nanxstats / deep-learning-recipes Goto Github PK

# Load and prepare MNIST data
mnist <- dataset_mnist()
x_train <- mnist$train$x
y_train <- mnist$train$y
x_test <- mnist$test$x
y_test <- mnist$test$y

dim(x_train) <- c(nrow(x_train), 784)
dim(x_test) <- c(nrow(x_test), 784)
x_train <- x_train / 255
x_test <- x_test / 255

ok <- which(y_train %in% c(1,2))
x_tmp <- x_train[ok,]
y_tmp <- y_train[ok]

ones <- which(y_tmp == 1)
ones <- sample(ones)
twos <- which(y_tmp == 2)
twos <- sample(twos)

tmp_one <- do.call(rbind, lapply(ones[1:5000], function(x){
  anchor <- x
  positive <- sample(setdiff(ones, x), 1)
  negative <- sample(twos, 1)
  c(anchor, positive, negative)
}))

tmp_two <- do.call(rbind, lapply(twos[1:5000], function(x){
  anchor <- x
  positive <- sample(setdiff(twos, x), 1)
  negative <- sample(ones, 1)
  c(anchor, positive, negative)
}))

triplets <- rbind(tmp_one, tmp_two)
triplets <- triplets[sample(1:nrow(triplets)), ]

# Define keras triplet model ####
loss_identity <- function(y_true, y_pred) k_mean(y_pred - 0 * y_true)

# margin-based triplet loss 
loss_margin_triplet <- function(x) {
  embed_user <- x[[1]]
  embed_item_positive <- x[[2]]
  embed_item_negative <- x[[3]]
  
  loss <- k_maximum(
    0.0,
    k_sum(embed_user * embed_item_negative, axis = -1, keepdims = TRUE) -
      k_sum(embed_user * embed_item_positive, axis = -1, keepdims = TRUE) +
      10.0
  )
  
  loss
}

# input layers for items (positive and negative)
input_item_anchor <- layer_input(shape = c(784), name = "input_item_anchor")
input_item_positive <- layer_input(shape = c(784), name = "input_item_positive")
input_item_negative <- layer_input(shape = c(784), name = "input_item_negative")
  
layer_embed_item <- layer_embedding(
  input_dim = 10000, output_dim = 2, name = "embed_item")

embed_item_anchor <- input_item_anchor %>%
  layer_embed_item() %>%
  layer_flatten()
embed_item_positive <- input_item_positive %>%
  layer_embed_item() %>%
  layer_flatten()
embed_item_negative <- input_item_negative %>%
  layer_embed_item() %>%
  layer_flatten()
  
# margin-based triplet loss is the output
loss <- list(embed_item_anchor, embed_item_positive, embed_item_negative) %>%
  layer_lambda(loss_margin_triplet, output_shape = c(2))
  
# define model inputs/outputs
model <- keras_model(
  inputs = c(input_item_anchor, input_item_positive, input_item_negative),
  outputs = loss
)

model %>% compile(loss = loss_identity, optimizer = optimizer_nadam())