ViT-ProtoNet: Few-Shot Learning with Vision Transformers and Prototypical Networks

A lightweight PyTorch implementation of Prototypical Networks using a ViT-Small/16 backbone for few-shot classification on standard benchmarks (Mini-ImageNet, CUB-200, CIFAR-FS, FC100). Includes both production scripts for training & evaluation and Jupyter notebooks for interactive analysis.

🏆 State-of-the-Art Few-Shot Results 🏆

---

🎯 Benchmarks at a Glance

Dataset	📈 Rank
CIFAR-FS	#1
FC100	#1
CUB-200	#4
Mini-ImageNet	#4

---

🚀 Features

• Lightweight Backbone
  Uses ViT-Small/16 (384-dim embeddings, 6 heads, 12 layers) for fast feature extraction.
• Prototypical Framework
  Episodic training with support & query sets; prototypes computed as mean CLS-token embeddings.
• Multiple Benchmarks
  Ready to run on Mini-ImageNet, CUB-200, CIFAR-FS, and FC100.
• Modular Codebase
  Clean separation between dataset loaders, model definitions, training loops, and utilities.
• Interactive Notebooks
  Explore experiments, visualize prototype distributions, and plot learning curves.

---

📝 Prerequisites

• Python 3.8 or higher
• NVIDIA GPU with CUDA 10.2+ (recommended for reasonable training speed)

---

⚙️ Installation

1. Clone the repository

   git clone https://github.com/abdulvahapmutlu/vit-protonet.git
   cd vit-protonet
   

2. Create & activate a virtual environment

   python -m venv .venv
   source .venv/bin/activate    # Linux / macOS
   .venv\Scripts\activate       # Windows
   

3. Install dependencies

   pip install --upgrade pip
   pip install -r requirements.txt
   

---

📥 Data Preparation

1. Download You must obtain the original benchmarks yourself (e.g. from official sources or Kaggle).

2. Organize Create a root folder, e.g. /path/to/datasets, with subdirectories:

   /path/to/datasets/
     ├── mini_imagenet/
     ├── CUB_200_2011/
     ├── cifar-fs/
     └── FC100/
   

3. Point your scripts Pass the dataset root path to train.py and evaluate.py via the --data_root argument.

---

▶️ Quick Start

1. Training

python src/train.py \
  --dataset cub \
  --data_root /path/to/datasets/CUB_200_2011 \
  --ways 5 \
  --shots 5 \
  --queries 15 \
  --episodes 1000 \
  --lr 1e-4 \
  --weight_decay 1e-4 \
  --batch_size 4 \
  --output_dir ./checkpoints/cub

• --dataset: one of mini, cub, cifarfs, fc100
• --ways/--shots/--queries: few-shot episode configuration
• --episodes: number of training episodes
• --lr & --weight_decay: AdamW hyperparameters
• --batch_size: number of episodes per gradient step
• --output_dir: where to save model checkpoints & logs

2. Evaluation

python src/evaluate.py \
  --dataset cub \
  --data_root /path/to/datasets/CUB_200_2011 \
  --checkpoint ./checkpoints/cub/best_model.pth \
  --ways 5 \
  --shots 5 \
  --queries 15 \
  --eval_episodes 100

• --checkpoint: path to the trained .pth file
• --eval_episodes: number of test episodes (e.g. 100 or 1000)

Results will be printed as mean accuracy ± 95% CI.

---

📒 Notebooks

For exploratory analysis and visualization, open one of the notebooks in the notebooks/ folder:

• CUB-200.ipynb
• FC100.ipynb
• CIFAR-FS.ipynb
• MiniImagenet.ipynb

Each notebook walks through:

1. Loading a trained checkpoint
2. Sampling episodes
3. Computing & plotting class prototypes
4. Visualizing support/query embeddings with t-SNE / PCA
5. Plotting learning curves

---

🔧 Utilities

• src/utils.py

  • Logging to console & CSV
  • Plotting loss & accuracy curves
  • CLI argument parsing

Feel free to extract or extend any helper functions for your own experiments.

---

📄 License

This project is licensed under the MIT License.

---

📧 Contact

For questions or suggestions, open an issue or contact me at abdulvahapmutlu1@gmail.com.