How would your summer season vacation’s photographs look had Edvard Munch painted them? (Maybe it’s higher to not know).
Let’s take a extra comforting instance: How would a pleasant, summarly river panorama look if painted by Katsushika Hokusai?
Type switch on photographs is just not new, however received a lift when Gatys, Ecker, and Bethge(Gatys, Ecker, and Bethge 2015) confirmed easy methods to efficiently do it with deep studying.
The principle thought is easy: Create a hybrid that could be a tradeoff between the content material picture we need to manipulate, and a type picture we need to imitate, by optimizing for maximal resemblance to each on the similar time.
If you happen to’ve learn the chapter on neural type switch from Deep Studying with R, you might acknowledge among the code snippets that observe.
Nonetheless, there is a vital distinction: This put up makes use of TensorFlow Keen Execution, permitting for an crucial method of coding that makes it simple to map ideas to code.
Identical to earlier posts on keen execution on this weblog, this can be a port of a Google Colaboratory pocket book that performs the identical job in Python.
As normal, please ensure you have the required bundle variations put in. And no want to repeat the snippets – you’ll discover the whole code among the many Keras examples.
Stipulations
The code on this put up is determined by the newest variations of a number of of the TensorFlow R packages. You may set up these packages as follows:
set up.packages(c("tensorflow", "keras", "tfdatasets"))You also needs to ensure that you might be working the very newest model of TensorFlow (v1.10), which you’ll be able to set up like so:
library(tensorflow)
install_tensorflow()There are further necessities for utilizing TensorFlow keen execution. First, we have to name tfe_enable_eager_execution() proper at the start of this system. Second, we have to use the implementation of Keras included in TensorFlow, relatively than the bottom Keras implementation.
Stipulations behind us, let’s get began!
Enter photographs
Right here is our content material picture – substitute by a picture of your personal:
# If in case you have sufficient reminiscence in your GPU, no must load the photographs
# at such small dimension.
# That is the dimensions I discovered working for a 4G GPU.
img_shape <- c(128, 128, 3)
content_path <- "isar.jpg"
content_image <- image_load(content_path, target_size = img_shape[1:2])
content_image %>%
image_to_array() %>%
`/`(., 255) %>%
as.raster() %>%
plot()
And right here’s the type mannequin, Hokusai’s The Nice Wave off Kanagawa, which you’ll be able to obtain from Wikimedia Commons:
We create a wrapper that hundreds and preprocesses the enter photographs for us.
As we might be working with VGG19, a community that has been educated on ImageNet, we have to rework our enter photographs in the identical method that was used coaching it. Later, we’ll apply the inverse transformation to our mixture picture earlier than displaying it.
load_and_preprocess_image <- perform(path) {
img <- image_load(path, target_size = img_shape[1:2]) %>%
image_to_array() %>%
k_expand_dims(axis = 1) %>%
imagenet_preprocess_input()
}
deprocess_image <- perform(x) {
x <- x[1, , ,]
# Take away zero-center by imply pixel
x[, , 1] <- x[, , 1] + 103.939
x[, , 2] <- x[, , 2] + 116.779
x[, , 3] <- x[, , 3] + 123.68
# 'BGR'->'RGB'
x <- x[, , c(3, 2, 1)]
x[x > 255] <- 255
x[x < 0] <- 0
x[] <- as.integer(x) / 255
x
}Setting the scene
We’re going to use a neural community, however we received’t be coaching it. Neural type switch is a bit unusual in that we don’t optimize the community’s weights, however again propagate the loss to the enter layer (the picture), in an effort to transfer it within the desired course.
We might be fascinated with two sorts of outputs from the community, comparable to our two objectives.
Firstly, we need to hold the mixture picture much like the content material picture, on a excessive degree. In a convnet, higher layers map to extra holistic ideas, so we’re choosing a layer excessive up within the graph to match outputs from the supply and the mixture.
Secondly, the generated picture ought to “seem like” the type picture. Type corresponds to decrease degree options like texture, shapes, strokes… So to match the mixture towards the type instance, we select a set of decrease degree conv blocks for comparability and combination the outcomes.
content_layers <- c("block5_conv2")
style_layers <- c("block1_conv1",
"block2_conv1",
"block3_conv1",
"block4_conv1",
"block5_conv1")
num_content_layers <- size(content_layers)
num_style_layers <- size(style_layers)
get_model <- perform() {
vgg <- application_vgg19(include_top = FALSE, weights = "imagenet")
vgg$trainable <- FALSE
style_outputs <- map(style_layers, perform(layer) vgg$get_layer(layer)$output)
content_outputs <- map(content_layers, perform(layer) vgg$get_layer(layer)$output)
model_outputs <- c(style_outputs, content_outputs)
keras_model(vgg$enter, model_outputs)
}Losses
When optimizing the enter picture, we are going to take into account three kinds of losses. Firstly, the content material loss: How completely different is the mixture picture from the supply? Right here, we’re utilizing the sum of the squared errors for comparability.
content_loss <- perform(content_image, goal) {
k_sum(k_square(goal - content_image))
}Our second concern is having the types match as carefully as attainable. Type is usually operationalized because the Gram matrix of flattened characteristic maps in a layer. We thus assume that type is said to how maps in a layer correlate with different.
