def testAuxWithoutLatent():
# raise(SkipTest('for quick results.'))
"Checks the posterior distribution on a Gaussian given an unlikely observation"
ripl = get_ripl()
ripl.assume("a", "(scope_include (quote A) 0 (normal 10.0 1.0))", label="pid")
ripl.observe("(scope_include (quote B) 0 (normal a 1.0))", 20.0)
# fake mu1=9, so that assumed_gibbs is doomed to fail the statistical test
MyNormalBiasedPosterior = partial(NormalExactPosterior, mu1=9, sigma1=1.0, sigma2=1.0)
ripl.register_proposal_program_class("MyNormalBiasedPosterior", MyNormalBiasedPosterior)
# Posterior for a is normal with mean 15, precision 2
# ripl.predict("(normal a 1.0)")
proposal_src = """
[declare {
"name":"normalbiased",
"class":"MyNormalBiasedPosterior",
"conditioned":[["B",0]],
"target":[["A",0]],
"ready":"yes",
"num_samples":0}]
"""
ripl.execute_program(proposal_src)
predictions = collectSamples(ripl,"pid",infer="(custommh normalbiased aux 1 20)")
cdf = stats.norm(loc=15, scale=math.sqrt(0.5)).cdf
return reportKnownContinuous(cdf, predictions, "N(15,sqrt(0.5))")
def testCustomTrainer():
# raise(SkipTest('for quick results.'))
"Train a proposal to inverse a QMR model with rare diseases, using custom trainer"
ripl = get_ripl()
ripl.assume("d0", "(scope_include (quote D) 0 (bernoulli 0.01))", label="D0")
ripl.assume("d1", "(scope_include (quote D) 1 (bernoulli 0.01))", label="D1")
ripl.assume("d2", "(scope_include (quote D) 2 (bernoulli 0.005))", label="D2")
ripl.assume("joint", "(+ (* d0 4) (* d1 2) d2)", label="pid")
ripl.observe("(scope_include (quote S) 0 (bernoulli (- 1.0001 (pow 0.5 (+ d0 d1)))))", 1.0)
ripl.observe("(scope_include (quote S) 1 (bernoulli (- 1.0001 (pow 0.5 (+ d0 d2)))))", 1.0)
ripl.observe("(scope_include (quote S) 2 (bernoulli (- 1.0001 (pow 0.5 (+ d1 d2)))))", 1.0)
ripl.register_proposal_program_class("LogisticRegressionProposalProgram", LogisticRegressionProposalProgram)
ripl.register_trainer_src("QMR_highprior", QMR_highprior)
# Posterior for d2 is Bernoulli with p(d2=1) = 0.5
proposal_src = """
[declare {
"name":"logreg",
"class":"LogisticRegressionProposalProgram",
"conditioned":[["S",0], ["S",1], ["S",2]],
"target":[["D",0], ["D",1], ["D",2]],
"trainer":"QMR_highprior",
"num_samples":1000}]
"""
ripl.execute_program(proposal_src)
predictions = collectSamples(ripl,"pid",infer="(custommh logreg aux 1 5)")
ans = [(0, 0.000001), (1, 0.003), (2, 0.006), (3, 0.25), (4, 0.006), (5, 0.25), (6, 0.5), (7, 0.004)]
return reportKnownDiscrete(ans, predictions)
def testAssumedGibbsWithLatent():
