Writing effective bayesian programs using TensorFlow and TFP
summary
Walks through concrete code patterns for Bayesian modeling with TensorFlow and Tensorflow Probability: uses JointDistributionSequential for generative and logprob paths, exploits batch-able chains so many MCMC chains run on one GPU/TPU, wraps samplers in tf.function with experimental_compile=True for XLA speedups, and applies bijector transforms for constrained parameters. Covers L‑BFGS and NUTS/HMC, ADVI for mass‑matrix init, common broadcasting shape bugs and fixes, and shows that these patterns actually run and scale rather than just being theoretical.
AI-written summary · about
chapters (13)
- 0:00Introduction
- 3:45Bayesian workflow basics
- 8:07Linear regression in TFP
- 13:11Distributions in TFP
- 25:11Bayesian model implementation
- 25:56Joint distribution API
- 37:49Debugging shape issues
- 41:31Batch processing & MCMC
- 55:47Robust regression
- 1:01:21Baseball player inference
- 1:05:38Mixture models & broadcasting
- 1:09:40Mixed effect models
- 1:15:59Summary & Q&A
related talks
01h 07mProbabilistic Programming and Readable Models
01h 03mA Personal Viewpoint on Probabilistic Programming
39mProbablistic Programming Data Science with PyMC3
01h 31mProbabilistic Programming and Bayesian Inference in Python
01h 22mDeep Probabilistic Methods with PyTorch
01h 26mProbabilistic Programming and Bayesian Computing with PyMC
47mThe State of the Art for Probabilistic Programming
39mThe Duct Tape of Heroes: Bayes Rule