Alexey Radul
Alexey Radul
- [ ] Synthesize parameters and indexes for testSameMarginal randomly (quickcheck style) - [ ] Dispatch on data type for sameness of distribution tests in testSameMarginal and testUCKernel - [...
against Monte Carlo assessment estimates from the (explicitly?) uncollapsed version. The former should be the mean of the latter.
Sampler agreement - [ ] Collapsed vs uncollapsed multivariate normal (need to take 1-D marginals, presmably) - [ ] Collapsed vs stick-breaking forms of Chinese Restaurant Process Write and test...
Consider a CRP whose current set of assignments is 1, 1, 3 (to wit, 2 had been assigned before, but was moved by inference). If the CRP is to emit...
By inspection (which may be mistaken?), it seems that the set of tables listed by `enumerate` will disagree with the set produced by `simulate` if the set of empty tables...
``` (assume f (make_lazy_hmm ...)) (observe (f 1) atom) (predict (f 7)) (resample 3) ``` produces ``` ... File "/home/axch/work/pcp/Venturecxx/build/lib.linux-x86_64-2.7/venture/lite/hmm.py", line 93, in simulateLatents if shouldRestore: aux.xs.append(latentDB[0]) VentureException: *** evaluation:...
``` $ venture puma venture[script] > assume foo = make_dir_mult(array(0.5, 0.5), array(true, false)) RuntimeError: make_dir_mult expects 1 arguments, not 2 ``` Likewise the uncollapsed and symmetric versions.
Right now it uses either one-sample or two-sample chi^2, depending on whether one of the desired distributions is assessable. I recently found myself investigating the wisdom and appropriateness of this...
For purposes of taking derivatives, a stochastic function (simulate) should be viewed as a deterministic function that takes the PRNG state as an additional argument. The derivative with respect to...
It should be really easy to get some Venture model, name some scope, and get what the current values there are and what Venture thinks the local posterior gradient there...