# raise(SkipTest('for quick results.'))
"Checks an exact proposal with latent randomness"
ripl = get_ripl()
ripl.assume("x", "(scope_include (quote X) 0 (normal 0 1.0))", label="pid")
ripl.observe("(scope_include (quote Y) 0 (normal (sqrt (* x x)) 1.0))", 10.0)
MyBimodalExactPosterior = partial(BimodalExactPosterior, mu1=0, sigma1=1.0, sigma2=1.0)
ripl.register_proposal_program_class("MyBimodalExactPosterior", MyBimodalExactPosterior)
# Posterior for a is bi-modal with two modes at -5, 5, both with precision 2
# ripl.predict("(normal a 1.0)")
proposal_src = """
[declare {
"name":"bimodalexact",
"class":"MyBimodalExactPosterior",
"conditioned":[["Y",0]],
"target":[["X",0]],
"ready":"yes",
"num_samples":0}]
"""
ripl.execute_program(proposal_src)
predictions = collectSamples(ripl,"pid",infer="(custommh bimodalexact assumed_gibbs 1 1)")
cdf = (lambda x:(stats.norm(loc=-5, scale=math.sqrt(0.5)).cdf(x)+stats.norm(loc=5, scale=math.sqrt(0.5)).cdf(x))*0.5)
return reportKnownContinuous(cdf, predictions, "0.5*N(-5,sqrt(0.5))+0.5*N(5,sqrt(0.5))")
def testAnnotateInferenceErrorInDefinedDo():
ripl = get_ripl(persistent_inference_trace=True)
ripl.define(
"act", """\
(do (assume x (normal 0 1))
(y <- (sample x))
(observe x (+ 1 badness)))""")
err.assert_error_message_contains(
"""\
(run act)
^^^^^^^^^
(do (assume x (normal 0 1)) (y <- (sample x)) (observe x (add 1 badness)))
^^^^^^^
""", ripl.infer, "act")
def testNormalWithObserve1(seed):
# Checks the posterior distribution on a Gaussian given an unlikely
# observation
ripl = get_ripl(seed=seed)
ripl.assume("a", "(normal 10.0 1.0)", label="pid")
ripl.observe("(normal a 1.0)", 14.0)
# Posterior for a is normal with mean 12, precision 2
(samples, weights) = collectLikelihoodWeighted(ripl, "pid")
for (s, w) in zip(samples, weights):
# The weights I have should be deterministically given by the likelihood
assert_almost_equal(math.exp(w), stats.norm(loc=14, scale=1).pdf(s))
# The test points should be drawn from the prior
return reportKnownGaussian(10, 1, samples)
def testAnnotateInModelError():
# Tests Github Issue #538.
ripl = get_ripl()
ripl.set_mode("venture_script")
err.assert_error_message_contains(
"""\
*** evaluation: Nested ripl operation signalled an error
(autorun (in_model (run (new_model)) (action (run (sample (add foo 1))))))
^^^^^^^^^^^^^^^^^^^^
Caused by
*** evaluation: Cannot find symbol 'foo'
(add foo 1.0)
^^^
""", ripl.evaluate, "in_model(run(new_model()), action(run(sample(foo + 1))))")
def testAnnotateInferenceErrorInDo():
# TODO I need the inference trace to be persistent to trigger the
# inference prelude did skipping hack :(
ripl = get_ripl(persistent_inference_trace=True)
expression = """\
(do (assume x (normal 0 1))
(observe x (+ 1 badness)))"""
err.assert_error_message_contains(
"""\
(run (do (assume x (normal 0 1)) (observe x (add 1 badness))))
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
(run (do (assume x (normal 0 1)) (observe x (add 1 badness))))
^^^^^^^
""", ripl.infer, expression)
def testHMMResampleSmoke():
ripl = get_ripl()
ripl.assume("f","""
(make_lazy_hmm
(simplex 0.5 0.5)
(matrix (array (array 0.7 0.3)
(array 0.3 0.7)))
(matrix (array (array 0.9 0.2)
(array 0.1 0.8))))
""")
ripl.observe("(f 1)","integer<0>")
ripl.predict("(f 7)")
ripl.infer("(resample 3)")
ripl.infer("(mh default one 10)")
def check(mode):
with tempfile.NamedTemporaryFile(prefix='serialized.ripl') as f:
v1 = get_ripl()
v1.assume('is_tricky', '(flip 0.2)')
v1.assume('theta', '(if is_tricky (beta 1.0 1.0) 0.5)')
v1.assume('flip_coin', '(lambda () (flip theta))')
v1.observe('(flip_coin)', 'true')
v1.infer(1)
result1 = v1.predict('theta', label='theta_prediction')
call(v1, 'save', f.name, mode)
v2 = get_ripl()
call(v2, 'load', f.name, mode)
result2 = v2.report('theta_prediction')
result3 = v2.predict('theta')
assert result1 == result2 and result1 == result3
text1 = v1.get_text(1)
text2 = v2.get_text(1)
assert text1 == text2
def testBinomial1(seed):
# A simple test that checks the interface of binomial and its
# simulate method
ripl = get_ripl(seed=seed)
p = 0.3
n = 4
ripl.assume("p", "(if (flip) %f %f)" % (p, p))
ripl.predict("(binomial %d p)" % n, label="pid")
predictions = collectSamples(ripl, "pid")
ans = [(x, scipy.stats.binom.pmf(x, n, p)) for x in range(n + 1)]
assert_almost_equal(sum([xx[1] for xx in ans]), 1)
return reportKnownDiscrete(ans, predictions)
def testResampling1(seed):
P = 10
ripl = get_ripl(seed=seed)
def a_sample():
ripl.clear()
ripl.infer("(resample %d)" % P)
ripl.assume("x", "(normal 0 1)")
ripl.observe("(normal x 1)", 2)
ripl.infer("(resample 1)")
return ripl.sample("x")
predictions = [a_sample() for _ in range(default_num_samples())]
return reportKnownGaussian(1, math.sqrt(0.5), predictions)
def check_beta_bernoulli(maker, action, seed):
if maker == "make_uc_beta_bernoulli" and action in [
'serialize', 'convert_lite', 'convert_puma'
]:
raise SkipTest(
"Cannot convert BetaBernoulliSP to a stack dictionary. Issue: https://app.asana.com/0/9277420529946/16149214487233"
)
v = get_ripl(seed=seed)
v.assume('a', '(normal 10.0 1.0)')
v.assume('f', '({0} a a)'.format(maker))
v.predict('(f)', label='pid')
for _ in range(20):
v.observe('(f)', 'true')
return _test_serialize_program(v, 'pid', action)
def checkEnumerativeGibbsXOR2(in_parallel, seed):
# Tests that an XOR chain mixes with enumerative gibbs.
ripl = get_ripl(seed=seed)
ripl.assume("x", "(tag 0 0 (bernoulli 0.0015))", label="pid")
ripl.assume("y", "(tag 0 0 (bernoulli 0.0005))")
ripl.assume("noisy_true",
"(lambda (pred noise) (flip (if pred 1.0 noise)))")
ripl.observe("(noisy_true (= (+ x y) 1) .000001)", "true")
infer = "(gibbs 0 0 %s %s)" % \
(default_num_transitions_per_sample(), in_parallel)
predictions = collectSamples(ripl, "pid", infer=infer)
ans = [(True, .75), (False, .25)]
return reportKnownDiscrete(ans, predictions)
def testMVNormalRandomWalkSoundness(seed):
# This exercises the subtlety involving block proposals and delta
# kernels described in the "joint-delta-kernels" footnote in
# doc/on-latents.md.
r = get_ripl(seed=seed)
r.assume("mean", "(multivariate_normal (array 0) (id_matrix 1))")
r.assume("y", "(multivariate_normal mean (id_matrix 1))")
predictions = [
c[0] for c in collectSamples(r,
"y",
infer="(resimulation_mh default all 50)",
num_samples=default_num_samples(10))
]
return reportKnownGaussian(0, math.sqrt(2), predictions)
def testCycleKernel(seed):
# Same example as testBlockingExample0, but a cycle kernel that
# covers everything should solve it
ripl = get_ripl(seed=seed)
ripl.assume("a", "(tag 0 0 (normal 10.0 1.0))", label="pid")
ripl.assume("b", "(tag 1 1 (normal a 1.0))")
ripl.observe("(normal b 1.0)", 14.0)
infer = "(repeat %s (do (mh 0 0 1) (mh 1 1 1)))" % \
default_num_transitions_per_sample()
predictions = collectSamples(ripl, "pid", infer=infer)
return reportKnownGaussian(34.0 / 3.0, math.sqrt(2.0 / 3.0), predictions)
def checkBrushScope(operator):
# Check that putting scope control in the brush doesn't cause
# particle Gibbs to crash.
ripl = get_ripl()
ripl.assume("x1", "(tag (quote state) 0 (normal 1 1))")
ripl.assume(
"t", "1") # This variable matters to get the block id into the brush.
