apunkt/GPy
1
0
Fork 0
mirror of https://github.com/SheffieldML/GPy.git synced 2026-05-07 02:52:40 +02:00
Code Issues Projects Releases Packages Wiki Activity Actions

more transformation checks

Browse source 
Need to check, which ones can be checked by kde?
This commit is contained in:
Max Zwiessele 2015-09-12 18:41:40 +01:00
parent b8b7ca1e17
commit 39e7e18c8b
1 changed files with 37 additions and 20 deletions
Download patch file Download diff file Expand all files Collapse all files

57
GPy/testing/rv_transformation_tests.py
View file

@ -24,24 +24,27 @@ class TestModel(GPy.core.Model):
class RVTransformationTestCase(unittest.TestCase): class RVTransformationTestCase(unittest.TestCase):
def _test_trans(self, trans): def _test_trans(self, trans, kde=True):
m = TestModel(trans.__class__.__name__) m = TestModel(trans.__class__.__name__)
prior = GPy.priors.LogGaussian(.5, 0.1) prior = GPy.priors.LogGaussian(.5, 0.1)
m.theta.set_prior(prior) m.theta.set_prior(prior)
m.theta.unconstrain() m.theta.unconstrain()
m.theta.constrain(trans) m.theta.constrain(trans)
# The PDF of the transformed variables np.random.seed(1234)
p_phi = lambda phi : np.exp(-m._objective_grads(phi)[0])
# To the empirical PDF of:
np.random.seed(0)
theta_s = prior.rvs(5e5) theta_s = prior.rvs(5e5)
phi_s = trans.finv(theta_s) if kde:
# which is essentially a kernel density estimation # The PDF of the transformed variables
kde = st.gaussian_kde(phi_s) p_phi = lambda phi : np.exp(-m._objective_grads(phi)[0])
# We will compare the PDF here: # To the empirical PDF of:
phi = np.linspace(phi_s.min(), phi_s.max(), 100) phi_s = trans.finv(theta_s)
# The transformed PDF of phi should be this: # which is essentially a kernel density estimation
pdf_phi = np.array([p_phi(p) for p in phi]) kde = st.gaussian_kde(phi_s)
# We will compare the PDF here:
phi = np.linspace(phi_s.min(), phi_s.max(), 100)
# The transformed PDF of phi should be this:
pdf_phi = np.array([p_phi(p) for p in phi])
# The following test cannot be very accurate
self.assertTrue(np.linalg.norm(pdf_phi - kde(phi)) / np.linalg.norm(kde(phi)) <= 1e-1)
# UNCOMMENT TO SEE GRAPHICAL COMPARISON # UNCOMMENT TO SEE GRAPHICAL COMPARISON
#import matplotlib.pyplot as plt #import matplotlib.pyplot as plt
#fig, ax = plt.subplots() #fig, ax = plt.subplots()
@ -53,8 +56,6 @@ class RVTransformationTestCase(unittest.TestCase):
#plt.legend(loc='best') #plt.legend(loc='best')
#plt.show(block=True) #plt.show(block=True)
# END OF PLOT # END OF PLOT
# The following test cannot be very accurate
self.assertTrue(np.linalg.norm(pdf_phi - kde(phi)) / np.linalg.norm(kde(phi)) <= 1e-1)
# Check the gradients at a few random points # Check the gradients at a few random points
for i in range(5): for i in range(5):
m.theta = theta_s[i] m.theta = theta_s[i]
@ -62,14 +63,30 @@ class RVTransformationTestCase(unittest.TestCase):
def test_Logexp(self): def test_Logexp(self):
self._test_trans(GPy.constraints.Logexp()) self._test_trans(GPy.constraints.Logexp())
def test_Exponent(self):
self._test_trans(GPy.constraints.Exponent()) self._test_trans(GPy.constraints.Exponent())
self._test_trans(GPy.constraints.LogexpNeg())
self._test_trans(GPy.constraints.NegativeLogexp()) def test_Logexpneg(self):
self._test_trans(GPy.constraints.LogexpNeg(False))
def test_neglogexp(self):
self._test_trans(GPy.constraints.NegativeLogexp(False))
def test_logexpclipped(self):
self._test_trans(GPy.constraints.LogexpClipped()) self._test_trans(GPy.constraints.LogexpClipped())
self._test_trans(GPy.constraints.NegativeExponent())
self._test_trans(GPy.constraints.LogexpNeg()) def test_NegExp(self):
self._test_trans(GPy.constraints.Square()) self._test_trans(GPy.constraints.NegativeExponent(False))
self._test_trans(GPy.constraints.Logistic())
def test_logexpNeg(self):
self._test_trans(GPy.constraints.LogexpNeg(False))
def test_Square(self):
self._test_trans(GPy.constraints.Square(False))
def test_logistic(self):
self._test_trans(GPy.constraints.Logistic(False))
if __name__ == '__main__': if __name__ == '__main__':
unittest.main() unittest.main()