blei-lab/edward

Built on TensorFlow's computational graphs, Edward combines directed graphical models with neural networks and implicit generative models, supporting diverse inference algorithms from black-box variational inference and HMC to GANs and message passing. It enables composable inference strategies like expectation-maximization and ABC methods, alongside posterior predictive checks for model criticism. The library integrates with Keras and tf.layers for neural network components while leveraging TensorFlow's automatic differentiation and distributed training capabilities.

4,843 stars. No commits in the last 6 months.