poomstas / kaggle_petfinder

Kaggle's PetFinder Competition where we create a CNN- and transformer-based regression model that takes an image and a boolean array as inputs.

PetFinder's Kaggle Competition

More information here.

Core Challenge

The challenge is to generate a regression model that predicts the score representing the popularity of the pet images (aka "Pawpularity"). What's a little more interesting is that the challenge also provides boolean metadata (see table below) that could help improve the regression accuracy. We will have to devise a way to incorporate both the image and the metadata to predict a continuous, numerical popularity value.

Novelty

• Use of embedding layers to...

Training, Validation, Testing Data Structure

The original dataset provides summarized .csv files for training and testing. In this work, I separate the training data into training and validation.

The helper function separate_train_val is used to separate the training data into training and validation, and store that information in the ./ data folder.

The final resulting folder is as follows: