google-research/uda

Combines consistency regularization with task-specific augmentation strategies—back-translation for text and RandAugment for images—to enforce prediction invariance on unlabeled data. Implements dual loss objectives weighting labeled and unlabeled examples, with confidence-based masking to filter low-confidence predictions. Supports both TensorFlow on GPUs and Google Cloud TPU, with pre-built implementations for BERT text classification and vision tasks (CIFAR-10, SVHN, ImageNet).

2,202 stars. No commits in the last 6 months.