Eric Lippert
Eric Lippert
Summary: The compiler should detect when a model uses an operation which we know for certain cannot be accumulated into the graph and immediately throw a helpful error message. This...
Summary: We represent edge requirements imposed by BMG by a requirement object; there is no reason why we ever need two or more instances of the object representing "there is...
Summary: We're getting to the end of this series of refactorings. In this diff I start adding *combinators* on *graph fixers*. Recall that a graph fixer is a pass which...
Summary: In this diff I complete the initial round of refactorings of the graph fixing passes, by refactoring the passes which turn an observe-true on a Bernoulli into a factor...
Summary: I'm continuing the work I began in the previous diff. We have numerous difficulties in the "problem fixing" phase where the graph is gradually rewritten from its accumulated form...
Summary: These three performance tests do only one thing: run inference with and without a certain optimization, and measure the change in the number of nodes and edges in the...
Summary: See https://github.com/facebookresearch/beanmachine/issues/1312 This user-supplied repro with a simple model produces an internal compiler error. What's happening here is: the graph is mutable, and large graphs may be expensive to...
Summary: This diff continues the work begun in the previous; I am adding location information to error messages when known. In this diff, violations of BMG type system rules are...
Summary: The compiler error messages were a little bit awkward: `The model uses a & operation unsupported by Bean Machine Graph.` Not wrong but hard to read. We always said...
Summary: I'm continuing my efforts to: * Refactor and simplify the operator handling code to unify it with the special-function-handling code. * Ensure that all the ways to represent an...