def testParallelVarianceOneAtATime(self):
x = np.random.randn(5, 10)
n, avg, m2 = 1, x[0], 0
for row in range(1, 5):
n, avg, m2 = tensor_normalizer.parallel_variance_calculation(
n_a=n, avg_a=avg, m2_a=m2, n_b=1, avg_b=x[row], m2_b=0)
var = m2 / n
self.assertAllClose(avg, x.mean(axis=0))
self.assertAllClose(var, x.var(axis=0))
def testParallelVarianceForOneGroup(self):
x = tf.constant(np.random.randn(5, 10))
n = 5
avg, var = tf.nn.moments(x, axes=[0])
m2 = var * n
new_n, new_avg, new_m2, _ = tensor_normalizer.parallel_variance_calculation(
n, avg, m2, n_b=0, avg_b=0, m2_b=0, m2_b_c=0)
new_var = new_m2 / n
(avg, var, new_avg, new_var) = self.evaluate(
(avg, var, new_avg, new_var))
self.assertEqual(new_n, 5)
self.assertAllClose(new_avg, avg)
self.assertAllClose(new_var, var)
def testParallelVarianceCombinesGroups(self):
x1 = tf.constant(np.random.randn(5, 10))
x2 = tf.constant(np.random.randn(15, 10))
n1 = 5
n2 = 15
avg1, var1 = tf.nn.moments(x1, axes=[0])
avg2, var2 = tf.nn.moments(x2, axes=[0])
m2_1 = var1 * n1
m2_2 = var2 * n2
n, avg, m2 = tensor_normalizer.parallel_variance_calculation(
n1, avg1, m2_1, n2, avg2, m2_2)
var = m2 / n
avg_true, var_true = tf.nn.moments(tf.concat((x1, x2), axis=0),
axes=[0])
avg, var, avg_true, var_true = self.evaluate(
(avg, var, avg_true, var_true))
self.assertAllClose(avg, avg_true)
self.assertAllClose(var, var_true)
def testParallelVarianceCombinesGroups(self):
x1 = tf.constant(np.random.randn(5, 10))
x2 = tf.constant(np.random.randn(15, 10))
n1 = 5
n2 = 15
avg1 = tf.math.reduce_mean(x1, axis=[0])
m2_1 = tf.math.reduce_sum(tf.math.squared_difference(x1, avg1),
axis=[0])
avg2 = tf.math.reduce_mean(x2, axis=[0])
m2_2 = tf.math.reduce_sum(tf.math.squared_difference(x2, avg2),
axis=[0])
m2_c = m2_2 * 0.
n, avg, m2, _ = tensor_normalizer.parallel_variance_calculation(
n1, avg1, m2_1, n2, avg2, m2_2, m2_c)
var = m2 / n
avg_true, var_true = tf.nn.moments(tf.concat((x1, x2), axis=0),
axes=[0])
avg, var, avg_true, var_true = self.evaluate(
(avg, var, avg_true, var_true))
self.assertAllClose(avg, avg_true)
self.assertAllClose(var, var_true)