thuml/LogME

LogME computes a lightweight transferability metric by fitting a Gaussian-naive-Bayes classifier to frozen pre-trained features, enabling model selection without fine-tuning hyperparameters. Beyond ranking models, it includes B-Tuning for ensemble-based adaptation across multiple heterogeneous pre-trained models, and provides implementations of comparative baselines (LEEP, NCE) with corrected algorithms. The library offers a scikit-learn compatible API supporting classification, multi-label, and regression tasks across standard vision model hubs.

211 stars. No commits in the last 6 months.