DL-Coastal_Images_Classification

Custom Deep Learning Pipeline for Coastal Image Classification

This project provides a streamlined deep learning framework for classifying coastal images using a ResNet50-based architecture. The pipeline is modular and fully configurable, making it easy to train, evaluate, and customize experiments.

Key Features

• ResNet50 Model: A custom-implemented ResNet50 model for image classification tasks.

• Flexible Data Handling: Utilities for dataset preparation, normalization, and basic data checks.

• Configurable Experiments: All key parameters (dataset path, batch size, epochs, optimizer settings) are controlled via YAML configs.

• Training & Evaluation Scripts: Easy-to-run shell scripts for both training and evaluation.

• Pixi Environment: Ensures fully reproducible environments via pixi.toml.

---

Installation

1️⃣ Clone the repository:

git clone https://github.com/toutsos/DL-Coastal_Images_Classification.git
cd DL-Coastal_Images_Classification

2️⃣ Set up the Pixi environment:

pixi install

---

Training

Run the training script:

bash scripts/COASTresnet50.sh

This will:

• Load the dataset using the configuration in configs/COASTresnet50.yaml
• Train the ResNet50 model
• Save checkpoints and logs

---

Evaluation

To evaluate the trained model:

bash scripts/evaluation.sh

---

Modifying Hyperparameters

Edit the YAML config file in configs/ (e.g., COASTresnet50.yaml) to change:

• Dataset path
• Batch size
• Number of epochs
• Learning rate & optimizer parameters
• Image size & normalization settings

Example snippet from the YAML:

batch_size: 32
num_epochs: 50
learning_rate: 0.001
dataset_path: /path/to/your/dataset

---

Metrics & Logging

During training and evaluation, you'll see:

• Training & Validation Accuracy
• Loss Curves
• Confusion Matrix and visualizations (via plot_utils.py)

Logs and outputs are saved to the specified directory in your config.

---

🗂️ Repository Structure

├── configs/
│   └── COASTresnet50.yaml          # Configuration file
├── scripts/
│   ├── COASTresnet50.sh            # Training script
│   └── evaluation.sh               # Evaluation script
├── src/
│   ├── train.py                    # Training entry point
│   ├── evaluation.py               # Evaluation logic
│   ├── datamodules/
│   │   └── dataset.py              # Dataset loader
│   └── modules/
│       └── ResNet50Model.py        # ResNet50 architecture
├── pixi.toml                       # Environment specification
├── README.md
└── .gitignore

---

How to Extend

• New Dataset: Modify the dataset loader in datamodules/dataset.py to handle new formats.

• Model Tweaks: Customize or extend modules/ResNet50Model.py to experiment with new architectures.

• New Configs: Duplicate the YAML config and adjust as needed for new experiments.

Project dependencies:

• pytorch
• lightning
• torchmetrics
• torchvision
• matplotlib
• pandas
• numpy
• pyrootutils
• hydra-core
• pytorch-cuda
• webdataset
• scikit-learn

More detais about the versions of each package you can find in the Packages-Versions