def test_dldtheta(self):
OK = np.zeros(5)
for i in range(5):
self.a = 1.0 * rand()
self.b = self.a + 10.0 * rand()
self.mu = (self.a + 10 * rand()) / 2.
self.s = 2.0 * rand() + 1.0
self.p = Distributions.TruncatedGaussian({
'a': self.a,
'b': self.b,
'mu': self.mu,
'sigma': self.s
})
p = self.p.copy()
p.primary = ['mu', 'sigma']
dat = p.sample(100)
def f(arr):
p.array2primary(arr)
return np.sum(p.loglik(dat))
def df(arr):
p.array2primary(arr)
return np.sum(p.dldtheta(dat), axis=1)
arr0 = p.primary2array()
arr0 = abs(np.random.randn(len(arr0)))
err = optimize.check_grad(f, df, arr0)
if err < 1e-02:
OK[i] = 1
M = np.max(OK)
self.assertTrue(
M > 0.5, 'Gradient error %.4g is greater than %.4g' % (err, 1e-02))
def setUp(self):
self.a = 1.0 * rand()
self.b = self.a + 10.0 * rand()
self.mu = (self.a + 10 * rand()) / 2.
self.s = 2.0 * rand() + 1.0
self.p = Distributions.TruncatedGaussian({
'a': self.a,
'b': self.b,
'mu': self.mu,
'sigma': self.s
})