Introduction

In this repository you will find Jupyter Notebooks with Python code to build a model that will succesfully identify 5 cetacean families (families of whales and dolphins) with 70% accuracy. You will also find a pdf of a Canva slideshow I used when presenting this project.

Repository Structure

The code (in the Code subdirectory) for this project can be split in two batches:

• Exploratory and experimental analysis
  • Species_and_Family_Image_Arrays.ipynb
  • First_Models_Family.ipynb
  • First_Models_Species.ipynb
• Final modelling
  • Resize_Images_and_Subset_to_Directories.ipynb
  • Image_Generator_Deep_Learning.ipynb

The exploratory and experimental batch of code starts with Species_and_Family_Image_Arrays.ipynb which outputs pickles of two sampled datasets, one with balanced classes for families and one for species. These pickles are read into First_Models_Family.ipynb and First_Models_Species.ipynb to build tentative Convolutional Neural Networks which predict classifications of images to either Species or Family level with varying success.

The final modelling uses an image data generator approach to modelling in order to palliate RAM overconsumption. It starts with Resize_Images_and_Subset_to_Directories.ipynb which downsizes image dimensions and reorganises them into training and testing directories. These directories are called from Image_Generator_Deep_Learning.ipynb by image generators to feed Convolutional Neural Networks in batches of images.

Important Notes

• The data used is not available on this repository as it is from the Happy Whale Dolphin and Whale Identification Kaggle Competition in 2022.
• The directories correspond to those in a Kaggle working environment
• Python dependencies are listed at the beginning of each notebook but can also be found in dependencies.txt. The version of Python used is: .