$4$ CNN For Image Classification

Let's see if this works on real data! We'll attempt to classify people in the Yale Faces dataset. Since

training a CNN can be very time consuming, we'll so the following:

Resize the images to be $4040$

Just use one image for each person for training $($we will not use any validation set$).$

As a result, your training data will be a $144040$ tensor.

The CNN architecture is as follows. All hyperparameter choice are up to you:

A single $99$ convolutional kernel.

A max$-$pool layer with width $=4$ and stride $=4.$

A flattened layer

A Fully Connected Layer

A Softmax activation function

A Cross Entropy Loss objective function.

Additional Implementation Details

One$-$hot encode your targets $($beware that the first ID is not $0)$

Make sure to either z$-$score your features or divide them by $255.$