FadelMamar/hierarchicalclassification

Master thesis work: Exploring hierarchical label structure for better classification of drivers of tropical deforestation

hierarchicalclassification

The current directory provides the code necessary to reproduce the results from my Master Thesis report. For any addtional question, send me an email: fadel.seydou@gmail.com.

Installing dependencies

• Open a terminal on Linux or a Anaconda command prompt on Windows
• Clone this repository
• Move to the repository with cd ./hierarchicalclassification
• Create a virtual environment
  • python -m venv .venv
• Activate a virtual environment
  • For Linux: source ../.venv/bin/activate
  • For Windows: .venv\bin\activate.bat
• Install dependencies
  • pip install -r requirements.txt

Project workflow

• Open the jupyter notebook: ./src/data_exploration.ipynb
• It will explain the general workflow and hypothesis behind the data prepration.

Data preparation

• Download the following files and move them to ./data:
  • brasil_coverage_2018_labelDist_2.csv
  • brasil_coverage_2018.tif
• Move to directory ./src
• Run python datapreparation.py to create needed files for downloading data
• Initialize to Google earth engine API: more here
• Run bash download_data.sh

Training

• In the terminal, move to ./src
• Create a free account on weight&biases to log training metrics
• Initialize weight&biases in this working directory as follows: wandb login and provide API key.
• Open train.sh and update the parameters. All parameters are explained in args.py
• Run bash train.sh