Abstract
LAMDA-SSL is open-sourced on GitHub and its detailed usage documentation is available at https://ygzwqzd.github.io/LAMDA-SSL/. This documentation introduces LAMDA-SSL in detail from various aspects and can be divided into four parts. The first part introduces the design idea, features and functions of LAMDA-SSL. The second part shows the usage of LAMDA-SSL by abundant examples in detail. The third part introduces all algorithms implemented by LAMDA-SSL to help users quickly understand and choose SSL algorithms. The fourth part shows the APIs of LAMDA-SSL. This detailed documentation greatly reduces the cost of familiarizing users with LAMDA-SSL toolkit and SSL algorithms.
Introduction
In real-world machine learning applications, labeled data is often scarce and expensive to obtain, while unlabeled data is abundant. Semi-supervised learning (SSL) provides a powerful paradigm to leverage this unlabeled data to improve model performance. However, the current software ecosystem for SSL is fragmented. Established libraries like scikit-learn focus primarily on a limited set of statistical methods, whereas newer toolkits like USB are specialized for deep learning classification tasks. LAMDA-SSL addresses this gap by providing a comprehensive, unified toolkit that integrates both statistical and deep learning SSL algorithms across a wide range of tasks, including classification, regression, and clustering.
System Architecture
LAMDA-SSL is designed with a modular architecture to ensure flexibility and ease of use. It consists of two primary components: the Data Module and the Model Module.
- Data Module: This module handles data management and transformation. It supports four major data types—tabular, image, text, and graph—and provides 45 different transformation methods and augmentations to prepare data for SSL algorithms.
- Model Module: This module houses the core algorithms. LAMDA-SSL implements 30 algorithms in total, categorized into 12 statistical SSL algorithms (such as SSGMM, TSVM, and Label Propagation) and 18 deep SSL algorithms (including FixMatch, FlexMatch, and SSVAE).
A key innovation is the Simple Interface, which adopts the familiar fit() and predict() API from scikit-learn. By utilizing a DeepModelMixin, the toolkit allows PyTorch-based deep learning models to behave like standard scikit-learn estimators, facilitating seamless integration with existing workflows while supporting GPU acceleration and distributed training.
Supported Tasks and Datasets
Unlike many SSL toolkits that focus solely on classification, LAMDA-SSL provides extensive support for regression (e.g., CoReg, ICT Reg) and clustering (e.g., Constrained K-Means). The toolkit allows for standardized evaluation across various benchmarks, including CIFAR-10 for image classification, Cora for graph-based learning, and the Boston dataset for regression. The implementations are designed to meet or exceed the performance benchmarks established in the original algorithmic publications.
Comparative Analysis
LAMDA-SSL stands out by offering a significantly broader range of algorithms and features compared to existing toolkits.
As shown in the comparison above, LAMDA-SSL provides 30 algorithms, more than doubling the offerings of its closest competitors. It is the only toolkit to simultaneously support hyper-parameter searching, GPU acceleration, and distributed learning. Furthermore, its support for graph data and diverse task types (regression and clustering) makes it a versatile tool for both researchers and practitioners.
Conclusion and Future Work
LAMDA-SSL bridges the gap between traditional and modern SSL, offering a unified framework that simplifies the deployment of complex algorithms. Future development will focus on incorporating emerging state-of-the-art methods and expanding the toolkit to address "open environment" challenges, such as handling unlabeled data that contains classes not present in the labeled training set.
The toolkit is open-source and available at https://github.com/YGZWQZD/LAMDA-SSL.
Summary updated