We due to this fact compute the Gram matrices of the layers we’re fascinated with (outlined above), for the supply picture in addition to the optimization candidate, and evaluate them, once more utilizing the sum of squared errors.
gram_matrix <- perform(x) {
options <- k_batch_flatten(k_permute_dimensions(x, c(3, 1, 2)))
gram <- k_dot(options, k_transpose(options))
gram
}
style_loss <- perform(gram_target, mixture) {
gram_comb <- gram_matrix(mixture)
k_sum(k_square(gram_target - gram_comb)) /
(4 * (img_shape[3] ^ 2) * (img_shape[1] * img_shape[2]) ^ 2)
}Thirdly, we don’t need the mixture picture to look overly pixelated, thus we’re including in a regularization element, the entire variation within the picture:
total_variation_loss <- perform(picture) {
y_ij <- picture[1:(img_shape[1] - 1L), 1:(img_shape[2] - 1L),]
y_i1j <- picture[2:(img_shape[1]), 1:(img_shape[2] - 1L),]
y_ij1 <- picture[1:(img_shape[1] - 1L), 2:(img_shape[2]),]
a <- k_square(y_ij - y_i1j)
b <- k_square(y_ij - y_ij1)
k_sum(k_pow(a + b, 1.25))
}The tough factor is easy methods to mix these losses. We’ve reached acceptable outcomes with the next weightings, however be happy to mess around as you see match:
content_weight <- 100
style_weight <- 0.8
total_variation_weight <- 0.01Get mannequin outputs for the content material and elegance photographs
We want the mannequin’s output for the content material and elegance photographs, however right here it suffices to do that simply as soon as.
We concatenate each photographs alongside the batch dimension, cross that enter to the mannequin, and get again an inventory of outputs, the place each aspect of the listing is a 4-d tensor. For the type picture, we’re within the type outputs at batch place 1, whereas for the content material picture, we’d like the content material output at batch place 2.
Within the beneath feedback, please word that the sizes of dimensions 2 and three will differ for those who’re loading photographs at a special dimension.
get_feature_representations <-
perform(mannequin, content_path, style_path) {
# dim == (1, 128, 128, 3)
style_image <-
load_and_process_image(style_path) %>% k_cast("float32")
# dim == (1, 128, 128, 3)
content_image <-
load_and_process_image(content_path) %>% k_cast("float32")
# dim == (2, 128, 128, 3)
stack_images <- k_concatenate(listing(style_image, content_image), axis = 1)
# size(model_outputs) == 6
# dim(model_outputs[[1]]) = (2, 128, 128, 64)
# dim(model_outputs[[6]]) = (2, 8, 8, 512)
model_outputs <- mannequin(stack_images)
style_features <-
model_outputs[1:num_style_layers] %>%
map(perform(batch) batch[1, , , ])
content_features <-
model_outputs[(num_style_layers + 1):(num_style_layers + num_content_layers)] %>%
map(perform(batch) batch[2, , , ])
listing(style_features, content_features)
}Computing the losses
On each iteration, we have to cross the mixture picture by way of the mannequin, get hold of the type and content material outputs, and compute the losses. Once more, the code is extensively commented with tensor sizes for straightforward verification, however please take into account that the precise numbers presuppose you’re working with 128×128 photographs.
compute_loss <-
perform(mannequin, loss_weights, init_image, gram_style_features, content_features) {
c(style_weight, content_weight) %<-% loss_weights
model_outputs <- mannequin(init_image)
style_output_features <- model_outputs[1:num_style_layers]
content_output_features <-
model_outputs[(num_style_layers + 1):(num_style_layers + num_content_layers)]
# type loss
weight_per_style_layer <- 1 / num_style_layers
style_score <- 0
# dim(style_zip[[5]][[1]]) == (512, 512)
style_zip <- transpose(listing(gram_style_features, style_output_features))
for (l in 1:size(style_zip)) {
# for l == 1:
# dim(target_style) == (64, 64)
# dim(comb_style) == (1, 128, 128, 64)
c(target_style, comb_style) %<-% style_zip[[l]]
style_score <- style_score + weight_per_style_layer *
style_loss(target_style, comb_style[1, , , ])
}
# content material loss
weight_per_content_layer <- 1 / num_content_layers
content_score <- 0
content_zip <- transpose(listing(content_features, content_output_features))
for (l in 1:size(content_zip)) {
# dim(comb_content) == (1, 8, 8, 512)
# dim(target_content) == (8, 8, 512)
c(target_content, comb_content) %<-% content_zip[[l]]
content_score <- content_score + weight_per_content_layer *
content_loss(comb_content[1, , , ], target_content)
}
# complete variation loss
variation_loss <- total_variation_loss(init_image[1, , ,])
style_score <- style_score * style_weight
content_score <- content_score * content_weight
variation_score <- variation_loss * total_variation_weight
loss <- style_score + content_score + variation_score
listing(loss, style_score, content_score, variation_score)
}Computing the gradients
As quickly as now we have the losses, acquiring the gradients of the general loss with respect to the enter picture is only a matter of calling tape$gradient on the GradientTape. Word that the nested name to compute_loss, and thus the decision of the mannequin on our mixture picture, occurs contained in the GradientTape context.
