable_weights)
freeze_weights(conv_base)
length(model$trainable_weights)



# Train model -------------------------------------------------------------

train_datagen = image_data_generator(
  rescale = 1/255,
  rotation_range = 40,
  width_shift_range = 0.2,
  height_shift_range = 0.2,
  shear_range = 0.2,
  zoom_range = 0.2,
  horizontal_flip = TRUE,
  fill_mode = "nearest"
)

# Note that the validation data shouldn't be augmented!
test_datagen <- image_data_generator(rescale = 1/255)  

train_generator <- flow_images_from_directory(
  train_dir,                  # Target directory  
  train_datagen,              # Data generator
  target_size = c(150, 150),  # Resizes all images to 150 × 150
  batch_size = 20,
  class_mode = "binary"       # binary_crossentropy loss for binary labels
)

validation_generator <- flow_images_from_directory(
  validation_dir,
  test_datagen,
  target_size = c(150, 150),
  batch_size = 20,
  class_mode = "binary"
)


# Compile model -----------------------------------------------------------

model %>% compile(
  loss = "binary_crossentropy",
  optimizer = optimizer_rmsprop(lr = 2e-5),
  metrics = c("accuracy")
)


# Evaluate  ---------------------------------------------------------------

history <- model %>% fit_generator(
  train_generator,
  steps_per_epoch = 100,
  epochs = 30,
  validation_data = validation_generator,
  validation_steps = 50
)


# Plot --------------------------------------------------------------------
plot(history)

The above example requires the external model to load (developed in Connet ImageNet) a dataset with the VGG16 architecture in order to be fully functional.

conv_base <- application_vgg16(
  weights = "imagenet",
  include_top = FALSE,
  input_shape = c(150, 150, 3)
)


model <- keras_model_sequential() %>% 
  conv_base %>%
  layer_flatten() %>% 
  layer_dense(units = 256, activation = "relu") %>% 
  layer_dense(units = 1, activation = "sigmoid")

cat/dog . VGG16 .

, . , , , VGG16 , .

( ) conv_base? , ?

Answer 1

... .

@Matias Valdenegro, , .

VGG16

-, , VGG16 , freeze_weights().

VGG16 . , , , .

VGG16, weights = "imagenet" weights=None ( R "", , , NULL). VGG16.

. , , , , ..

: VGG16 . . , , , , / ( , - ).

, Conv2D + Conv2D + MaxPooling2D. 2 3 , .

python , , , , R, , . .

Answer 2

, VGG-16 , , , , . , , , , , , , - (, ), , , , , , , . , inceptionv3 . .