This assignment is the machine learning project for classifying images. You will have to train a machine learning model to predict the $20$ possible classes of images. The

training set contains these labels $($along with the number of images in each class$)$: Class ID Class Name Count

$0$ book $739$

$1$ bottle $715$

$2$ car $500$

$3$ cat $564$

$4$ chair $508$

$5$ computermouse $534$

$6$ cup $774$

$7$ dog $464$

$8$ flower $875$

$9$ fork $605$

$10$ glass $572$

$11$ glasses $436$

$12$ headphones $459$

$13$ knife $790$

$14$ laptop $446$

$15$ pen $954$

$16$ plate $433$

$17$ shoes $899$

$18$ spoon $625$

$19$ tree $591$ A subset of the images collected in Assignment $1,$ Exercise $1(12483$ images$)$ is provided for training

$($download$),$ the remaining images serve as hidden test set.

Note that the images have been validated and thus have numbered file names, where the labels are

stored separately in the accompanying labels.csv file. Furthermore, the images have been resized

to $100$ pixels $($either width or height, depending on the aspect ratio$)$ Your submission for the challenge server has to be a ZIP $($only zip $-$ no tar, $7$z$,$ etc.$)$ archive that

contains two files:

$$ architecture.py: this script contains your model class and the instantiated object. Note

that the name has to be exactly architecture.py$,$ because this is expected by the evaluation

script on the server. Also, the class name has to be MyCNN$,$ the variable name containing the

instance has to be model. Example:

import torch

import torch.nn as nn

class MyCNN$($nn$.$Module$)$:

def $\_\_$init$\_\_()$:

super$().\_\_$init$\_\_()$

def forward$($self$,$ input$\_$images: torch.Tensor$)->$ torch.Tensor:

model $=$ MyCNN$()$

$$ model.pth: this file contains your trained model. Again, use this exact same file name. To create this file, use torch.save$($model$.$state$\_$dict$(),$ "model.pth$").$ So in your training loop

you need to store the trained model using this code $($preferably the one with the lowest loss$).$

You can load the file again with model.load$\_$state$\_$dict$($torch$.$load$($trained$\_$model$)).$ On

the website you can see the code of the evaluation script, so you can make sure that your submission is compatible.

The training set images have been reduced to a size of either $100\backslash$times $\_$ or $\_\backslash$times $100$ pixels, depending

on the format of the image. For the evaluation on the hidden test set on the challenge server, note

that the images need to have a dimensionality of $100\backslash$times $100$ pixels. To create the data set, class

ImageDataset, along with the functions to$\_$grayscale and prepare$\_$image were used. Import them

using the file dataset.py $-$ note that ImageDataset is slightly different from the version in assignment $3($normalization and tensor conversion was added$),$ so use this one to ensure compatibility

with how the test set is generated.

Your task is to predict the labels for the images of the test set. The evaluation criterion is overall

accuracy on the test set.

The following restrictions apply:

$$ You only have $10$ valid attempts to upload a model.

$$ Invalid attempts $($e$.$g$.,$ error parsing your submission files$)$ are not counted.

$$ Your best attempt will be used as final attempt.