This has to be done in Python

$1.$ Load an mnist dataset with $3$ classes $($any will do$,$ generally these are handwritten digits$)$

$2.$ Partition the dataset into $80\%$ training and $20\%$ as testing

$3.$ Run $3$ different Autoencoders on the dataset with the following input dimensions for each network:

a$.50->2->50$

b$.100->50->2->50->100$

c$.200->100->50->2->50->100->200$

$4.$ Now for each dimension the bottlenet is set to 'linear' and the other dimensions are set to tanh$.$ So the $\text{'}2\text{'}$ is the bottlenet and the rest are the tanh$.$

$5.$ NEXT! And this is the tricky part!! When you've got autoencoders created with $3$ different layers, dimensions etc you have to calculate to $5$ nearest neighbor accuracy for each network.

$6.$ Report this in a table and explain the results.

$7.$ Do this $10$ times and calculate the mean accuracy and standard deviation.

$8.$ Comment on the results of the $3$ different networks

$**$NOTE$**$ Nearest Neighbor is a simple and straightforward classifier. To predict a class label of a test sample you find k$-$nearest neighbor where k is user supplied. Assign the label of the test set by majority vote. Use Euclidean distance.

Ensure code compiles, show the code and plots and tables. Explain the results.