This repository provides the implementation for the Reptile meta-learning algorithm, focusing on finding good initializations for few-shot learning tasks. It is intended for researchers and practitioners in meta-learning and few-shot learning who want to reproduce or extend the results from the paper "On First-Order Meta-Learning Algorithms."

How It Works

Reptile is a first-order meta-learning algorithm that operates by repeatedly sampling a task, training on that task for a few steps, and then updating the meta-model's initialization towards the task-specific weights. This process aims to find an initialization that is broadly applicable across a distribution of tasks, enabling rapid adaptation to new, unseen tasks with minimal data.

Quick Start & Requirements

• Install/Run: Clone the repository and run the provided Python scripts (e.g., run_omniglot.py, run_miniimagenet.py).
• Prerequisites: Python, requires downloading datasets (~5GB) using the fetch_data.sh script.
• Setup Time: Data download takes 10-20 minutes. Training runs can be lengthy, with the paper's examples using up to 100,000 meta-iterations.
• Links: Paper (implied by repository name and description).

Highlighted Details

• Implements Reptile, a first-order meta-learning algorithm.
• Supports transductive and non-transductive evaluation modes.
• Includes scripts for reproducing paper results on Omniglot and Mini-ImageNet datasets.
• Allows comparison of different inner-loop gradient update strategies.

Maintenance & Community

• Status: Archived (code is provided as-is, no updates expected).
• No community links (Discord, Slack, etc.) are provided in the README.

Licensing & Compatibility

• The README does not explicitly state a license.

Limitations & Caveats

The repository is archived and will not receive further updates. The current implementation does not support resuming training from checkpoints.