log_joint_at crashes for nodes whose existence is conditional

[marcoct]

[assume mu (mem (lambda (j) (normal 0.0 10.0)))]
[assume z (mem (lambda (i) (uniform_discrete 0 2)))]
[observe (normal (mu (z 1)) 1.0) 0.321]
[infer global_log_joint]

Traceback (most recent call last):
  File "/usr/local/bin/venture", line 4, in <module>
    __import__('pkg_resources').run_script('venture==0.4.1.post313+ge9980af', 'venture')
  File "/usr/local/lib/python2.7/site-packages/pkg_resources/__init__.py", line 742, in run_script
    self.require(requires)[0].run_script(script_name, ns)
  File "/usr/local/lib/python2.7/site-packages/pkg_resources/__init__.py", line 1667, in run_script
    exec(code, namespace, namespace)
  File "/usr/local/lib/python2.7/site-packages/venture-0.4.1.post313+ge9980af-py2.7-macosx-10.9-x86_64.egg/EGG-INFO/scripts/venture", line 178, in <module>
    lite(args)
  File "/usr/local/lib/python2.7/site-packages/venture-0.4.1.post313+ge9980af-py2.7-macosx-10.9-x86_64.egg/EGG-INFO/scripts/venture", line 57, in lite
    local_ripl(args, v.Lite())
  File "/usr/local/lib/python2.7/site-packages/venture-0.4.1.post313+ge9980af-py2.7-macosx-10.9-x86_64.egg/EGG-INFO/scripts/venture", line 61, in local_ripl
    (interactive, r, files, plugins) = build()
  File "/usr/local/lib/python2.7/site-packages/venture-0.4.1.post313+ge9980af-py2.7-macosx-10.9-x86_64.egg/EGG-INFO/scripts/venture", line 86, in build
    r.execute_program_from_file(f)
  File "/usr/local/lib/python2.7/site-packages/venture-0.4.1.post313+ge9980af-py2.7-macosx-10.9-x86_64.egg/venture/ripl/ripl.py", line 284, in execute_program_from_file
    self.execute_program(f.read())
  File "/usr/local/lib/python2.7/site-packages/venture-0.4.1.post313+ge9980af-py2.7-macosx-10.9-x86_64.egg/venture/ripl/ripl.py", line 264, in execute_program
    return self.execute_parsed_program(*self.parse_program(program_string))
  File "/usr/local/lib/python2.7/site-packages/venture-0.4.1.post313+ge9980af-py2.7-macosx-10.9-x86_64.egg/venture/ripl/ripl.py", line 272, in execute_parsed_program
    vals.append(self.execute_instruction(instruction))
  File "/usr/local/lib/python2.7/site-packages/venture-0.4.1.post313+ge9980af-py2.7-macosx-10.9-x86_64.egg/venture/ripl/ripl.py", line 158, in execute_instruction
    self._raise_annotated(e, instruction)
  File "/usr/local/lib/python2.7/site-packages/venture-0.4.1.post313+ge9980af-py2.7-macosx-10.9-x86_64.egg/venture/ripl/ripl.py", line 154, in execute_instruction
    stringable_instruction)
  File "/usr/local/lib/python2.7/site-packages/venture-0.4.1.post313+ge9980af-py2.7-macosx-10.9-x86_64.egg/venture/ripl/ripl.py", line 170, in _execute_parsed_instruction
    return self.sivm.execute_instruction(parsed_instruction)
  File "/usr/local/lib/python2.7/site-packages/venture-0.4.1.post313+ge9980af-py2.7-macosx-10.9-x86_64.egg/venture/sivm/venture_sivm.py", line 69, in execute_instruction
    response = self._call_core_sivm_instruction(instruction)
  File "/usr/local/lib/python2.7/site-packages/venture-0.4.1.post313+ge9980af-py2.7-macosx-10.9-x86_64.egg/venture/sivm/venture_sivm.py", line 130, in _call_core_sivm_instruction
    response = self.core_sivm.execute_instruction(desugared_instruction)
  File "/usr/local/lib/python2.7/site-packages/venture-0.4.1.post313+ge9980af-py2.7-macosx-10.9-x86_64.egg/venture/sivm/core_sivm.py", line 46, in execute_instruction
    return f(instruction)
  File "/usr/local/lib/python2.7/site-packages/venture-0.4.1.post313+ge9980af-py2.7-macosx-10.9-x86_64.egg/venture/sivm/core_sivm.py", line 160, in _do_infer
    (did, val) = self.engine.infer(e)
  File "/usr/local/lib/python2.7/site-packages/venture-0.4.1.post313+ge9980af-py2.7-macosx-10.9-x86_64.egg/venture/engine/engine.py", line 234, in infer
    return self.raw_evaluate([v.sym("run"), program])
  File "/usr/local/lib/python2.7/site-packages/venture-0.4.1.post313+ge9980af-py2.7-macosx-10.9-x86_64.egg/venture/engine/engine.py", line 176, in raw_evaluate
    did = self._do_raw_evaluate(program)
  File "/usr/local/lib/python2.7/site-packages/venture-0.4.1.post313+ge9980af-py2.7-macosx-10.9-x86_64.egg/venture/engine/engine.py", line 184, in _do_raw_evaluate
