Pyro is Uber AI lab open source programming language a depth of probability (PPL), based on Python and PyTorch, focusing on variation reasoning, while supporting the inference algorithm can be combined. pyro 0.3.4 version has been released, new features are as follows:
- A more flexible module of EasyGuide
- Automatic initialization wizard of different methods
- More standardized stream - Block Neural Autoregressive Flow, and squares, Sylvester flow, deepELUFlow, House Holder stream, RealNVP
- Support ReduceLROnPlateau Scheduler
- MCMC new interfaces:
- Potential function can be specified directly, rather than Pyro model HMC / NUTS kernel
- MCMC.summary () method provides site-level summaries and diagnostic information
- Instead of the predicted class of utility functions TracePredictive
- Add the divergent information to MCMC.diagnostics ()
- DiscreteHMM distributed parallel training for fast distribution of any observed discrete state Hidden Markov Model
Repair bug:
- pickling problem-solving Pyro processing program, making it possible to choose a larger model classes possible
- MCMC use of multi-processing errors more repair. Using the new interface, memory consumption is very low, thus allowing to collect more examples
- The model has a number of performance enhancements Sample Site
For more information, please see the release log .