asyml/texar

Provides modularized, reusable components (encoders, decoders, attention mechanisms) decomposed along a Learning-Inference-Model architecture, with built-in support for multiple training paradigms including maximum likelihood, reinforcement learning, and adversarial learning. Integrates pre-trained models like BERT and GPT-2 with uniform interfaces, and maintains parallel PyTorch and TensorFlow implementations with largely compatible APIs for cross-framework flexibility.

2,392 stars and 225 monthly downloads. No commits in the last 6 months. Available on PyPI.