Unstructured Data Analysis Final Project

Imperial College London - Machine Learning and Data Science

Problem Statement

• Train a Convolutional Neural Network classifier for fire image classification.
• Investigate the effect of the preprocessing, hidden and output layers of the trained CNN model on an example image.
• Evaluate the Accuracy, Precision and Recall of the model on Validation, Testing and Out-of-Sample images.

Repository Contents

• ~/data/
  • FLAME_Dataset_subset/: A 100MB subset of the 1.46GB FLAME Dataset used for model training in sample testing
  • create_data_subset.ipynb: Python script used to create the subset of the larger FLAME dataset
  • Kaggle FIRE dataset/: A dataset used for out of sample testing
• ~/figures/: Figures exported and saved throughout the analysis, and used in the main report.
• ~/latex_report/
  • UDA_FinalProject_Batek.pdf: The PDF report to be submitted on Coursera.
  • UDA_FinalProject_Batek.tex: LaTeX script used to compile the PDF report.
• ~/models/:
  • dcnn_model_checkpoints/: Directory containing model copies saved after each epoch during training.
  • dcnn_model_training.hist: A Python Dictionary containing the training metrics (TP, FP, TN, FN, Accuracy, Precision, Recall, AUC for Training and Validation) for each epoch during training.
  • dcnn_model.keras: A trained Xception Classifier model, trained using the entire 1 GB FLAME Training set
  • dcnn_model_simple.keras: A trained Xception Classifier model, trained using only the training data from ~/data/FLAME_Dataset_subset/
• ~/reference_material/: Directory contained reference article PDFs and the references.bib file for the assignment.
• ~/scripts/: Directory containing the assignment script UDA_FinalProject_Batek.ipynb

Technical Specifications

Hardware

The project was compiled exclusively on my Miscrosoft Surface Laptop 4, which has the following specifications:

Throughout the project, I relied on CPU computing only, as I did not have access to GPU resources.

Dependencies

Please consult the environment.yml, requirement.txt and spec-file.txt for a specific list detailing the Python environment dependencies. The key dependencies are:

• Python v3.7.12
• Tensorflow v2.3.0
• Keras v2.4

The list above is not exhaustive - please refer to the aforementioned files if necessary. To replicate the python environment, execute:

conda create --name <env_name> --file spec-file.txt

with conda or

pip install -r requirements.txt

after creating and activating a virtual environment with virtualenv

Instructions

The Python Notebook containing the code for the assignment is located in ~/scripts/UDA_FinalProject_Batek.ipynb. This notebook was written so that each section separated by markdown headings can be run independently. However, it can also be run sequentially all at once.

Before doing so, be sure to chance the dataset root_path variables throughout the notebook:

Runtime for the model training section of the script was 5 hours and 35 minutes with the hardware specifications above.