def testAInvUpdateEmptyObservations(self, batch_size, context_dim):
a_array = 2 * np.eye(context_dim) + np.array(
range(context_dim * context_dim)).reshape(
(context_dim, context_dim))
a_array = a_array + a_array.T
a_inv_array = np.linalg.inv(a_array)
expected_a_inv_update_array = np.zeros([context_dim, context_dim],
dtype=np.float32)
a_inv = tf.constant(a_inv_array,
dtype=tf.float32,
shape=[context_dim, context_dim])
x = tf.constant([], dtype=tf.float32, shape=[0, context_dim])
a_inv_update = linalg.update_inverse(a_inv, x)
self.assertAllClose(expected_a_inv_update_array,
self.evaluate(a_inv_update))
def testAInvUpdate(self, batch_size, context_dim):
a_array = 2 * np.eye(context_dim) + np.array(
range(context_dim * context_dim)).reshape(
(context_dim, context_dim))
a_array = a_array + a_array.T
a_inv_array = np.linalg.inv(a_array)
x_array = np.array(range(batch_size * context_dim)).reshape(
(batch_size, context_dim))
expected_a_inv_updated_array = np.linalg.inv(
a_array + np.matmul(np.transpose(x_array), x_array))
a_inv = tf.constant(a_inv_array,
dtype=tf.float32,
shape=[context_dim, context_dim])
x = tf.constant(x_array,
dtype=tf.float32,
shape=[batch_size, context_dim])
a_inv_update = linalg.update_inverse(a_inv, x)
self.assertAllClose(expected_a_inv_updated_array,
self.evaluate(a_inv + a_inv_update))