    self.infer_trace.eval(did, program)
  File "/usr/local/lib/python2.7/site-packages/venture-0.4.1.post313+ge9980af-py2.7-macosx-10.9-x86_64.egg/venture/untraced/trace.py", line 53, in eval
    val = evaluator.eval(addr.Address(addr.List(id)), py_exp, self.env)
  File "/usr/local/lib/python2.7/site-packages/venture-0.4.1.post313+ge9980af-py2.7-macosx-10.9-x86_64.egg/venture/untraced/evaluator.py", line 54, in eval
    val = apply(address, nodes, env)
  File "/usr/local/lib/python2.7/site-packages/venture-0.4.1.post313+ge9980af-py2.7-macosx-10.9-x86_64.egg/venture/untraced/evaluator.py", line 78, in apply
    req_nodes = [evalRequest(req_args, spr, r) for r in requests.esrs]
  File "/usr/local/lib/python2.7/site-packages/venture-0.4.1.post313+ge9980af-py2.7-macosx-10.9-x86_64.egg/venture/untraced/evaluator.py", line 119, in evalRequest
    ans = node.Node(new_addr, eval(new_addr, r.exp, r.env))
  File "/usr/local/lib/python2.7/site-packages/venture-0.4.1.post313+ge9980af-py2.7-macosx-10.9-x86_64.egg/venture/untraced/evaluator.py", line 50, in eval
    v = eval(addr,subexp,env)
  File "/usr/local/lib/python2.7/site-packages/venture-0.4.1.post313+ge9980af-py2.7-macosx-10.9-x86_64.egg/venture/untraced/evaluator.py", line 54, in eval
    val = apply(address, nodes, env)
  File "/usr/local/lib/python2.7/site-packages/venture-0.4.1.post313+ge9980af-py2.7-macosx-10.9-x86_64.egg/venture/untraced/evaluator.py", line 80, in apply
    return applyPSP(spr.sp.outputPSP, OutputArgs(address, nodes[1:], env, req_nodes, requests))
  File "/usr/local/lib/python2.7/site-packages/venture-0.4.1.post313+ge9980af-py2.7-macosx-10.9-x86_64.egg/venture/untraced/evaluator.py", line 109, in applyPSP
    val = psp.simulate(args)
  File "/usr/local/lib/python2.7/site-packages/venture-0.4.1.post313+ge9980af-py2.7-macosx-10.9-x86_64.egg/venture/lite/psp.py", line 358, in simulate
    return self.f_type.wrap_return(self.psp.simulate(self.f_type.unwrap_args(args)))
  File "/usr/local/lib/python2.7/site-packages/venture-0.4.1.post313+ge9980af-py2.7-macosx-10.9-x86_64.egg/venture/lite/inference_sps.py", line 92, in simulate
    ans = getattr(engine, self.name)(*self.operands)
  File "/usr/local/lib/python2.7/site-packages/venture-0.4.1.post313+ge9980af-py2.7-macosx-10.9-x86_64.egg/venture/engine/inference.py", line 102, in log_joint_at
    return self.engine.model.traces.map('log_joint_at', scope, block)
  File "/usr/local/lib/python2.7/site-packages/venture-0.4.1.post313+ge9980af-py2.7-macosx-10.9-x86_64.egg/venture/multiprocess.py", line 201, in map
    return self.handle_result_list(res)
  File "/usr/local/lib/python2.7/site-packages/venture-0.4.1.post313+ge9980af-py2.7-macosx-10.9-x86_64.egg/venture/multiprocess.py", line 234, in handle_result_list
    raise exception_handler.gen_exception()
venture.exception.VentureException: *** evaluation: Cannot deterministically propose values for nodes whose existence is conditional
**************************************************
Stack trace from worker:
**************************************************
Traceback (most recent call last):
  File "/usr/local/lib/python2.7/site-packages/venture-0.4.1.post313+ge9980af-py2.7-macosx-10.9-x86_64.egg/venture/multiprocess.py", line 104, in wrapped
    res = f(*args, **kwargs)
  File "/usr/local/lib/python2.7/site-packages/venture-0.4.1.post313+ge9980af-py2.7-macosx-10.9-x86_64.egg/venture/multiprocess.py", line 340, in <lambda>
    return safely(lambda *args,**kwargs: getattr(self.obj, attrname)(*args, **kwargs))
  File "/usr/local/lib/python2.7/site-packages/venture-0.4.1.post313+ge9980af-py2.7-macosx-10.9-x86_64.egg/venture/lite/trace.py", line 572, in log_joint_at
    registerDeterministicLKernels(self, scaffold, pnodes, currentValues)
  File "/usr/local/lib/python2.7/site-packages/venture-0.4.1.post313+ge9980af-py2.7-macosx-10.9-x86_64.egg/venture/lite/infer/mh.py", line 32, in registerDeterministicLKernels
    raise Exception("Cannot deterministically propose values for nodes whose existence is conditional")
Exception: Cannot deterministically propose values for nodes whose existence is conditional

