def test_ask_tell_flow(self):
# Tests the ask-and-tell pattern
# Create problem
x0 = np.array([2, 4])
log_pdf = pints.toy.GaussianLogPDF([2, 4], [[1, 0], [0, 3]])
# Ask must return initial point
mcmc = pints.SliceStepoutMCMC(x0)
self.assertTrue(np.all(mcmc.ask() == x0))
# Ask can't be called twice
self.assertRaises(RuntimeError, mcmc.ask)
# Tell
fx0 = log_pdf(x0)
x1, fx1, ac = mcmc.tell(fx0)
self.assertTrue(np.all(x0 == x1))
self.assertTrue(fx0 == fx1)
# Check that the new slice has been constructed appropriately
self.assertTrue(mcmc.current_slice_height() < fx1)
# Can't tell twice
self.assertRaises(RuntimeError, mcmc.tell, log_pdf(x0))
# First point must be finite
mcmc = pints.SliceStepoutMCMC(x0)
mcmc.ask()
self.assertRaises(ValueError, mcmc.tell, np.inf)
def test_run(self):
# Test a short run
# Create log pdf
log_pdf = pints.toy.MultimodalGaussianLogPDF(
modes=[[0, 2], [0, 7], [5, 0], [4, 4]])
# Create and configure sampler
mcmc = pints.SliceStepoutMCMC(np.array([1, 1]))
mcmc.set_prob_overrelaxed(0.6)
mcmc.set_width(30)
mcmc.set_bisection_steps(2)
# Run a few iterations of the sampler
np.random.seed(456)
n = 20
chain = []
while len(chain) < n:
x = mcmc.ask()
fx = log_pdf(x)
reply = mcmc.tell(fx)
if reply is not None:
y, fy, ac = reply
chain.append(y)
self.assertEqual(fy, log_pdf(y))
self.assertEqual(np.shape(chain), (n, 2))
def test_bisection(self):
# Test bisection methods
mcmc = pints.SliceStepoutMCMC(np.array([1, 2]))
mcmc.set_bisection_steps(40)
self.assertEqual(mcmc.bisection_steps(), 40)
with self.assertRaises(ValueError):
mcmc.set_bisection_steps(-30)
def test_expansion(self):
# Test expansion step methods
mcmc = pints.SliceStepoutMCMC(np.array([1, 2]))
mcmc.set_expansion_steps(3)
self.assertEqual(mcmc.expansion_steps(), 3.)
with self.assertRaises(ValueError):
mcmc.set_expansion_steps(-1)
def test_hyper_parameters(self):
# Test the hyper parameter interface
mcmc = pints.SliceStepoutMCMC(np.array([1, 2]))
self.assertEqual(mcmc.n_hyper_parameters(), 4)
mcmc.set_hyper_parameters([3, 100, .7, 50])
self.assertTrue((np.all(mcmc.width() == np.array([3, 3]))))
self.assertEqual(mcmc.expansion_steps(), 100)
self.assertEqual(mcmc.prob_overrelaxed(), 0.7)
self.assertEqual(mcmc.bisection_steps(), 50)
def test_overrelaxed(self):
# Test overrelaxation methods
mcmc = pints.SliceStepoutMCMC(np.array([1, 2]))
mcmc.set_prob_overrelaxed(0.5)
self.assertEqual(mcmc.prob_overrelaxed(), 0.5)
with self.assertRaises(ValueError):
mcmc.set_prob_overrelaxed(-1)
with self.assertRaises(ValueError):
mcmc.set_prob_overrelaxed(4)
def test_first_run(self):
# Tests the very first run of the sampler.
# Create log pdf
log_pdf = pints.toy.GaussianLogPDF([2, 4], [[1, 0], [0, 3]])
# Create mcmc
x0 = np.array([2., 4.])
