TripletMAML: A metric-based model-agnostic meta-learning algorithm for few-shot classification

TripletMAML implements a novel few-shot learning approach by integrating triplet-loss-based metric learning within the Model-Agnostic Meta-Learning (MAML) framework. This repository supports both classification and retrieval tasks while providing reproducible, transparent, and well-documented experiments.

---

Table of Contents

• Citation
• Overview
• Features
• Installation & Dependencies
• Dataset Preparation
• Usage
• Directory Structure
• Detailed Documentation
  • Algorithm Details
  • Code Modules
  • Step-by-Step Usage Guides
• Permanent Links & Reproducibility

Citation

If you find TripletMAML useful in your research, please use the following BibTeX entry for citation.

@misc{gulcu2025tripletmaml,
  author =       {Ayla Gülcü, Zeki Kuş, İsmail Taha Samed Özkan and Osman Furkan Karakuş},
  title =        {TripletMAML: A metric-based model-agnostic meta-learning algorithm for few-shot classification},
  howpublished = {\url{https://github.com/aylagulcu/TripletMAML}},
  journal =      {The Visual Computer}
  year =         {2025}
}

Overview

TripletMAML leverages the strengths of both meta-learning and metric-based learning to tackle the challenges of few-shot classification. By incorporating a triplet loss into the MAML framework, the algorithm learns robust and discriminative feature representations even when data is scarce. This repository includes:

• Classification Tasks: Training and evaluation scripts for few-shot classification.
• Retrieval Tasks: Scripts to test the retrieval capability using learned embeddings.
• Interactive Experimentation: A Jupyter Notebook for task control and hyperparameter tuning.

---

Features

• Few-Shot Learning: Designed specifically for few-shot scenarios.
• Metric-Based Learning: Uses triplet loss to improve representation learning.
• Meta-Learning Framework: Builds on MAML for fast adaptation to new tasks.
• Comprehensive Dataset Support: Includes scripts and instructions for Omniglot, MiniImageNet, CUB-200-2011, and CIFAR Few-Shot.
• Detailed Documentation: Clear instructions, algorithm details, and permanent link assurance for reproducibility.

---

Installation & Dependencies

Prerequisites

• Operating System: Ubuntu 20.04 LTS is recommended.
• Python Version: 3.9.7 (Anaconda distribution is suggested).

Required Libraries

• python == 3.9.7
• learn2learn == 0.1.7
• numpy == 1.20.3
• pytorch == 1.10.2 (compatible with py3.9_cuda11.3_cudnn8.2.0_0)
• scikit-learn == 1.2.0

Setup Instructions

git clone https://github.com/aylagulcu/TripletMAML.git
cd TripletMAML
python3 -m venv env
source env/bin/activate
pip install -r requirements.txt

---

Dataset Preparation

Dataset	Download	Download Needed
Omniglot	download	NO
MiniImageNet	download	NO
CUB-200-2011	download	YES
CIFAR Few-Shot	download	YES

• for Omniglot

  TripletOmniglot.py Files will handle downloading automatically if download mode is set as True.
  

• for MiniImageNet

  ./data/MiniImageNet/MiniImageNet_downloader.py will handle downloading test-train-validation files accordingly. 
  No needed to install another dataset.
  

• for CUB-200-2011

  CUB-200-2011 files are needed to be downloaded from given link. Then replaced into corresponding folder which is ./data/CUB/
  Choose the generator of your preference so they can create the test-train-validation files accordingly.
  

• for CIFAR Few-Shot

  CIFAR Few-Shot files are needed to be downloaded from given link. Then replaced into corresponding folder which is ./data/CIFARFS100/
  First run the processor then you can use generator to create the test-train-validation files accordingly.
  

Usage

Classification

python TripletMAML/maml_triplet_train_test_val.py

Retrieval

python TripletMAML/maml_triplet_test_retrieval.py

Interactive Notebook

Open:

TripletMAML/TaskControl.ipynb

Directory Structure

|—— .gitignore
|—— data
|    |—— Generators
|        |—— CIFARFS100
|            |—— CIFARFS100_generator.py
|            |—— CIFARFS100_processor.py
|        |—— CUB
|            |—— CUB_BB_NoResize_generator.py
|            |—— CUB_BB_Resize_generator.py
|            |—— CUB_NoBB_generator.py
|        |—— Flowers
|            |—— Flowers_generator.py
|        |—— MiniImageNet
|            |—— MiniImageNet_downloader.py
|—— HPO
|    |—— backbone.py
|    |—— de.py
|    |—— losses.py
|    |—— model.py
|    |—— rs.py
|    |—— train.py
|    |—— Triplets
|        |—— TripletCUB.py
|        |—— TripletFlowers.py
|        |—— TripletFSCIFAR100.py
|        |—— TripletMiniImageNet.py
|        |—— TripletOmniglot.py
|        |—— __init__.py
|—— TripletMAML
|    |—— backbone.py
|    |—— losses.py
|    |—— maml_triplet_test_retrieval.py
|    |—— maml_triplet_train_test_val.py
|    |—— TaskControl.ipynb
|    |—— Triplets
|        |—— TripletCUB.py
|        |—— TripletFlowers.py
|        |—— TripletFSCIFAR100.py
|        |—— TripletMiniImageNet.py
|        |—— TripletOmniglot.py
|        |—— __init__.py

Detailed Documentation

Algorithm Details

• Triplet Loss: Enhances feature separability across classes.
• MAML: Allows quick adaptation using meta-learning.
• Integration: Triplet loss is incorporated in the inner loop of MAML.

Code Modules

• backbone.py: Defines the neural network architecture.
• losses.py: Implements the triplet loss.
• maml_triplet_train_test_val.py: Training and validation logic.
• maml_triplet_test_retrieval.py: Retrieval evaluation logic.

Step-by-Step Usage

1. Prepare dataset.
2. Configure parameters in the relevant script.
3. Run the experiment (see Usage section).

---

Permanent Links & Reproducibility

All dataset and code links are verified to be permanent. This documentation is designed for full reproducibility by other researchers.