def testLogAddSub(self):
for i in range(100):
f1 = kaldi_math.rand() % 10000
f2 = kaldi_math.rand() % 20
add1 = kaldi_math.exp(
kaldi_math.log_add(kaldi_math.log(f1), kaldi_math.log(f2)))
add2 = kaldi_math.exp(
kaldi_math.log_add(kaldi_math.log(f2), kaldi_math.log(f1)))
add = f1 + f2
thresh = add * 0.00001
self.assertAlmostEqual(add, add1, delta=thresh)
self.assertAlmostEqual(add, add2, delta=thresh)
try:
f2_check = kaldi_math.exp(
kaldi_math.log_sub(kaldi_math.log(add),
kaldi_math.log(f1)))
self.assertAlmostEqual(f2, f2_check, delta=thresh)
except:
# self.assertEqual(0, f2)
pass
def testDefines(self):
self.assertAlmostEqual(0.0,
kaldi_math.exp(kaldi_math.K_LOG_ZERO_FLOAT))
self.assertAlmostEqual(0.0,
kaldi_math.exp(kaldi_math.K_LOG_ZERO_DOUBLE))
# TODO:
# How to test these in Python?
# den = 0.0
# self.assertTrue(kaldi_math.KALDI_ISNAN(0.0 / den))
# self.assertFalse(kaldi_math.KALDI_ISINF(0.0 / den))
# self.assertFalse(kaldi_math.KALDI_ISFINITE(0.0 / den))
# self.assertFalse(kaldi_math.KALDI_ISNAN(1.0 / den))
# self.assertTrue(kaldi_math.KALDI_ISINF(1.0 / den))
# self.assertFalse(kaldi_math.KALDI_ISFINITE(1.0 / den))
# self.assertTrue(kaldi_math.KALDI_ISFINITE(0.0))
# self.assertFalse(kaldi_math.KALDI_ISINF(0.0))
# self.assertFalse(kaldi_math.KALDI_ISNAN(0.0))
self.assertTrue(1.0 != 1.0 + kaldi_math.DBL_EPSILON)
self.assertTrue(1.0 == 1.0 + 0.5 * kaldi_math.DBL_EPSILON)
# self.assertNotAlmostEqual(1.0, 1.0 + kaldi_math.FLT_EPSILON, places = 7)
# self.assertAlmostEqual(1.0, 1.0 + 0.5 * kaldi_math.FLT_EPSILON, places = 6)
self.assertAlmostEqual(0.0, math.fabs(math.sin(kaldi_math.M_PI)))
self.assertAlmostEqual(1.0, math.fabs(math.cos(kaldi_math.M_PI)))
self.assertAlmostEqual(0.0, math.fabs(math.sin(kaldi_math.M_2PI)))
self.assertAlmostEqual(1.0, math.fabs(math.cos(kaldi_math.M_2PI)))
self.assertAlmostEqual(
0.0,
math.fabs(math.sin(kaldi_math.exp(kaldi_math.M_LOG_2PI))),
places=5)
self.assertAlmostEqual(
1.0,
math.fabs(math.cos(kaldi_math.exp(kaldi_math.M_LOG_2PI))),
places=5)
def showPerplexity(fst, data):
num_phones = 0
tot_loglike = 0.0
for i in data:
num_phones += len(i)
linear_fst = StdVectorFst()
make_linear_acceptor(i, linear_fst)
composed_fst = compose(linear_fst, fst)
weight = shortestdistance(composed_fst)[-1]
tot_loglike -= weight.value
perplexity = kaldi_math.exp(-(tot_loglike / num_phones))
print("Perplexity over {} phones (of training data) is {}".format(
num_phones, perplexity))
def init_rand_diag_gmm(gmm):
num_comp, dim = gmm.num_gauss(), gmm.dim()
weights = Vector([kaldi_math.rand_uniform() for _ in range(num_comp)])
tot_weigth = weights.sum()
for i, m in enumerate(weights):
weights[i] = m / tot_weigth
means = Matrix([[kaldi_math.rand_gauss() for _ in range(dim)] for _ in range(num_comp)])
vars_ = Matrix([[kaldi_math.exp(kaldi_math.rand_gauss()) for _ in range(dim)] for _ in range(num_comp)])
vars_.invert_elements_()
gmm.set_weights(weights)
gmm.set_inv_vars_and_means(vars_, means)
gmm.perturb(0.5 * kaldi_math.rand_uniform())
gmm.compute_gconsts()