AIPI-540 Module 1

This is the repository for the first module of the AIPI-540 course of the MEng in Artificial Intelligence at Duke University. The module is focused on applying Computer Vision techniques to solve a problem.

Description

Prior to this project there was no way to classify images of the species of snakes that exist in North Carolina. Previous efforts created big datasets of images of snakes, but they were from all over the world, not including som of the specied that are found in North Carolina. This project brings novelty by creating a new dataset of images of the 38 most common snake species that are found in North Carolina and to train a model to classify them.

Getting Started

Dependencies

This project needs the dependencies defined in "requirements.txt" as well as GPU support for faster training (optional - the saved weights in the repository can be used). Some of the dependencies are:

• Keras,
• PyTorch
• scikit-learn
• MatplotLib
• Pillow
• Streamlit

Installing

The inference code in this repository is ready to be run. There is no need to perform data sourcing or training. However, all the code that was used to produce the dataset and train the model is available in the repository.
Below are two instruction flows:

1. Source and pre-process data and train model
2. Run application in inference-only mode

Source and pre-process data and train model

1. Clone the repository
2. Install the dependencies (pip install -r requirements.txt)
3. (Optional) - Run the scripts/data_sourcing/make_data.py script to generate the images dataset.
4. (Optional) - Run the scripts/segmentation/segment_preparation.py to install Grounding Dino dependencies.
5. (Optional) - Run the scripts/segmentation/segment_process.py to create image dataset.
6. Perform transfer learning to train and save the model weights by running the scripts/training/resnet50_torch.py script.
7. Voilà! You have a trained model ready to be used!

Note: the optional steps above are not necessary if you directly download the image dataset from the links provided in the "Image Dataset Links" section below.

Image Dataset Links

• Original Raw Image Dataset
• Segmented Image Dataset (optimized)

Run application for inference

1. Optional (if ran previous flow) - Clone the repository
2. Optional (if ran previous flow) - Install the dependencies (pip install -r requirements.txt)
3. Optional (if ran previous flow) - Run the scripts/segmentation/segment_preparation.py to install Grounding Dino dependencies.
4. Run streamlit run main.py script to start the Streamlit application.

Authors

• João Cardoso
• Leslie Dees
• Zihan Xing

Acknowledgments

Inspiration, code snippets, etc.

• Grounding Dino