(run global_log_joint)
^^^^^^^^^^^^^^^^^^^^^^
Caused by
Cannot deterministically propose values for nodes whose existence is conditional
**************************************************
Stack trace from worker:
**************************************************
Traceback (most recent call last):
  File "/usr/local/lib/python2.7/site-packages/venture-0.4.1.post313+ge9980af-py2.7-macosx-10.9-x86_64.egg/venture/multiprocess.py", line 104, in wrapped
    res = f(*args, **kwargs)
  File "/usr/local/lib/python2.7/site-packages/venture-0.4.1.post313+ge9980af-py2.7-macosx-10.9-x86_64.egg/venture/multiprocess.py", line 340, in <lambda>
    return safely(lambda *args,**kwargs: getattr(self.obj, attrname)(*args, **kwargs))
  File "/usr/local/lib/python2.7/site-packages/venture-0.4.1.post313+ge9980af-py2.7-macosx-10.9-x86_64.egg/venture/lite/trace.py", line 572, in log_joint_at
    registerDeterministicLKernels(self, scaffold, pnodes, currentValues)
  File "/usr/local/lib/python2.7/site-packages/venture-0.4.1.post313+ge9980af-py2.7-macosx-10.9-x86_64.egg/venture/lite/infer/mh.py", line 32, in registerDeterministicLKernels
    raise Exception("Cannot deterministically propose values for nodes whose existence is conditional")
Exception: Cannot deterministically propose values for nodes whose existence is conditional

[axch]

There is a semantic logic to this problem. Imagine a model with an unknown number of continuous latents, like an uncollapsed DPMM with continuous cluster parameters. One might imagine computing the global posterior at any given configuration of parameters, but what does that actually mean? The number is a probability density, that must be interpreted with respect to something like counting measure on all the discrete variables and Lebesgue measure on the continuous ones. What would it mean to compare two such numbers, if the number of clusters is different? Formally, the "probability" of a (infinite-precision) configuration with more clusters is infinitely smaller than the "probability" of one with fewer. But the density in both cases is a finite number, and in fact, the information of how many Lebesgue-measured quantities were involved is lost by the time that number is presented.

