PlaidML_Keras_R_layer_output_test_5_4.R

# tensorflow backend
#  original: https://github.com/jjallaire/deep-learning-with-r-notebooks/blob/master/notebooks/5.4-visualizing-what-convnets-learn.Rmd
library(keras)

model <- application_vgg16(
  weights = "imagenet", 
  include_top = FALSE
)

layer_name <- "block3_conv1"
filter_index <- 1

summary(model)

layer_output <- get_layer(model, layer_name)$output

loss <- k_mean(layer_output[,,,filter_index])

grads <- k_gradients(loss, model$input)[[1]] 

grads <- grads / (k_sqrt(k_mean(k_square(grads))) + 1e-5)


# PlaidML-Keras ver.
library(keras)
use_backend(backend = "plaidml")

model <- application_vgg16(
  weights = "imagenet", 
  include_top = FALSE
)

layer_name <- "block3_conv1"
filter_index <- 1

summary(model)

layer_output <- get_layer(model, layer_name)$output

loss <- k_mean(layer_output)

grads <- k_l2_normalize(k_gradients(loss, model$input)[[1]])

# grads <- grads / (k_sqrt(k_mean(k_square(grads))) + 1e-5) cannot work


# generate_pattern() function is also changed as below

generate_pattern <- function(layer_name, filter_index, size = 150) {
  
  # Build a loss function that maximizes the activation
  # of the nth filter of the layer considered.
  layer_output <- model$get_layer(layer_name)$output
  loss <- k_mean(layer_output) 
  
  # Compute the gradient of the input picture wrt this loss
  grads <- k_l2_normalize(k_gradients(loss, model$input)[[1]])
    
  # This function returns the loss and grads given the input picture
  iterate <- k_function(list(model$input), list(loss, grads))
  
  # We start from a gray image with some noise
  input_img_data <- 
    array(runif(size * size * 3), dim = c(1, size, size, 3)) * 20 + 128
  
  # Run gradient ascent for 40 steps
  step <- 1
  for (i in 1:40) {
    c(loss_value, grads_value) %<-% iterate(list(input_img_data))
    input_img_data <- input_img_data + (grads_value * step) 
  }
  
  img <- input_img_data[1,,,]
  deprocess_image(img) 
}

.pngは出力されたがオリジナルと違って全部均一なノイズだからうまく言ってるのかは分からない。とりあえず走った。

k_l2_normalize()でnormalizeとかどうだろうか。結果は変わらなかったが…。