LibFewShot: A Comprehensive Library for Few-Shot Learning.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Few-shot learning, particularly image classification, is advancing rapidly.
• Generic techniques like data augmentation and pre-training significantly improve performance.
• Inconsistent software platforms and architectures hinder fair comparisons and reproducibility.

Purpose of the Study:

• To develop a comprehensive library (LibFewShot) for few-shot learning.
• To provide standardized re-implementations of state-of-the-art methods.
• To conduct thorough evaluations of training tricks and meta-training necessity.

Main Methods:

• Re-implemented 18 state-of-the-art few-shot learning methods in a unified PyTorch codebase.
• Conducted comprehensive evaluations on multiple benchmarks.
• Assessed the impact of various training tricks and backbone architectures.

Main Results:

• LibFewShot offers a standardized framework for reproducible research.
• Evaluations highlight the significant impact of common training techniques.
• Meta- or episodic-training is confirmed as necessary, especially when combined with pre-training.

Conclusions:

• LibFewShot lowers entry barriers for few-shot learning research.
• The study elucidates the effects of various training strategies.
• Meta-training remains essential for effective few-shot image classification.