In this task, you are going to cluster MNIST dataset using a pretrained ResNet 18 model and k-means.

Load the dataset¶

Load the MNIST dataset bellow. You can use either torchvision.datasets.MNIST or sklearn.datasets.fetch_openml() or any other way to load the dataset. Note that you won't need a validation set.

Get ResNet 18¶

Instantiate ResNet 18 model (pretrained on imagenet) from torchvision's model zoo.

Remove the decider layer¶

Replace the fc layer of the resnet with an Identity layer so that we use the hidden features of the last layer.

Cluster the data¶

Use MiniBatchKMeans to cluster the MNIST dataset in 64 clusters.

Assign clusters to test set.¶

Predict cluster for test set samples.

Draw clusters¶

Draw 10 random samples per each cluster from the test set.