def get_weight_matrix(data_input: di.DataInput) -> tf.Tensor:
distances = dist.euclidian_distance(data_input.data_x_train, data_input.inducting_x_train)
reduce_min_1 = tf.reshape(tf.reduce_min(distances, axis=1), [-1, 1])
min_1_condition = distances == reduce_min_1
mask_1_distances = tf.reduce_max(distances) * tf.cast(min_1_condition, dtype=tf.float64)
distances_masked = distances + mask_1_distances
reduce_min_2 = tf.reshape(tf.reduce_min(distances_masked, axis=1), [-1, 1])
min_2_condition = distances_masked == reduce_min_2
weight_i = 1 - reduce_min_1 / (reduce_min_1 + reduce_min_2)
w_matrix = tf.zeros_like(distances) + weight_i * tf.cast(min_1_condition, dtype=tf.float64) + \
(1 - weight_i) * tf.cast(min_2_condition, dtype=tf.float64)
return w_matrix
def get_tf_tensor(self, hyper_parameter: List[tf.Tensor],
x_vector: np.ndarray,
x_vector_: np.ndarray) -> tf.Tensor:
assert x_vector is not None and x_vector_ is not None, "Input vectors x and x_ uninitialized: " + str(
self)
assert len(hyper_parameter) == self.get_number_of_hyper_parameter(
), "Invalid hyper_param size: " + str(self)
with tf.name_scope("SquaredExponentialKernel"):
dist: tf.Tensor = aux_dist.euclidian_distance(x_vector, x_vector_)
squared_distance: tf.Tensor = tf.square(dist,
name="squared_distance")
result: tf.Tensor = \
tf.math.exp(-0.5 * tf.math.divide(squared_distance, tf.square(hyper_parameter[0], name="l_squared")),
name="SquaredExponentialKernel")
if global_param.p_scaled_base_kernel:
result = tf.multiply(hyper_parameter[1], result)
self.last_hyper_parameter = hyper_parameter
return result