ripl.assume(
"x2", """
(if (> x1 1)
(tag (quote state) t (normal 2 1))
(tag (quote state) t (normal 0 1)))
""")
ripl.infer("(%s 'state ordered 4 3)" % operator)
def testAnnotateErrorTriggeredByInferenceOverProgrammaticAssume():
# Do not use the do macro yet
ripl = get_ripl()
ripl.infer("(assume control (flip))")
ripl.infer("(force control true)")
ripl.infer("(predict (if control 1 badness))")
# TODO Solve the double macroexpansion problem
err.assert_error_message_contains(
"""\
((biplex control (make_csp (quote ()) (quote 1.0)) (make_csp (quote ()) (quote badness))))
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
((biplex control (make_csp (quote ()) (quote 1.0)) (make_csp (quote ()) (quote badness))))
^^^^^^^
""", ripl.infer, "(resimulation_mh default one 50)")
def testOccasionalRejectionBrush(seed):
# Another version, this time with explicit brush creating the mix mh
# correction.
# Note that in this case, the correction is sound: The number of
# random choices available in the default one scope really is
# changing, whereas in `testOccasionalRejection` it is not.
# To see why, consider what transition the operator (gibbs default
# one 1) induces on this model (assuming the current behavior of
# always claiming the proposal weight is 0).
# - False, False is not possible
# - From False, True:
# - With probability 50%, enumerate the second coin, and propose
# to keep it at True.
# - Else, enumerate the first coin, find that both states are
# equally good, and
# - With probability 50%, propose to leave it
# - Else, propose to change it to True, which is accepted (with
# or without the correction)
# - Ergo, move to the True state 25% of the time.
# - From True, enumerate the first coin
# - With probability 50%, the second coin comes up False in the brush;
# propose to stay in the True state.
# - Else, both states equally good
# - With probability 50%, propose to stay in the True state
# - Else, propose to move to the False, True state
# - If the correction is applied, this proposal will be
# rejected with probability 50%.
# - Ergo, move to the False, True state 25% (no correction) or
# 12.5% (correction) of the time.
# - The former will induce a 50/50 stationary distribution on the
# value of flip1, whereas the right answer is 2:1 odds in favor of
# True.
r = get_ripl(seed=seed)
r.execute_program("""
(assume flip1 (flip))
(assume flip1_or_flip2
(if flip1 true (flip)))
(observe (exactly flip1_or_flip2) true)
;; Reject with a non-negligible probability per transition, which would
;; cause a crash if Gibbs couldn't handle rejection
(gibbs default one 50 false)
""")
infer = "(gibbs default one %s false)" % default_num_transitions_per_sample(
)
predictions = collectSamples(r, address="flip1", infer=infer)
ans = [(True, 2.0 / 3), (False, 1.0 / 3)]
return reportKnownDiscrete(ans, predictions)
def testCategorical1(seed):
# A simple test that checks the interface of categorical and its
# simulate method
ripl = get_ripl(seed=seed)
ripl.assume("x", "(categorical (simplex 0.1 0.2 0.3 0.4) (array 1 2 3 4))")
ripl.assume("y", "(categorical (simplex 0.2 0.6 0.2) (array 1 2 3))")
