Tutorials

We list here a view small examples, that are easy to run locally on a CPU, to show how to use posteriors:

• Visualizing VI and SGHMC
  
  
  
• EKF for Premier League football
  
  
  
• Autoencoder with Lightning

Additionally, we provide more in depth examples in the examples directory on GitHub, many of which use larger models that require a GPU to run:

• bayes_llama3: Uses posteriors and lightning to train a bayesian ensemble language model Llama-3-8b on the TQA dataset.

• continual_lora: Uses posteriors.laplace.diag_fisher to avoid catastrophic forgetting in fine-tuning Llama-2-7b on a series of books from the pg19 dataset.

• yelp: Compares a host of posteriors methods (highlighting the easy exchangeability) on a sentiment analysis task adapted from the Hugging Face tutorial.

• imdb: Investigates cold posterior effect for a range of approximate Bayesian methods with a CNN-LSTM model on IMDB data, with some interesting takeaways.

• continual_regression: Variational continual learning notebook for a simple regression task that's easy to visualize.

• pyro_pima_indians_sghmc: An accessible notebook demonstrating the use of posteriors with a pyro-defined Bayesian logistic regression model, as well as convergence diagnostics from pyro.