TorchCP is a Python toolbox for conformal prediction research in deep learning, targeting researchers and practitioners who need to quantify uncertainty in model predictions. It provides a PyTorch-based framework with GPU acceleration for various tasks, including classification, regression, graph node classification, and LLM applications, enabling more reliable and interpretable AI systems.

How It Works

TorchCP implements a wide array of representative conformal prediction methods, categorized into post-hoc and training-time approaches. The framework is built upon AdverTorch, leveraging its structure for efficient implementation. It supports diverse tasks by offering specialized modules for classification, regression, graph neural networks, and language models, allowing users to select and apply appropriate methods for their specific deep learning architectures and data types.

Quick Start & Requirements

• Install via pip: pip install torchcp
• Prerequisites: Python 3.9, PyTorch 2.0.1. GPU acceleration is strongly recommended.
• Documentation and examples are available.

Highlighted Details

• Implements methods from numerous top-tier ML conferences (NeurIPS, ICML, ICLR).
• Covers classification, regression, graph node classification, and LLM tasks.
• Built on PyTorch with strong GPU acceleration.
• Claims 100% unit test coverage.

Maintenance & Community

• Maintained by Hongxin Wei's research group at SUSTech.
• Open to comments, issues, contributions, and collaborations.
• Recent update (1.0.2) includes refactoring, new classifiers, and bug fixes.

Licensing & Compatibility

• Licensed under LGPL.
• Permissive for commercial use and linking with closed-source applications.

Limitations & Caveats

The codebase is still under construction, with some features marked as TODO. While it supports GPU acceleration, specific CUDA version requirements are not detailed.