State of the Tree / CSM

[dehann]

This issue is meant to bring together most relevant things to "Tree/CSM" as they stand now 3Q20.

We can edit and update here as needed.

References:

• ::BayesTree is full of cliques (::TreeClique),
  • Each clique contains a data container ::BayesTreeNodeData
• When solving, a Clique State Machine (CSM) is launched for each clique,
  • One common entrypoint tree, csmtasks, hists = solveTree!(fg [,tree], ...).
• CSM is built on FSM,
  • Here is the FSM loop,
  • Here is the same CSM loop point,
• FSM has early termination features like limititer and break-on-function which are patched through from CSM,
  • e.g. limititers per CSM: Documenting CSM and inference steps for debugging #443 (comment)
• Each CSM has its own ::CliqStateMachineContainer of data references
  • Most references are shared memory
  • csmc.cliqSubFg is the main local subgraph copy workhorse (it's reference is only local to each CSM individually).
• Each CSM interacts with other CSMs via either an up or down atomic ::Channel{LikelihoodMessage} (Consolidate belief messages and message channels  #459),
  • The Channels are stored in each clique's ::BayesTreeNodeData (Decide if msg channels in BTND or edges only #675),
    • e.g. upMsgChannel
    • A "pull" model is used to get a message from another CSM (CSM message passing push or pull model #674, pictorial Decide if msg channels in BTND or edges only #675 (comment))
  • Channels fetch (peek) vs. take! model fetch vs take! on message channels #855 (currently undecided on path forward)
• Format of data messages being exchanged are ::LikelihoodMessage for all solvers (parametric or nonparametric)
• Likelihood message are added and removed from cliqSubFg as either CSM tags e.g LIKELIHOODMESSAGE must be exclusive  #782
  • LIKELIHOODMESSAGE for all message factors,
  • UPWARD_DIFFERENTIAL for the deconvolved relative information between joint variables in a message,
  • UPWARD_COMMON for the gauge information during upsolve,
  • DOWNWARD_COMMON for the complete posterior gauge information during downsolve.
• A discussion on how to debug with CSM is available here Documenting CSM and inference steps for debugging #443,
  • And snapshot of a particular hex tree-init development example.
    • And can produce animations of whats happening like this.
  • See this issue for a read along tutorial style example of the debugging process (CSM init endless cycle MWE fail case  #754).