ReefNet Training Code

This repository contains data loading utilities used in the ReefNet project for training coral reef classification models using timm, a PyTorch-based image modeling library. The provided dataset classes enable flexible, split-aware loading of coral patch images annotated with genus-level labels.

For model details, benchmarks, and data, visit the project page: https://reefnet-project.github.io/reefnet-2025/

Directory Structure

bash pytroch-models/ reefnet_dataset.py # General-purpose loader dataset_reefnet_v2.py # main data loader for classification tasks used in ReefNet code dataset_rsg.py # Al-wajh dataset loader ... # rest of the files like the original repo README.md ... (integration with timm)

Overview of Data Loaders

ReefNetDataset

• General-purpose dataset class supporting arbitrary label, path, and split column names.
• Outputs: (image_tensor, label_index)
• CSV must contain columns like patch_path, split, and your chosen label column (e.g., genus or Experimental_label).

CoralPatchCSVLoader

• Simplified loader for standard train/val/test splits using genus as the label column.
• Prints summary statistics on load: class counts, unique labels, etc.

RSGPatchCSVLoader

• Specialized loader for Al-wajh testing dataset.
• Filters samples based on a predefined list of genera (e.g., Porites, Acropora, Pocillopora).
• Use this for domain-specific generalization or Red Sea evaluation tasks.

Usage

Example (Standalone PyTorch Training)

```python from loaders.reefnet_dataset import ReefNetDataSet from torchvision import transforms

transform = transforms.Compose([ transforms.Resize((224, 224)), transforms.ToTensor() ])

dataset = ReefNetDataSet( csvfile='reefnetmetadata.csv', split='train', labelcolumn='genus', pathcolumn='patch_path', transform=transform )

image, label = dataset[0] print(image.shape, label) ```

CSV Format

All loaders expect a CSV with at least these columns: - patch_path: path to the coral patch image - genus or Experimental_label: genus-level coral label - split: one of train, val, or test

For RSGPatchCSVLoader, an additional column Our_Labels is used to filter classes.

Citing ReefNet

If you use these loaders or any ReefNet data or models in your research, please cite:

bibtex @article{battach2025reefnet, title={ReefNet: A Large-scale, Taxonomically Enriched Dataset and Benchmark for Coral Reef Classification}, author={Battach, Yahia and Felemban, Abdulwahab and Khan, Faizan Farooq and Radwan, Yousef A. and Li, Xiang and Silva, Luis and Suka, Rhonda and Gonzalez, Karla and Marchese, Fabio and Williams, Ivor D. and Jones, Burton H. and Beery, Sara and Benzoni, Francesca and Elhoseiny, Mohamed}, journal={arXiv preprint arXiv:XXXX.XXXXX}, year={2025} }

Maintainer

• Yahia Battach Creator and Maintainer

License

This code is licensed under the CC BY 4.0 License.

Questions?

Feel free to open an issue or visit our project homepage for more resources.