What should global_log_joint to do in this circumstance?

(Note: global_log_likelihood remains sensible, because it's only the observations whose base measure is relevant, and that's presumably the same across all points of comparison.)

[lenaqr]

global_log_likelihood remains sensible

Well except when it already isn't. observe if (flip()) { uniform_continuous(0, 10) } else { uniform_discrete(0, 10) } = 5

[marcoct]

Despite the measure issues, it is possible to adapt the code to compute this quantity (each random choice can report its log probability (density) given its parents, plus global log likelihood) anyway, with a large warning in the documentation for this method indicating that results may be difficult to interpret or compare in certain cases?

Having this quantity, in addition to the global-log-likelihood, will help us to understand how our KL divergences and probability ratios compare to this quantity, and the semantic issues with this quantity may even be an important for highlighting the importance of using a technique like our KL-based measurement technique instead of log joint as a debugging tool.

Request: Create a second version of global_log_joint that has a scary sounding name like unsafe_global_log_joint (to discourage casual use) that doesn't crash when it encounters measure-theoretic problems.

[axch]

The structural problem is that regen and detach autmatically repropose brush from from the prior, and do not compute the weight.

Ways this request could be satisfied anyway:

• Generalize regen and detach to accept other kinds of proposals (e.g. the constant "use the current values") and compute their weights. This would be the most principled way to do it, but I don't want to architect it on a tight schedule.
• Hack regen and/or detach to accept a flag to just evaluate the prior density, please. I kinda don't want to do this, because regen and detach are subtle and covered in hacks already.
• Perform a custom trace traversal, evaluating the log density at each (random) node, perhaps insensitive to data dependencies. This seems the most expedient, especially since we already have an attempt at that in code, in the form of the getGlobalLogScore method of lite.Trace and new_cxx.Trace (the latter defined in pytrace.cxx).

Issues

• Base-measure compatibility as discussed above
  • Resolution: suspend disbelief
• Likelihood-free SPs, by assumption, cannot report likelihoods. Two options are to count 0 log likelihood (possibly with a printed warning) or crash. @marcoct : preference?
• The custom traversal plan will get the wrong answer in the presence of (exchangeably) coupled SPs, because each application will report its log density conditioned on all the other applications, whereas the correct thing is to count them in order.
  • Proposal: The common case is that the maker SP can report the log density of all applications of the made SP at once with the logDensityOfData method. So we can just invoke that on AAA nodes, and count 0 logDensity for nodes whose operator is an AAA node.
  • Alternative: Could unincorporate all nodes and then compute logDensity in tandem with reincorporating. This may actually be better, but I am currently afraid of it because it mutates the trace and I don't know whether it needs to be done in an order that respects the model's data dependencies.

Details, if adopting the custom proposal plan

• Rename the getGlobalLogScore method with a more specific (and perhaps cautionary) name.
• Implement the SP solution in getGlobalLogScore.
  • Any tests that see a different exact number for the value of getGlobalLogScore should probably just be updated.
  • Port same to Puma.
• Expose the new getGlobalLogScore as an inference SP named something like unsafe_global_log_joint.
  • Doesn't need to round-trip through the ripl, so can be an engine_method_sp.
  • Possibly include the above measure-theoretic caveats about comparability in its documentation.
  • Also document that it will still be wrong for likelihood-free SPs, and coupled SPs that are not AAA.
  • Could warn about likelihood-free SPs, but no reliable way to detect coupled but non-AAA SPs.

[axch]

@riastradh-probcomp, could you implement this? If the above discussion is not enough to go on, we can talk about it.

[axch]

Fixed by #572.