ripl.predict("(+ x y)", label="pid")
predictions = collectSamples(ripl, "pid")
ans = [(2, 0.1 * 0.2), (3, 0.1 * 0.6 + 0.2 * 0.2),
(4, 0.1 * 0.2 + 0.2 * 0.6 + 0.3 * 0.2),
(5, 0.2 * 0.2 + 0.3 * 0.6 + 0.4 * 0.2), (6, 0.3 * 0.2 + 0.4 * 0.6),
(7, 0.4 * 0.2)]
return reportKnownDiscrete(ans, predictions)
def testCMVN2D_mu1(seed):
if backend_name() != "lite": raise SkipTest("CMVN in lite only")
ripl = get_ripl(seed=seed)
ripl.assume("m0", "(array 5.0 5.0)")
ripl.assume("k0", "7.0")
ripl.assume("v0", "11.0")
ripl.assume("S0", "(matrix (array (array 13.0 0.0) (array 0.0 13.0)))")
ripl.assume("f", "(make_niw_normal m0 k0 v0 S0)")
ripl.predict("(f)", label="pid")
predictions = collectSamples(ripl, "pid")
mu1 = [p[0] for p in predictions]
return reportKnownMean(5, mu1)
def test_profiling_likelihoodfree():
# Make sure profiling doesn't break with likelihood-free SP's
class TestPSP(LikelihoodFreePSP):
def simulate(self, args):
x = args.operandValues()[0]
return x + stats.distributions.norm.rvs()
tester = typed_nr(TestPSP(), [t.NumberType()], t.NumberType())
ripl = get_ripl()
ripl.bind_foreign_sp('test', tester)
prog = '''
[ASSUME x (test 0)]
[INFER (mh default one 10)]'''
ripl.profiler_enable()
ripl.execute_program(prog)
def testGPLogscore1():
# Is this actually a valid test? The real solution to this problem
# (and to the corresponding bug with unincorporate) is to wrap the
# gp in a mem. This could be done automatically I suppose, or better
# through a library function.
raise SkipTest(
"GP logDensity is broken for multiple samples of the same input.")
ripl = get_ripl()
prep_ripl(ripl)
ripl.assume('gp', '(exactly (make_gp zero sq_exp))')
ripl.predict('(gp (array 0 0))')
ripl.get_global_logscore()
def test_gradients(seed):
ripl = get_ripl(seed=seed)
ripl.assume('mu_0', '(normal 0 1)')
ripl.assume('mean', '(gp_mean_const mu_0)')
ripl.assume('gs_expon_1',
'(lambda () (- 0. (log_logistic (log_odds_uniform))))')
ripl.assume('s2', '(gs_expon_1)')
ripl.assume('alpha', '(gs_expon_1)')
ripl.assume('cov', '(gp_cov_scale s2 (gp_cov_se alpha))')
ripl.assume('gp', '(make_gp mean cov)')
ripl.observe('(gp 0)', '1')
ripl.observe('(gp 1)', '2')
ripl.observe('(gp 2)', '4')
ripl.observe('(gp 3)', '8')
ripl.infer('(gradient_ascent default one 0.01 5 5)')
def testWishartPrior1(seed):
# Confirm that the diagonal elements of a Wishart are a chi-squared
# distribution.
if inParallel() and backend_name() == "puma":
raise SkipTest("The Lite SPs in Puma interface is not thread-safe, and wishart comes from Lite.")
ripl = get_ripl(seed=seed)
ripl.assume("s", "(matrix (array (array 2 -1) (array -1 3)))")
ripl.assume("m", "(wishart s 5)")
ripl.predict("(lookup m (pair 0 0))", label="prediction")
predictions = collectSamples(ripl, "prediction")
cdf = scipy.stats.chi2(df=5, scale=2).cdf
return reportKnownContinuous(cdf, predictions)
def testInvWishartPrior2(seed):
# Confirm that the diagonal elements of an inverse Wishart are an
# inverse Gamma distribution.
if inParallel() and backend_name() == "puma":
raise SkipTest("The Lite SPs in Puma interface is not thread-safe, and wishart comes from Lite.")
