Sharath33/LETTER

Siamese Network for Few-Shot Image Recognition

Few-shot image recognition using a Siamese Network trained on Omniglot. Recognises new character classes from as little as a single example.

Results

Configuration	Accuracy
5-way 1-shot	95.10%
5-way 5-shot	97.07%
10-way 1-shot	90.05%
10-way 5-shot	94.83%

Evaluated on 145 unseen test classes (never seen during training).

Architecture

• Backbone: ResNet-18 pretrained, final FC stripped → 512-d features
• Embedding head: Linear(512→256) → BN → ReLU → Linear(256→128) → L2 norm
• Loss: Contrastive loss with margin=1.0
• Distance: Cosine similarity on unit-sphere embeddings

Project Structure

siamese-few-shot/
├── src/
│   ├── dataset.py        # SiamesePairDataset + EpisodeDataset
│   ├── model.py          # EmbeddingNet + SiameseNet
│   ├── loss.py           # ContrastiveLoss
│   ├── train.py          # Training + validation loop
│   ├── run_training.py   # Main training entry point
│   ├── eval.py           # N-way K-shot episodic evaluation
│   └── demo.py           # Gradio demo
├── checkpoints/
│   ├── best.pt
│   └── siamese_embedding.onnx
├── data/
│   └── class_split.json
├── requirements.txt
└── README.md

Quickstart

git clone https://huggingface.co/<your-username>/siamese-few-shot
cd siamese-few-shot
pip install -r requirements.txt

# Run Gradio demo
cd src && python demo.py

# Run episodic evaluation
cd src && python eval.py

# Retrain from scratch
cd src && python run_training.py

Training Details

• Dataset: Omniglot (background split, 964 classes)
• Train / val / test split: 70% / 15% / 15% of classes
• Epochs: 30
• Batch size: 32
• Optimiser: Adam lr=1e-3
• Scheduler: CosineAnnealingLR
• Augmentation: RandomCrop, HorizontalFlip, ColorJitter

Requirements

torch>=2.0
torchvision>=0.15
timm
gradio
onnx
onnxruntime-gpu
pillow
numpy
matplotlib
scikit-learn
tqdm
wandb

Demo

Upload any two handwritten character images. The model returns a cosine similarity score and a same / different class decision.

Trained on Latin, Greek, Cyrillic, Japanese, and 25 other alphabets via the Omniglot dataset. Also tested on Indian script characters (Tamil, Hindi, Telugu, Kannada, Bengali, Malayalam, Gujarati, Punjabi).