class Gradient_Tests(unittest.TestCase):
"""
Check that our gradients match empirical gradients
"""
def setUp(self):
self.s1 = 'c a t a'
self.s2 = 'g a t t a'
self.X1 = np.array([[self.s1]])
self.X2 = np.array([[self.s2]])
self.alphabet = ['a', 'c', 'g', 't']
self.maxlen = 28
def test_gradient_gap_no_norm(self):
self.kern = StringKernel(implementation="numpy",
alphabet=self.alphabet,
order_coefs=[0.1, 0.1],
normalize=False,
gap_decay=0.8,
match_decay=0.8,
maxlen=self.maxlen)
result = self.kern.K(self.X1, self.X2)
true_grads = self.kern.gap_grads
E = 1e-4
self.kern.gap_decay.constrain_fixed(0.8 + E)
g_result1 = self.kern.K(self.X1, self.X2)
self.kern.gap_decay.constrain_fixed(0.8 - E)
g_result2 = self.kern.K(self.X1, self.X2)
g_result = (g_result1 - g_result2) / (2 * E)
self.assertAlmostEqual(true_grads[0][0], g_result[0][0], places=2)
def test_gradient_gap_norm(self):
self.kern = StringKernel(implementation="numpy",
alphabet=self.alphabet,
order_coefs=[0.1, 0.1],
normalize=True,
gap_decay=0.8,
match_decay=0.8,
maxlen=self.maxlen)
result = self.kern.K(self.X1, self.X2)
true_grads = self.kern.gap_grads
E = 1e-4
self.kern.gap_decay.constrain_fixed(0.8 + E)
g_result1 = self.kern.K(self.X1, self.X2)
self.kern.gap_decay.constrain_fixed(0.8 - E)
g_result2 = self.kern.K(self.X1, self.X2)
g_result = (g_result1 - g_result2) / (2 * E)
self.assertAlmostEqual(true_grads[0][0], g_result[0][0], places=2)
def test_gradient_coefs_dim_1_no_norm(self):
self.kern = StringKernel(implementation="numpy",
alphabet=self.alphabet,
order_coefs=[0.1, 0.1],
normalize=False,
gap_decay=0.8,
match_decay=0.8,
maxlen=self.maxlen)
result = self.kern.K(self.X1, self.X2)
true_grads = self.kern.coef_grads
E = 1e-4
self.kern.order_coefs.constrain_fixed([0.1 + E, 0.1])
g_result1 = self.kern.K(self.X1, self.X2)
self.kern.order_coefs.constrain_fixed([0.1 - E, 0.1])
g_result2 = self.kern.K(self.X1, self.X2)
g_result = (g_result1 - g_result2) / (2 * E)
self.assertAlmostEqual(true_grads[0][0][0], g_result[0][0], places=2)
def test_gradient_coefs_dim_2_no_norm(self):
self.kern = StringKernel(implementation="numpy",
alphabet=self.alphabet,
order_coefs=[0.1, 0.1],
normalize=False,
gap_decay=0.8,
match_decay=0.8,
maxlen=self.maxlen)
result = self.kern.K(self.X1, self.X2)
true_grads = self.kern.coef_grads
E = 1e-4
self.kern.order_coefs.constrain_fixed([0.1, 0.1 + E])
g_result1 = self.kern.K(self.X1, self.X2)
self.kern.order_coefs.constrain_fixed([0.1, 0.1 - E])
g_result2 = self.kern.K(self.X1, self.X2)
g_result = (g_result1 - g_result2) / (2 * E)
self.assertAlmostEqual(true_grads[0][0][1], g_result[0][0], places=2)
def test_gradient_coefs_dim_1_norm(self):
self.kern = StringKernel(implementation="numpy",
alphabet=self.alphabet,
order_coefs=[0.1, 0.1],
normalize=True,
gap_decay=0.8,
match_decay=0.8,
maxlen=self.maxlen)
result = self.kern.K(self.X1, self.X2)
true_grads = self.kern.coef_grads
E = 1e-4
self.kern.order_coefs.constrain_fixed([0.1 + E, 0.1])
g_result1 = self.kern.K(self.X1, self.X2)
self.kern.order_coefs.constrain_fixed([0.1 - E, 0.1])
g_result2 = self.kern.K(self.X1, self.X2)
g_result = (g_result1 - g_result2) / (2 * E)
self.assertAlmostEqual(true_grads[0][0][0], g_result[0][0], places=2)
def test_gradient_coefs_dim_2_norm(self):
self.kern = StringKernel(implementation="numpy",
alphabet=self.alphabet,
order_coefs=[0.1, 0.1],
normalize=True,
gap_decay=0.8,
match_decay=0.8,
maxlen=self.maxlen)
result = self.kern.K(self.X1, self.X2)
true_grads = self.kern.coef_grads
E = 1e-4
self.kern.order_coefs.constrain_fixed([0.1, 0.1 + E])
g_result1 = self.kern.K(self.X1, self.X2)
self.kern.order_coefs.constrain_fixed([0.1, 0.1 - E])
