🌸 BloomLens: Few-shot Learning for Fine-grained Flower Classification with Prototypical Networks

📖 Paper | 🎯 Models | 📊 Results |

A Course Project for Nanyang Technological University, SC4001 CE/CZ4042: Neural Networks and Deep Learning

🌟 Highlights

• 🚀 93.64% accuracy on 5-way 1-shot tasks
• 🎯 85.51% accuracy on 20-way 1-shot tasks
• 📈 Scales to 40-way tasks with 78.29% accuracy
• 🔄 Progressive training from 5-way to 20-way
• 🤖 Transformer-enhanced feature adaptation
• 🎨 Smart augmentation with MixUp and CutMix

🤖 Model Components

📊 Performance on Oxford Flowers-102

Note: the specific data split is explained in the paper

Model	5-way 1-shot	5-way 5-shot	20-way 1-shot	20-way 5-shot
AlexNet	41.95 ± 2.01	52.16 ± 2.16	17.13 ± 0.76	22.75 ± 0.78
ResNet18	57.59 ± 2.18	68.61 ± 2.29	31.39 ± 1.07	42.61 ± 0.97
ResNet50	54.21 ± 2.23	63.95 ± 2.30	27.90 ± 0.94	38.16 ± 0.98
DenseNet121	55.16 ± 2.08	67.61 ± 2.06	31.61 ± 1.08	43.69 ± 0.96
DenseNet201	58.52 ± 2.36	69.51 ± 2.06	31.97 ± 1.20	44.47 ± 1.05
Bayesian Prompt	70.40 ± 1.80	73.50 ± 1.50	-	-
BloomLens (Ours)	93.64 ± 6.86	95.88 ± 5.20	85.51 ± 5.77	89.66 ± 4.00

🚀 Quick Start

Installation

# Clone the repository
git clone https://github.com/Ry3nG/BloomLens.git

# Create conda environment
conda env create -f environment.yml

# Activate conda environment
conda activate bloomlens

Training

python src/training/train_prototypical.py

Testing

# Testing Prototypical Network
python src/evaluation/evaluate_prototypical.py
# Testing Baseline Model
python scripts/baseline_comparison_multimodel.py

Project Structure

BloomLens/
├── 📂 results/
├── 📂 scripts/
├── 📂 src/
│   ├── 📂 data/
│   ├── 📂 evaluation/
│   ├── 📂 models/
│   └── 📂 training/
├── 📂 docs/
│   └── 📂 diagrams/
├── 📄 environment.yml
└── 📄 README.md

📊 Monitoring

Training progress can be monitored using wandb.

wandb login # login to wandb

import wandb
wandb.init(project="bloomlens")