KOF-Mamba

This repository contains the training code and dataset files for faba bean chocolate spot severity grading.

Contents

• src/kof_mamba/config.py: paths and training configuration.
• src/kof_mamba/data.py: dataset loading and image transformations.
• src/kof_mamba/model.py: KOF-Mamba model, including AFMamba, DWT, and FFT branches.
• src/kof_mamba/losses.py: ordinal-aware loss and semantic alignment loss.
• src/kof_mamba/engine.py: training loop and data loader construction.
• src/kof_mamba/evaluate.py: checkpoint loading and test report generation.
• data/origin_data/: field leaf images organized by severity grade folders 0, 1, 3, 5, 7, and 9.
• data/public_test/: public test subset with 10 images per severity grade.
• data/class/task_0_1_3_5_7_9/: train, validation, and test split files.
• data/captions.json: image-level disease attribute descriptions used for semantic-assisted training.

Usage

Install dependencies:

pip install -r requirements.txt

Run training:

python -m src.kof_mamba.train

By default, the script uses relative paths inside this repository:

• images: data/origin_data
• split files: data/class/task_0_1_3_5_7_9
• captions: data/captions.json
• outputs: outputs/training_results_kof_mamba_6class

These paths can also be overridden with environment variables:

KOF_IMAGE_DIR=/path/to/images \
KOF_TASK_DIR=/path/to/task_0_1_3_5_7_9 \
KOF_CAPTIONS_JSON=/path/to/captions.json \
KOF_RESULTS_DIR=/path/to/outputs \
python -m src.kof_mamba.train