Exercise: Aerial Photos of Hurricane Damage

ACTL3143 & ACTL5111 Deep Learning for Actuaries

Your task is to classify, using a convolutional neural network, whether an image is of a damaged or an undamaged property from aerial photographs following a hurricane. The dataset is contained in hurricane.zip.

Hurricane (Source: Zurich Insurance Group)

I highly recommend you train on a GPU for this task. For the model I fitted, for just one epoch of training it took 13 seconds on Colab’s GPU compared to 13 minutes on Colab’s CPU (& 2 mins on my PC’s decent CPU).

The data

Download the data first, either manually or using the wget and unzip trick from the lectures.

# TODO

Convolutional neural network

Train a convolutional neural network to solve this task.

# TODO

Data augmentation

Add in some data augmentation (that is, get Keras to make fake training data by rotating & zooming the original training images).

# TODO

Hyperparameter tuning

The second part is to use keras-tuner to search for at least one optimal hyperparameter (e.g. find the optimal number of filters in your first convolutional layer, and/or find the optimal number of layers to use before or after flattening the inputs). Don’t go overboard on the second one by searching over a large set of hyperparameters (unless you want to stare at progress bars for a week).

# TODO