ripl = get_ripl(seed=seed)
ripl.assume("s", "(matrix (array (array 2 -1) (array -1 3)))")
ripl.assume("m", "(inv_wishart s 4.2)")
ripl.predict("(lookup m (pair 1 1))", label="prediction")
predictions = collectSamples(ripl, "prediction")
cdf = scipy.stats.invgamma(a=1.6, scale=1.5).cdf
return reportKnownContinuous(cdf, predictions)
def testReferences1(seed):
# Checks that the program runs without crashing. At some point, this
# program caused the old CXX backend to fire an assert. When the
# (flip) had a 0.0 or 1.0 it didn't fail.
ripl = get_ripl(seed=seed)
ripl.assume("draw_type0", "(make_crp 1.0)")
ripl.assume("draw_type1", "(if (flip) draw_type0 (lambda () atom<1>))")
ripl.assume("draw_type2", "(make_dir_cat (array 1.0 1.0))")
ripl.assume("class", "(if (flip) (lambda (name) (draw_type1)) (lambda (name) (draw_type2)))")
ripl.predict("(class 1)")
ripl.predict("(flip)", label="pid")
predictions = collectSamples(ripl,"pid")
ans = [(True,0.5), (False,0.5)]
return reportKnownDiscrete(ans, predictions)
def checkSliceNormalWithObserve2a(slice_method, seed):
# Checks the posterior distribution on a Gaussian given an unlikely
# observation. The difference between this and 1 is an extra
# predict, which apparently has a deleterious effect on mixing.
if (backend_name() != "lite") and (slice_method == 'slice_doubling'):
raise SkipTest(
"Slice sampling with doubling only implemented in Lite.")
ripl = get_ripl(seed=seed)
ripl.assume("a", "(normal 10.0 1.0)", label="pid")
ripl.observe("(normal a 1.0)", 14.0)
# Posterior for a is normal with mean 12, precision 2
ripl.predict("(normal a 1.0)")
predictions = myCollectSamples(ripl, slice_method)
return reportKnownGaussian(12, math.sqrt(0.5), predictions)
def testCollectLogScore():
# In the presence of likelihood-free SP's, the calling "collect" or
# "printf" should not crash the program.
class TestPSP(LikelihoodFreePSP):
def simulate(self, args):
x = args.operandValues()[0]
return x + stats.distributions.norm.rvs()
tester = typed_nr(TestPSP(), [t.NumberType()], t.NumberType())
ripl = get_ripl()
ripl.bind_foreign_sp('test', tester)
prog = '''
[ASSUME x (test 0)]
[ASSUME y (normal x 1)]
[infer (collect x)]'''
ripl.execute_program(prog)
def checkMakeSymDirCatAppControlsFlip(maker_1, maker_2, seed):
# Two AAA SPs with same parameters, where their applications control
# which are applied
ripl = get_ripl(seed=seed)
ripl.assume("a", "(normal 10.0 1.0)")
ripl.assume("f", "(%s a 4)" % maker_1)
ripl.assume("g", "(%s a 4)" % maker_2)
ripl.predict("(f)", label="pid")
ripl.predict("(g)")
for _ in range(5):
ripl.observe("(g)", "integer<1>")
ripl.predict("(if (eq (f) integer<1>) (g) (g))")
ripl.predict("(if (eq (g) integer<1>) (f) (f))")
return checkDirichletCategoricalAAA(ripl, "pid", infer="mixes_slowly")
def testPrintf2():
# Intercept stdout and make sure the message read what we expect
ripl = get_ripl()
pattern = make_pattern()
ripl.infer('(resample 2)')
ripl.assume('x', 2.1)
old_stdout = sys.stdout
result = StringIO()
sys.stdout = result
ripl.infer(
'(repeat 2 (do (resimulation_mh default one 1) (printf (run (collect x (labelled 3.1 foo))))))'
)
sys.stdout = old_stdout
res = result.getvalue()
assert pattern.match(res) is not None
def testBernoulliIfNormal2(seed):
# A simple program with bernoulli, if, and an absorbing application of normal
if not collect_iid_samples():
raise SkipTest("This test should not pass without reset.")
