example

##### Set directory and load extracted data
setwd("/Users/ju0/Desktop/Projects/Devices-General/DBH-Camera-Dev/dbh_images/images/")
mydata <- read.csv("DBHImageExtraction_185kVect977Elements-2023-01-18.csv")
mydata[1:5,1:8]

mydata <- cbind(paste(mydata[,2], "-", mydata[,3], sep = ""), mydata)
colnames(mydata)[1] <- "tmpSort"
mydata <- mydata[order(mydata[,2], decreasing = T), ]
dim(mydata)
mydata <- mydata[!duplicated(mydata$tmpSort), ]
mydata <- mydata[,-1]
mydata[1:5,1:8]
dim(mydata)

##### Set random seed and then split the data for testing and training.
set.seed(5)
data.nms <- mydata[!duplicated(mydata[,2]),2]
data.nms <- data.nms[sample(length(data.nms), size = length(data.nms), replace = F)]
length(data.nms)

##### Slit names list into two distinct sets of samples
data.nms.1 <- data.nms[1:round(length(data.nms)*0.5)]
data.nms.2 <- data.nms[(round(length(data.nms)*0.5)+1):(length(data.nms))]
length(data.nms.1)
length(data.nms.2)

##### Create training set data
mydata.nms.1 <- mydata[complete.cases(!match(mydata[,2], data.nms.1)), ]
dim(mydata.nms.1)
mydata.nms.1[1:3,1:8]
mydata[1:3,1:8]
dim(mydata.nms.1)
mydata.1 <- mydata.nms.1

##### Format data to run through Keras
mydata.nms.1 <- paste(mydata.1[,2], mydata.1[,3], sep = "_")
mydata.nms.1 <- cbind(mydata.nms.1, mydata.1)
mydata.nms.1[1:5,1:8]
mydata.nms.1 <- mydata.nms.1[!duplicated(mydata.nms.1[,1]), ]
rownames(mydata.nms.1) <- mydata.nms.1[,1]
mydata.nms.1 <- mydata.nms.1[,-1]
range(mydata.1[,1])
mydata.nms.1 <- mydata.nms.1[,-(2:5)]
mydata.nms.1[1:5,1:8]
set.seed(358)
mydata.nms.1 <- mydata.nms.1[sample(nrow(mydata.nms.1), size = nrow(mydata.nms.1), replace = F), ]
dim(mydata.nms.1)
plot(mydata.nms.1[1:100,1])
str(mydata.nms.1)
data.1 <- mydata.nms.1
str(data.1)
data.1[1:3,1:8]


##### Create testing/prediction data
mydata.nms.2 <- mydata[complete.cases(!match(mydata[,2], data.nms.2)), ]
dim(mydata.nms.2)
mydata.nms.2[1:3,1:8]
dim(mydata.nms.2)
mydata.2 <- mydata.nms.2

##### Format data to run through Keras
mydata.nms.2 <- paste(mydata.2[,2], mydata.2[,3], sep = "_")
mydata.nms.2 <- cbind(mydata.nms.2, mydata.2)
mydata.nms.2[1:5,1:8]
mydata.nms.2 <- mydata.nms.2[!duplicated(mydata.nms.2[,1]), ]
rownames(mydata.nms.2) <- mydata.nms.2[,1]
mydata.nms.2 <- mydata.nms.2[,-1]
range(mydata.2[,1])
mydata.nms.2 <- mydata.nms.2[,-(2:5)]
mydata.nms.2[1:5,1:8]
set.seed(13)
mydata.nms.2 <- mydata.nms.2[sample(nrow(mydata.nms.2), size = nrow(mydata.nms.2), replace = F), ]
dim(mydata.nms.2)
plot(mydata.nms.2[1:100,1])
str(mydata.nms.2)
data.2 <- mydata.nms.2
data.2[1:3,1:8]
str(data.2)

##### Check dimensions and formats of data.1 and data.2
data.1[1:3,1:8]
data.2[1:3,1:8]
dim(data.1)
dim(data.2)

pythonCNN <- rbind(data.1, data.2)
write.csv(pythonCNN, file = "mytransformerdata.csv", quote = F, col.names = F)


##### Change names back to standard names
data.1.names <- rownames(data.1)
data.1.names <- gsub("mirror_", "mirror-", data.1.names)
data.2.names <- rownames(data.2)
data.2.names <- gsub("mirror_", "mirror-", data.2.names)
rownames(data.1) <- data.1.names
rownames(data.2) <- data.2.names

##### Check that all names do not match between data sets, output of match should be all NAs
data.rw.nms.1 <- t(matrix(unlist(strsplit(rownames(data.1), split = "_")), nrow = 5))
data.rw.nms.2 <- t(matrix(unlist(strsplit(rownames(data.2), split = "_")), nrow = 5))
match(unique(data.rw.nms.1[,1]), unique(data.rw.nms.2[,1]))


################################################################################
########################### Load the required libraries ########################
################################################################################

library(keras)


##### Load the training data from 'mydata.txt'
# data <- read.table("mydata.txt", header = F)

##### Convert the data to a form that can be used with a CNN
x_train <- as.matrix(data.1[, -1])
y_train <- as.matrix(data.1[, 1])

##### Reshape the data for use with a 1D CNN
x_train <- array_reshape(x_train, c(nrow(x_train), ncol(x_train), 1))

##### Create the CNN model
model <- keras_model_sequential() %>%
  layer_conv_1d(filters = 64, kernel_size = 9, activation = "relu", input_shape = c(ncol(x_train), 1)) %>%
  layer_max_pooling_1d(pool_size = 2) %>%
  layer_flatten() %>%
  layer_dense(units = 1, activation = "sigmoid") # try 'softmax' or 'elu', not bad idea to try all activations just because learning.

##### Compile the model
model %>% compile(
  optimizer = "adam",
  loss = "binary_crossentropy",
  metrics = c("accuracy")
)

##### Train the model
history <- model %>% fit(
  x_train, y_train,
  epochs = 100,
  batch_size = 16)