g_result2 = self.kern.K(self.X1, self.X2)
g_result = (g_result1 - g_result2) / (2 * E)
self.assertAlmostEqual(true_grads[0][0][1], g_result[0][0], places=2)
def test_gradient_match_no_norm(self):
self.kern = StringKernel(implementation="numpy",
alphabet=self.alphabet,
order_coefs=[0.1, 0.1],
normalize=False,
gap_decay=0.8,
match_decay=0.8,
maxlen=self.maxlen)
result = self.kern.K(self.X1, self.X2)
true_grads = self.kern.match_grads
E = 1e-4
self.kern.match_decay.constrain_fixed(0.8 + E)
g_result1 = self.kern.K(self.X1, self.X2)
self.kern.match_decay.constrain_fixed(0.8 - E)
g_result2 = self.kern.K(self.X1, self.X2)
g_result = (g_result1 - g_result2) / (2 * E)
self.assertAlmostEqual(true_grads[0][0], g_result[0][0], places=2)
def test_gradient_match_norm(self):
self.kern = StringKernel(implementation="numpy",
alphabet=self.alphabet,
order_coefs=[0.1, 0.1],
normalize=True,
gap_decay=0.8,
match_decay=0.8,
maxlen=self.maxlen)
result = self.kern.K(self.X1, self.X2)
true_grads = self.kern.match_grads
E = 1e-4
self.kern.match_decay.constrain_fixed(0.8 + E)
g_result1 = self.kern.K(self.X1, self.X2)
self.kern.match_decay.constrain_fixed(0.8 - E)
g_result2 = self.kern.K(self.X1, self.X2)
g_result = (g_result1 - g_result2) / (2 * E)
self.assertAlmostEqual(true_grads[0][0], g_result[0][0], places=2)
class Kern_Cal_Tests(unittest.TestCase):
"""
Test the kernel calculations on some simple hardcoded examples
Checking that shortcuts (e.g diag and sym) work properly
"""
def setUp(self):
self.s1 = 'c a t a'
self.s2 = 'g a t t a'
self.s3 = 'c g t a g c t a g c g a c g c a g c c a a t c g a t c g'
self.s4 = 'c g a g a t g c c a a t a g a g a g a g c g c t g t a'
self.X = np.array([[self.s1], [self.s2], [self.s3], [self.s4]])
self.alphabet = 'acgt'
self.maxlen = 28
def test_k_1(self):
self.kern = StringKernel(implementation="numpy",
alphabet=self.alphabet,
order_coefs=[1.] * 5,
normalize=False,
gap_decay=2.0,
match_decay=2.0,
maxlen=self.maxlen)
expected = 504.0
result = self.kern.K(self.X, self.X)
self.assertAlmostEqual(result[0][1], expected)
def test_k_2(self):
self.kern = StringKernel(implementation="numpy",
alphabet=self.alphabet,
order_coefs=[1.] * 5,
normalize=False,
gap_decay=0.8,
match_decay=0.8,
maxlen=self.maxlen)
expected = 5.943705
result = self.kern.K(self.X, self.X)
self.assertAlmostEqual(result[0][1], expected, places=4)
def test_diag_non_diag(self):
self.kern = StringKernel(implementation="numpy",
alphabet=self.alphabet,
order_coefs=[1.] * 5,
normalize=False,
gap_decay=2.0,
match_decay=2.0,
maxlen=self.maxlen)
result1 = np.diag(self.kern.K(self.X))
result2 = self.kern.Kdiag(self.X)
self.assertAlmostEqual(np.sum(result1) / 1000, np.sum(result2) / 1000)
def test_sym_non_sym(self):
self.kern = StringKernel(implementation="numpy",
alphabet=self.alphabet,
order_coefs=[1.] * 5,
normalize=False,
gap_decay=2.0,
match_decay=2.0,
maxlen=self.maxlen)
result1 = self.kern.K(self.X)
result2 = self.kern.K(self.X, self.X)
self.assertAlmostEqual(np.sum(result1) / 1000, np.sum(result2) / 1000)
def test_norm_diag(self):
self.kern = StringKernel(implementation="numpy",
alphabet=self.alphabet,
order_coefs=[1.] * 5,
normalize=True,
gap_decay=2.0,
match_decay=2.0,
maxlen=self.maxlen)
result = self.kern.Kdiag(self.X)
self.assertTrue(np.array_equal(result, np.ones((len(self.X)))))
def test_norm_no_diag(self):
self.kern = StringKernel(implementation="numpy",
alphabet=self.alphabet,
order_coefs=[1.] * 5,
normalize=True,
gap_decay=2.0,
match_decay=2.0,
maxlen=self.maxlen)
result = np.diag(self.kern.K(self.X))
self.assertTrue(np.array_equal(result, np.ones((len(self.X)))))