compute_grads <-
perform(mannequin, loss_weights, init_image, gram_style_features, content_features) {
with(tf$GradientTape() %as% tape, {
scores <-
compute_loss(mannequin,
loss_weights,
init_image,
gram_style_features,
content_features)
})
total_loss <- scores[[1]]
listing(tape$gradient(total_loss, init_image), scores)
}Coaching part
Now it’s time to coach! Whereas the pure continuation of this sentence would have been “… the mannequin,” the mannequin we’re coaching right here is just not VGG19 (that one we’re simply utilizing as a device), however a minimal setup of simply:
- a
Variablethat holds our to-be-optimized picture - the loss capabilities we outlined above
- an optimizer that may apply the calculated gradients to the picture variable (
tf$practice$AdamOptimizer)
Beneath, we get the type options (of the type picture) and the content material characteristic (of the content material picture) simply as soon as, then iterate over the optimization course of, saving the output each 100 iterations.
In distinction to the unique article and the Deep Studying with R e-book, however following the Google pocket book as an alternative, we’re not utilizing L-BFGS for optimization, however Adam, as our aim right here is to supply a concise introduction to keen execution.
Nonetheless, you could possibly plug in one other optimization methodology for those who needed, changing
optimizer$apply_gradients(listing(tuple(grads, init_image)))
by an algorithm of your selection (and naturally, assigning the results of the optimization to the Variable holding the picture).
run_style_transfer <- perform(content_path, style_path) {
mannequin <- get_model()
stroll(mannequin$layers, perform(layer) layer$trainable = FALSE)
c(style_features, content_features) %<-%
get_feature_representations(mannequin, content_path, style_path)
# dim(gram_style_features[[1]]) == (64, 64)
gram_style_features <- map(style_features, perform(characteristic) gram_matrix(characteristic))
init_image <- load_and_process_image(content_path)
init_image <- tf$contrib$keen$Variable(init_image, dtype = "float32")
optimizer <- tf$practice$AdamOptimizer(learning_rate = 1,
beta1 = 0.99,
epsilon = 1e-1)
c(best_loss, best_image) %<-% listing(Inf, NULL)
loss_weights <- listing(style_weight, content_weight)
start_time <- Sys.time()
global_start <- Sys.time()
norm_means <- c(103.939, 116.779, 123.68)
min_vals <- -norm_means
max_vals <- 255 - norm_means
for (i in seq_len(num_iterations)) {
# dim(grads) == (1, 128, 128, 3)
c(grads, all_losses) %<-% compute_grads(mannequin,
loss_weights,
init_image,
gram_style_features,
content_features)
c(loss, style_score, content_score, variation_score) %<-% all_losses
optimizer$apply_gradients(listing(tuple(grads, init_image)))
clipped <- tf$clip_by_value(init_image, min_vals, max_vals)
init_image$assign(clipped)
end_time <- Sys.time()
if (k_cast_to_floatx(loss) < best_loss) {
best_loss <- k_cast_to_floatx(loss)
best_image <- init_image
}
if (i %% 50 == 0) {
glue("Iteration: {i}") %>% print()
glue(
"Whole loss: {k_cast_to_floatx(loss)},
type loss: {k_cast_to_floatx(style_score)},
content material loss: {k_cast_to_floatx(content_score)},
complete variation loss: {k_cast_to_floatx(variation_score)},
time for 1 iteration: {(Sys.time() - start_time) %>% spherical(2)}"
) %>% print()
if (i %% 100 == 0) {
png(paste0("style_epoch_", i, ".png"))
plot_image <- best_image$numpy()
plot_image <- deprocess_image(plot_image)
plot(as.raster(plot_image), foremost = glue("Iteration {i}"))
dev.off()
}
}
}
glue("Whole time: {Sys.time() - global_start} seconds") %>% print()
listing(best_image, best_loss)
}Able to run
Now, we’re prepared to begin the method:
c(best_image, best_loss) %<-% run_style_transfer(content_path, style_path)In our case, outcomes didn’t change a lot after ~ iteration 1000, and that is how our river panorama was wanting:
… positively extra inviting than had it been painted by Edvard Munch!
Conclusion
With neural type switch, some fiddling round could also be wanted till you get the consequence you need. However as our instance exhibits, this doesn’t imply the code needs to be sophisticated. Moreover to being simple to understand, keen execution additionally allows you to add debugging output, and step by way of the code line-by-line to examine on tensor shapes.
Till subsequent time in our keen execution sequence!
