This project implements a Siamese neural network trained using triplet loss for image similarity estimation. The model learns to embed images into a feature space where similar images are close together and dissimilar images are far apart. The model is built in TensorFlow/Keras and leverages distributed training across GPUs for faster convergence.
The project includes data handling, model creation, training with freezing and fine-tuning, and detailed evaluation metrics.
✅ TensorFlow tf.data pipeline for efficient data loading.
✅ Distributed training using tf.distribute.Strategy.
✅ Modular and scalable training script.
✅ Data pipeline using triplet generators.
✅ Triplet loss function (anchor, positive, negative).
✅ Progressive training (frozen backbone → fine-tuning).
✅ Using transfer learning with ResNet50 for embeddings.
✅ Comprehensive evaluation with confusion matrices and per-class metrics.
✅ Visualization of training loss and evaluation results.
✅ Easily adaptable for your own image datasets.
Your dataset directory should be organized as follows:
root_dataset_dir/
│
├── ID_1/
│ ├── image1.jpg
│ ├── image2.jpg
│ └── ...
│
├── ID_2/
│ ├── image1.jpg
│ ├── image2.jpg
│ └── ...
│
└── ID_N/
├── image1.jpg
├── image2.jpg
└── ...
- Each subfolder corresponds to one unique identity/class (
ID_1,ID_2, …). - Each subfolder contains multiple images of that identity.
- The code selects triplets (anchor, positive, negative) based on these folders.
Make sure you have Python 3.7+ and the following packages installed:
pip install tensorflow seaborn matplotlib numpy pandas
-
Prepare your dataset: Organize your image dataset as described above.
-
Configure paths and parameters: Edit the pathData variable and other parameters inside the notebook as needed.
-
Run training: Execute the training notebook to train the Siamese network.
-
Evaluate the model: Use the provided evaluation functions and cells to compute similarity metrics and visualize results.
- SiameseNetwork.ipynb: Jupyter notebook containing the full pipeline (data loading, training, evaluation)
- Keras Example: Siamese Network
- TensorFlow Guide: Distributed Training with Keras
- TensorFlow Guide: Distributed Training Overview
- Keras Guide: Transfer Learning & Fine-tuning
MIT License