This project included myself and Mastoorah Faizi

Programming Language: This project was developed in Python, using the following key libraries:

• torch (PyTorch)
• torchvision
• timm (for loading the pretrained MegaDescriptor model)
• pandas, numpy, matplotlib
• scikit-learn
• wildlife_datasets (custom dataset API)
• kagglehub (for dataset download)

Code Structure: The project consists of a single Jupyter Notebook:

• 4_23_MegaDesc_Animals_Subset.ipynb This notebook includes:
  • Dataset filtering and preprocessing
  • Custom TripletDataset class definition
  • MegaDescriptor model loading
  • Triplet loss training loop
  • Evaluation using cosine similarity, balanced accuracy, and confusion matrix
  • Visualization of results

Additional files for submission:

• Wildlife CNN Report.pdf (project report)
• demo.mp4 or equivalent (recorded demonstration)
• readme.txt (this file)

How to Run the Code (with detailed steps):

1. Set up a Python environment

2. Install dependencies pip install torch torchvision timm pandas numpy matplotlib scikit-learn pip install git+https://github.com/WildlifeDatasets/wildlife-datasets@develop pip install kagglehub

3. Download and extract the dataset (run code at beginning of notebook)

4. GPU Use (Recommended)

   • This notebook is computationally intensive. For training and evaluation to complete in a reasonable time, a CUDA-enabled GPU is highly recommended.
   • In Colab, go to Runtime > Change runtime type > GPU before running.

5. Run the rest of the notebook

   • Open 4_23_MegaDesc_Animals_Subset.ipynb