ripl = get_ripl(seed=seed)
ripl.assume("b", "(bernoulli 0.3)")
ripl.predict("(normal (if b 0.0 10.0) 1.0)", label="pid")
predictions = collectSamples(ripl, "pid")
def cdf(x):
return 0.3 * stats.norm.cdf(x, loc=0, scale=1) + \
0.7 * stats.norm.cdf(x, loc=10, scale=1)
return reportKnownContinuous(cdf, predictions, "0.3*N(0,1) + 0.7*N(10,1)")
def testBinomial2(seed):
# A simple test that checks the binomial logdensity
ripl = get_ripl(seed=seed)
b = 0.7
p1 = 0.3
p2 = 0.4
n = 4
ripl.assume("p","(if (flip %f) %f %f)" % (b,p1,p2))
ripl.predict("(binomial %d p)" % n,label="pid")
predictions = collectSamples(ripl,"pid")
ans = [(x,b * scipy.stats.binom.pmf(x,n,p1) + (1 - b) * scipy.stats.binom.pmf(x,n,p2)) for x in range(n+1)]
assert_almost_equal(sum([xx[1] for xx in ans]),1)
return reportKnownDiscrete(ans, predictions)
def initBasicPFripl2(seed):
ripl = get_ripl(seed=seed)
ripl.assume(
"f", """
(mem (lambda (i)
(tag 0 i
(normal (if (eq i 0) 0 (f (- i 1))) 1))))
""")
ripl.assume("g", """
(mem (lambda (i)
(normal (f i) 1.0)))
""")
return ripl
def testDefaultTrainer():
# raise(SkipTest('for quick results.'))
"Train a proposal to inverse a joint Gaussian model, using default (target) trainer"
ripl = get_ripl()
ripl.assume("a", "(scope_include (quote A) 0 (normal 0.0 1.0))", label="pid")
ripl.observe("(scope_include (quote B) 0 (normal (+ (* a 2) 2) 1.0))", 4.5)
ripl.register_proposal_program_class("LinearRegressionProposalProgram", LinearRegressionProposalProgram)
# Posterior for a is normal with mean 1.0, precision 5
proposal_src = """
[declare {
"name":"linreg",
"class":"LinearRegressionProposalProgram",
"conditioned":[["B",0]],
"target":[["A",0]],
"num_samples":1000}]
"""
ripl.execute_program(proposal_src)
predictions = collectSamples(ripl,"pid",infer="(custommh linreg aux 1 5)")
cdf = stats.norm(loc=1.0, scale=math.sqrt(0.2)).cdf
return reportKnownContinuous(cdf, predictions, "N(1.0,sqrt(0.2))")
def testCustommhSaveLoad():
# raise(SkipTest('for quick results.'))
"Checks the posterior distribution on a Gaussian given an unlikely observation"
ripl = get_ripl()
ripl.assume("a", "(scope_include (quote A) 0 (normal 10.0 1.0))", label="pid")
ripl.observe("(scope_include (quote B) 0 (normal a 1.0))", 20.0)
MyNormalExactPosterior = partial(NormalExactPosterior, mu1=10.0, sigma1=1.0, sigma2=1.0)
ripl.register_proposal_program_class("MyNormalExactPosterior", MyNormalExactPosterior)
# Posterior for a is normal with mean 15, precision 2
# ripl.predict("(normal a 1.0)")
proposal_src_1 = """
[declare {
"name":"normalexact1",
"class":"MyNormalExactPosterior",
"conditioned":[["B",0]],
"target":[["A",0]],
"ready":"yes",
"save_to":"test_custommh_save_load.npy",
"num_samples":0}]
"""
ripl.execute_program(proposal_src_1)
proposal_src_2 = """
[declare {
"name":"normalexact2",
"class":"MyNormalExactPosterior",
"conditioned":[["B",0]],
"target":[["A",0]],
"load_from":"test_custommh_save_load.npy",
"num_samples":0}]
"""
ripl.execute_program(proposal_src_2)
predictions = collectSamples(ripl,"pid",infer="(custommh normalexact2 assumed_gibbs 1 1)")
cdf = stats.norm(loc=15, scale=math.sqrt(0.5)).cdf
return reportKnownContinuous(cdf, predictions, "N(15,sqrt(0.5))")