INPUT$\_$SHAPE $=(256,256,3)$

inputs $=$ keras.layers.Input$($shape$=$INPUT$\_$SHAPE$)$

conv$1=$ SeparableConv$2$D$(64,3,$ activation$=$'relu', padding$=$'same'$)($inputs$)$

conv$1=$ BatchNormalization$()($conv$1)$

pool$1=$ MaxPooling$2$D$($pool$\_$size$=(2,2))($conv$1)$

apool$1=$ AveragePooling$2$D$($pool$\_$size$=(2,2))($conv$1)$

am$1=$ keras.layers.Add$()([$apool$1,$ pool$1])$

out$1=$ am$1$

conv$2=$ SeparableConv$2$D$(128,3,$ activation$=$'relu', padding$=$'same'$)($out$1)$

conv$2=$ BatchNormalization$()($conv$2)$

pool$2=$ MaxPooling$2$D$($pool$\_$size$=(2,2))($conv$2)$

apool$2=$ AveragePooling$2$D$($pool$\_$size$=(2,2))($conv$2)$

am$2=$ keras.layers.Add$()([$apool$2,$ pool$2])$

out$2=$ am$2$

conv$3=$ SeparableConv$2$D$(256,3,$ activation$=$'relu', padding$=$'same'$)($out$2)$

conv$3=$ BatchNormalization$()($conv$3)$

conv$4=$ SeparableConv$2$D$(256,3,$ activation$=$'relu', padding$=$'same'$)($conv$3)$

conv$4=$ BatchNormalization$()($conv$4)$

drop$3=$ Dropout$(0\mathrm{.}3)($conv$4)$

pool$3=$ MaxPooling$2$D$($pool$\_$size$=(2,2))($drop$3)$

gpool$3=$ AveragePooling$2$D$($pool$\_$size$=(2,2))($drop$3)$

kernel$\_$initializer$=\text{'}$he$\_$uniform'$)($pconv$5)$

n$1=$ keras.layers.Add$()([$gpool$3,$ pool$3])$

n$2=$ keras.activations.sigmoid$($n$1)$

n$2$p $=$ Multiply$()([$n$2,$ pool$3])$

n$2$gp $=$ Multiply$()([$n$2,$ gpool$3])$

out $=$ Add$()([$n$2$p$,$ n$2$gp$])$

def squeeze$\_$excite$\_$block$($tensor$,$ ratio$=1)$:

init $=$ tensor

channel$\_$axis $=1$ if K$.$image$\_$data$\_$format$()==$ "channels$\_$first" else $-1$

filters $=$ init.shape$[$channel$\_$axis$]$

se$\_$shape $=(1,1,$ filters$)$

se $=$ GlobalAveragePooling$2$D$()($init$)$

se $=$ Reshape$($se$\_$shape$)($se$)$

se $=$ Dense$($filters $//$ ratio, activation$=$'relu', kernel$\_$initializer$=\text{'}$he$\_$uniform', use$\_$bias$=$False$)($se$)$

se $=$ Dense$($filters$,$ activation$=$'sigmoid', kernel$\_$initializer$=\text{'}$he$\_$uniform', use$\_$bias$=$False$)($se$)$

if K$.$image$\_$data$\_$format$()==$ 'channels$\_$first':

se $=$ Permute$((3,1,2))($se$)$

x $=$ multiply$([$init$,$ se$])$

return x

se $=$ squeeze$\_$excite$\_$block$($out$,$ ratio$=1)$

kernel$\_$initializer$=\text{'}$he$\_$uniform'$)($conv$4)$

kernel$\_$initializer$=\text{'}$he$\_$normal'$)($conv$4)$

flat $=$ Flatten$()($se$)$

dropout$1=$ Dropout$(0\mathrm{.}3)($flat$)$

output $=$ Dense$(1,$ activation$=$'sigmoid'$)($dropout$1)$

loss $=$ keras.losses.BinaryFocalCrossentropy$($apply$\_$class$\_$balancing$=$True, gamma$=2,$ from$\_$logits$=$True$)$

model $=$ Model$($inputs$=$inputs, outputs$=$output$)$

optimizer $=$ Adam$($learning$\_$rate$=0\mathrm{.}00003)$

model.compile$($optimizer$=$optimizer,

loss$=$loss,

metrics$=[\text{'}$accuracy$\text{'}])$

print$($model$.$summary$())$

callback $=$ tf$.$keras.callbacks.EarlyStopping$($monitor$=$'val$\_$loss', mode$=$'auto', verbose$=1,$ patience$=250)$

result $=$ model.fit$($X$\_$train,

y$\_$train,

batch$\_$size$=16,$

verbose$=2,$

epochs$=300,$

validation$\_$split$=0\mathrm{.}3,$ callbacks$=[$callback$])$ # $10$ percent of data for validation

acc $=$ result.history$[\text{'}$accuracy$\text{'}]$

val$\_$acc $=$ result.history$[\text{'}$val$\_$accuracy'$]$

loss $=$ result.history$[\text{'}$loss$\text{'}]$

val$\_$loss $=$ result.history$[\text{'}$val$\_$loss'$]$

epochs $=$ range$(1,$ len$($loss$)+1)$

y$\_$pred $=$ model.predict$($X$\_$test$)$

# Save the plots

save$\_$dir $=\text{'}$C:$//$MASIEL$//$RESULTS$\text{'}$ # Set your save directory here

# Data to save

# Data to save

plot$\_$data $=\{$

'history': $\{$

'epochs': list$($range$(1,$ len$($acc$)+1)),$

'accuracy': acc,

'val$\_$accuracy': val$\_$acc,

'loss': loss,

'val$\_$loss': val$\_$loss,

$\},$

'accuracy': accuracy$\_$score$($y$\_$test, y$\_$pred.round$()),$

'confusion$\_$matrix': confusion$\_$matrix$($y$\_$test, y$\_$pred.round$()),$

$\text{'}$f$1\_$score': f$1\_$score$($y$\_$test, y$\_$pred.round$()),$

$\text{'}$y$\_$test': y$\_$test.tolist$(),$ # Convert numpy array to list for pickling

$\text{'}$y$\_$pred': y$\_$pred.tolist$()$ # Convert numpy array to list for pickling

$\}$

# Save directory

save$\_$dir $=\text{'}$C:$//$MASIEL$//$RESULTS$\text{'}$

data$\_$path $=$ os$.$path.join$($save$\_$dir, 'model$\_$results.pkl$\text{'})$

# Save the data

with open$($data$\_$path, $\text{'}$wb$\text{'})$ as file:

pickle.dump$($plot$\_$data, file$)$

# Save Test Accuracy and F$1$ Score to a text file

test$\_$accuracy $=$ accuracy$\_$score$($y$\_$test, y$\_$pred.round$())$

f$1=$ f$1\_$score$($y$\_$test, y$\_$pred.round$())$

with open$($os$.$path.join$($save$\_$dir, 'test$\_$metrics.txt$\text{'}),\text{'}$w$\text{'})$ as file:

file.write$($f$\text{'}$Test Accuracy: $\{$test$\_$accuracy$\}$

$\text{'})$

file.write$($f$\text{'}$F$1$ Score: $\{$f$1\}$

$\text{'})$