Wildfire Prediction

Link to the Kaggle Challenge. Deadline: 21st of February 2025.

Additional rule: we are not allowed to use the dataset/train dataset labels.

The "valid" validation dataset was split into a training and testing set, defined in train.csv (80%) and test.csv (20%).

Dataset	Fire	No Fire	Fire %	No Fire %
Train	2811	2266	55.37%	44.63%
Test	669	554	54.70%	45.30%

Quickstart

Install the necessary python dependencies:

pip install -r requirements.txt

Download the dataset using the following command:

kaggle datasets download -d abdelghaniaaba/wildfire-prediction-dataset

Then unpack it into a dataset/ directory:

unzip -q wildfire-prediction-dataset.zip -d dataset

CLI Commands

The following cli commands are available. Run python main.py --help for more information.

Command	Description
test	Test a classifier model
train	Train a classifier model
train-fixmatch	Train a classifier model using FixMatch
train-mean-teacher	Train a classifier model using Mean Teacher
logs	Plot classifier training logs

Testing a model

Make sure that you have the relevant model checkpoints.
To test a FixMatch model, simply use the same command with a FixMatch checkpoint.

python main.py test --classifier resnext --checkpoints model.pth

Training a model

python main.py train --classifier resnext --epochs 30 --learning-rate 0.0001 --batch-size 8

With FixMatch :

python main.py train-fixmatch --classifier resnext --epochs 30 --learning-rate 0.00005 --batch-size 8 --threshold 0.999

With Mean Teacher :

python main.py train-mean-teacher --classifier resnext --epochs 30 --learning-rate 0.0001 --batch-size 32 --teacher_student_loss mse_scaled --temperature 0.7

Plot training logs

python main.py logs --path training_logs.jsonl

Annotating dataset with Ensemble

python main.py annotate --dataset_path train_unlabeled.csv --output_path <output.csv> --classifier <classifier> --checkpoints <path_to_checkpoints_directory>