This is a classification problem with $6$ classes. It comes with a predefined train and test dataset $($ignore the seg$\_$pred sub directory$).$ Your task is to develop a good

classifier for these images. To do that you should experiment with both:

$2$ pre$-$trained models that you fine$-$tune, as well as

$2$ model architectures that you train from scratch.

For all four approaches always do some hyper parameter tuning. Split the train data into $80\%$ real train and $20\%$ validation.

Submission:

Submit a report in pdf format explaining your four approaches. Make sure to include network diagrams or listings, e$.$g$.$ the output from model.summary$()$ in

Keras. For each approach produce a table that lists the validation set accucary for the different hyper parameter combination you have explored.

For the best combination from each approach also list the respective accucary on the test set. For these $4$ best combinations also supply plots $($learning curves$)$ of

the training and validation error over "time" $($have number of epochs on the x$-$axis, and both errors on the y$-$axis, e$.$g$.$ like in Figure $10\mathrm{.}8.$a$)$

For the single best model identify the worst misclassified example of each of the six classes. "Worst" is defined as having the max difference between

prob$\_$of$\_$predicted$\_$but$\_$wrong$\_$class and prob$\_$of$\_$correct$\_$class. Plot these six images, and show the two labels $($predicted and correct$)$ as a title for each image.

When you look at these errors, are there any potential explanations $($like the official label was wrong, or the image is ambiguous, or the image has some other

unusual properties $...)?$

Notes:

Use either Keras$/$Tensorflow$,$ or Pytorch, whichever you like better. Both libraries provide for many pretrained image classification models, which are usually trained

on ImageNet.