Create, train and evaluate three extra CNN models by extending the Model$\_0.$ For each model, you are required to print the summary of model structure, train the model over $60$ epochs, plot the training and validation accuracies and losses, and evaluate the trained model with the test data. Explain what you observe.

Model$\_0$:

from tensorflowkeras.models import Sequential

from tensorflowkeras.layers import Corv$2$D$,$ MaxPooling$2$D$,$ Dropout, Flatten, Dense

# Buld the CNN model

model$\_0=$ Sequential $0$

# Layer $1$

model$\_0.$add$($Corv$2$O $,$ activetion$=$'relu', input$\_$shepe$-(144,144,3)))$

model$\_0.$add$($MaxPooling$2$O$((2,2)))$

# Layer $2$

model$\_0.$addi$($Com$2$O $,$ activation$=$"relu'$))$

model$\_0.$add$($MaxPooling$20((2,2)))$

# Layer $3$a

model$\_0.$add$($Com$2$O $,$ activation$=$'relu'$))$

# Layer $3b$

model$\_0.$add$($Corv$2$D $,$ activation$=$"relu'$))$

# Layer $3$c

model$\_0.$add$($Corv$2$O $,$ activation$=$'relu'$))$

# Layer $4$

model$\_0.$addi$($Corn$2$O $,$ activation$=$'relun'$))$

# Layer $5$

model$\_0.$add$($Com $,$ actlwation$=$'rel$))$

# Layer $6$

model$\_0.$add$($MaxPooling$2$D$((2,2)))$

model$\_0.$add$($Dropout$(0\mathrm{.}5))$

# Flatten layer

model$\_0.$add$($Flatten$0)$

$"$ Dense layers

model$\_0.$add$($Dense$(128,$ activation$=$'relu'$))$

model$\_0.$add$($Dense$(6,$ activation$=$'softmax'$))$

# Compile the model

model$\_0,$compile$($optimizer$=$"adam', loss$=$'sparse$\_$categorical$\_$crossentropy, metrics$=[\text{'}$accuracy$]]$

Table $1$

i$.$Model$\_1$: Take the structure of Model$\_0,$ but quadruple the number of filters in each convolutional layer $($so the first layer has $8$ filters and the last convolutional layer has $32$ filters$).$ Leave the dense layers as they are. Save your model as model$\_1.$h$5.$

ii$.$Model$\_2$: Based on what you have learnt from residual connections:

modify Model$\_0$ to add a residual link from the output of the layer marked $($a$)$ in Table $1$ to the output of the layer marked $($b$)($i$.$e$.$ skipping the convolution layers with output shapes $(35,35,4))$

add another residual link from the output of the layer marked $($c$)$ to the output of the layer marked $($d$)($i$.$e$.$ skipping the final convolution layer$)$

Save your model as model$\_2.$h$5.$

iii.Model$\_3$: Based on what you have learnt from data augmentation, apply two data augmentation methods $($e$.$g$.$ normalisation, horizontal translation, or rotation$)$ to generate augmented training data. Create Model$\_3$ using the same structure as Model$\_1$ above. Train Model$\_3$ with the augmented training data. Save your model as model$\_3.$h$5.$

iv$.$Compare the performance of the four models $($Model$\_0$ from Q$1$a$,$ and the three models in Q$1$b$)$ and provide a brief discussion on how the different mode structures affected their performance.