HPWREN_Experiments

This repository is designed to post-process and analyze the classification and internal representation data produced by the LLaVA-Based Wildfire Detection pipeline. The code is provided in a single Jupyter notebook, Analyze Data.ipynb, which demonstrates how to:

• Load saved checkpoints (checkpoint.txt) and internal embeddings (internal_rep.pt)
• Reduce the dimensionality of embeddings using t-SNE
• Plot images on a scatter plot to visualize clusters and identify nearest neighbors
• Display a grid of the nearest images for a chosen coordinate in the t-SNE space
• Compute performance metrics (like an F-measure)

Repository Contents

1. Analyze Data.ipynb

   • Reads classification results from output.csv
   • Loads internal representations from internal_rep.pt
   • Performs a t-SNE (2D) dimensionality reduction for visualization of data clusters
   • Plots the data with color-coding for different classes
   • Shows a 5×4 grid of the nearest-neighbor images for a chosen point in t-SNE space (see the sample image below)
   • Computes the F-measure to evaluate classification performance

2. HPWREN_RESULTS/ (Mentioned in the Notebook)

   • checkpoint.txt: Tracks which images have been processed
   • internal_rep.pt: A PyTorch tensor of internal feature representations for analysis
   • output.csv: Classification results (e.g., file_path, predicted_label)

3. HPWREN/

   • A folder (or symlink) containing the original images used for visualization. The exact path is read from output.csv.

Getting Started

Prerequisites

You will need:

• Python 3.8+
• Jupyter Notebook or JupyterLab
• Python libraries:
  • numpy
  • pandas
  • matplotlib
  • scikit-learn (for t-SNE)
  • torch (for PyTorch tensor handling)
  • Pillow (for image handling)

If you used a conda environment or the requirements.txt from the main LLaVA-based repository, you likely have most dependencies in place already.

Folder Structure

A typical folder structure might look like this:

HPWREN_Experiments/
├── Analyze Data.ipynb
├── HPWREN_RESULTS/
│   ├── checkpoint.txt
│   ├── internal_rep.pt
│   ├── output.csv
│   └── ...
├── HPWREN/
│   ├── ...
│   └── (images or symbolic links to them)
└── README.md

Make sure:

• checkpoint.txt, internal_rep.pt, output.csv reside in HPWREN_RESULTS.
• The images referred to in output.csv match paths in the HPWREN/ folder (or whichever folder you configure).

Usage

1. Clone or download this repository:

   git clone https://github.com/your-username/HPWREN_Experiments.git
   cd HPWREN_Experiments

2. Install dependencies (if not already installed):

   pip install -r requirements.txt

   or install them manually.

3. Launch Jupyter Notebook:

   jupyter notebook

   Then open Analyze Data.ipynb in your browser.

4. Run the notebook cells in sequence:

   • The notebook loads the embeddings from internal_rep.pt.
   • It reads classification results from output.csv.
   • It performs a t-SNE and plots the data points, color-coding each class.
   • It can generate a 5×4 grid of images showing the nearest neighbors to a chosen point in t-SNE space.
   • Finally, it computes the F-measure based on ground-truth labels encoded in the filenames.

Visualization & Results

• t-SNE Scatter Plot:
  Shows each sample’s 2D embedding. Black dots could be class “0,” red dots could be class “1.”

• Nearest Neighbor Image Grid:
  The notebook allows you to pick a 2D coordinate in the t-SNE space and retrieve the closest 20 samples. They are displayed in a 5×4 grid (as shown in the sample image below), providing a quick way to visually check classification consistency.
  

• Metrics:

  • F-Measure is computed using true positives, false positives, and false negatives.

Contributing

1. Fork this repository.
2. Create a new branch for your feature or bugfix.
3. Make changes and test thoroughly.
4. Submit a pull request describing your modifications.

Acknowledgments

• HPWREN for open data and camera imagery HPWREN.
• LLaVA developers for the underlying vision-language models. LLAVA
• Matplotlib, pandas, PyTorch, scikit-learn, and the open-source community for their invaluable tools.