from matplotlib.offsetbox import OffsetImage, AnnotationBbox

np$.$random.seed$(0)$

plt$.$figure$($figsize$=(40,40))$

# Scatter plot to help with positioning images

plt$.$scatter$($tsne$\_$results$[$:$,0],$ tsne$\_$results$[$:$,1],$ alpha$=0\mathrm{.}5)$

# Loop over each image in the subset and plot at the corresponding t$-$SNE position

for i in range$($subset$\_$size$)$:

$$ # Get the image corresponding to this t$-$SNE point

$$ image $=$ train$\_$images$[$i$].$reshape$(32,32,3)$

$$ # Create an OffsetImage object

$$ imagebox $=$ OffsetImage$($image$,$ zoom$=0\mathrm{.}7)$

$$ # Create the annotation box with the image

$$ ab $=$ AnnotationBbox$($imagebox$,($tsne$\_$results$[$i$,0],$ tsne$\_$results$[$i$,1]),$ frameon$=$False$)$

$$ # Add it to the plot

$$ plt$.$gca$().$add$\_$artist$($ab

# Title and labels

plt$.$title$(\text{'}$t$-$SNE Visualization with Images'$)$

plt$.$xlabel$(\text{'}$t$-$SNE Component $1\text{'})$

plt$.$ylabel$(\text{'}$t$-$SNE Component $2\text{'})$

# Show the plot

plt$.$show$()$Q: Based on the last part, how many clusters do you think is good for kMeans? Why?

T: Set up KMeans with some rasonable k based on the previous part $($no worries: there is no single best answer$).$

Tip: to speed things up$,$ import and use MiniBatchKMeans instead of KMeans. Browse the documentation to know how it differs.

$[]$

#$...$

T: Fit kMeans with your data.

$[]$

#$...$

Plot tSNE embedding using clusters' labels

Let's see to what extent the kMeans clusters resemble the structure of the tSNE output.

T: Plot the tSNE embedding again $--$ but this time assign colors corresponding to the kMeans cluster of each image.

Q: Can you see significant groups of points with the same color $($label$)?($If not, something is wrong.$)$ How many do you see, roughly?

$[]$

#$...$

T: Repeat the plot above but define the color of each point as the mean color of the images in the cluster to which the image belongs to$.$

Hint: You should see some blue and orange parts. Also some almost white and quite dark parts? If yes $--$ good.

If you don't see them $--$ something is likely wrong. Maybe too few iterations? If everything is gray, something is very wrong $--$ maybe way too few clusters $($k$).$ Tune the parameters until happy. Do you see why we wanted to use the faster, approximate version of k$-$means? Data analysis is often done iteratively$/$interactively $--$ so efficient algorithms save your time.

$[]$

#$...$

If you're not satisfied with the quality you can tune the parameters some more.

T: If everything looks acceptable, rerun kMeans on the full dataset $--$ something which we couldn't realistically do with tSNE!

$[]$

#$...$

Let's veriify if the clusters we got on the entire dataset are reasonable.

T: For each cluster center, plot, say, $10$ images which are closest in the sense of the Euclidean metric to it$.$

Q: Looks good? Or maybe you see sometihng suspicious?

For example: if any cluster center look like a single image in the dataset, you likely chose too many clusters!

$[]$

#$...$