mcmc = pints.SliceStepoutMCMC(x0)
# Ask() should fail if _ready_for_tell flag is True
x = mcmc.ask()
with self.assertRaises(RuntimeError):
mcmc.ask()
# Check whether first iteration of ask() returns x0
self.assertTrue(np.all(x == x0))
# Calculate log pdf for x0
fx = log_pdf.evaluateS1(x0)[0]
# Test first iteration of tell(). The first point in the chain
# should be x0
sample = mcmc.tell(fx)
self.assertTrue(np.all(sample == x0))
# We update the _current_log_pdf value used to generate the new slice
self.assertEqual(mcmc.current_log_pdf(), fx)
# Check that the new slice has been constructed appropriately
self.assertTrue(mcmc.current_slice_height() < mcmc.current_log_pdf())
# Tell() should fail when fx is infinite
mcmc = pints.SliceStepoutMCMC(x0)
mcmc.ask()
with self.assertRaises(ValueError):
fx = np.inf
mcmc.tell(fx)
# Tell() should fail when _ready_for_tell is False
with self.assertRaises(RuntimeError):
mcmc.tell(fx)
def test_width(self):
# Test width methods
mcmc = pints.SliceStepoutMCMC(np.array([1, 2]))
# Test set_width(), width()
mcmc.set_width(2)
self.assertTrue(np.all(mcmc.width() == np.array([2, 2])))
mcmc.set_width([3, 5])
self.assertTrue(np.all(mcmc.width() == np.array([3, 5])))
with self.assertRaises(ValueError):
mcmc.set_width(-1)
with self.assertRaises(ValueError):
mcmc.set_width([3, 3, 3, 3])
def test_basic(self):
# Test basic methods of the class.
# Create mcmc
x0 = np.array([1, 1])
mcmc = pints.SliceStepoutMCMC(x0)
# Test name
self.assertEqual(mcmc.name(), 'Slice Sampling - Stepout')
# Test set_width(), width()
mcmc.set_width(2)
self.assertTrue(np.all(mcmc.width() == np.array([2, 2])))
mcmc.set_width([3, 5])
self.assertTrue(np.all(mcmc.width() == np.array([3, 5])))
with self.assertRaises(ValueError):
mcmc.set_width(-1)
with self.assertRaises(ValueError):
mcmc.set_width([3, 3, 3, 3])
# Test set_expansion_steps(), expansion_steps()
mcmc.set_expansion_steps(3)
self.assertEqual(mcmc.expansion_steps(), 3.)
with self.assertRaises(ValueError):
mcmc.set_expansion_steps(-1)
# Test set_prob_overrelaxed(), prob_overrelaxed()
mcmc.set_prob_overrelaxed(.5)
self.assertEqual(mcmc.prob_overrelaxed(), .5)
with self.assertRaises(ValueError):
mcmc.set_prob_overrelaxed(-1)
with self.assertRaises(ValueError):
mcmc.set_prob_overrelaxed(4)
# Test set_bisection_steps(), bisection_steps()
mcmc.set_bisection_steps(40)
self.assertEqual(mcmc.bisection_steps(), 40)
with self.assertRaises(ValueError):
mcmc.set_bisection_steps(-30)
# Test number of hyperparameters
self.assertEqual(mcmc.n_hyper_parameters(), 4)
# Test setting hyperparameters
mcmc.set_hyper_parameters([3, 100, .7, 50])
self.assertTrue((np.all(mcmc.width() == np.array([3, 3]))))
self.assertEqual(mcmc.expansion_steps(), 100)
self.assertEqual(mcmc.prob_overrelaxed(), 0.7)
self.assertEqual(mcmc.bisection_steps(), 50)
def test_multimodal_overrelaxed_run(self):
# Create log pdf
log_pdf = pints.toy.MultimodalGaussianLogPDF(
modes=[[0, 2], [0, 7], [5, 0], [4, 4]])
# Create mcmc
x0 = np.array([1, 1])
mcmc = pints.SliceStepoutMCMC(x0)
mcmc.set_prob_overrelaxed(0.6)
mcmc.set_width(30)
mcmc.set_bisection_steps(2)
# Run multiple iterations of the sampler
chain = []
while len(chain) < 100:
x = mcmc.ask()
fx = log_pdf.evaluateS1(x)[0]
sample = mcmc.tell(fx)
if sample is not None:
chain.append(np.copy(sample))
self.assertEqual(np.shape(chain), (100, 2))
def test_name(self):
# Tests the sampler's name
mcmc = pints.SliceStepoutMCMC(np.array([1, 2]))
self.assertEqual(mcmc.name(), 'Slice Sampling - Stepout')
