An undercomplete autoencoder is built as follows:

$```$

input$\_$size $=784$

hidden$\_$size $=128$

code$\_$size $=32$

input$\_$img $=$ lnput$($shape$=($input$\_$size,$))$

hidden$\_1=$ Dense$($hidden$\_$size, activation$=$'relu'$)($input$\_$img$)$

code $=$ Dense$($code$\_$size, activation$=$'relu'$)($hidden$\_1)$

hidden$\_2=$ Dense$($hidden$\_$size, activation$=$'relu'$)($code$)$

output$\_$img $=$ Dense$($input$\_$size, activation$=$'sigmoid'$)($hidden$\_2)$

autoencoder $=$ Model$($input$\_$img, output$\_$img$)$

$```$

In the training phase, the weights and the bias parameters of the autoencoder are optimized via a gradient descent to minimize the loss function between the reconstructed output and the original data. The total number of parameters $($weight$+$bias$)$ in this autoencoder is:

A$)201776$

B$)210752$

C$)208896$

D$)209968$