def __init__()

in ma_policy/normalizers.py [0:0]

• 26 lines of code
• 3 McCabe index (conditional complexity)

    def __init__(self, beta, scope="ema", reuse=None, epsilon=1e-6, per_element_update=False, shape=(), version=1):
        self._version = version
        self._per_element_update = per_element_update
        with tf.variable_scope(scope, reuse=reuse):
            # Expected value of x
            self._biased_mean = tf.get_variable(
                dtype=tf.float32,
                shape=shape,
                initializer=tf.constant_initializer(0.0),
                name="mean",
                trainable=False)
            # Expected value of x^2
            self._biased_sq = tf.get_variable(
                dtype=tf.float32,
                shape=shape,
                initializer=tf.constant_initializer(0.0),
                name="sq",
                trainable=False)
            # How to integrate observations of x over time
            self._one_minus_beta = 1.0 - beta
            # Weight placed on ema[-1] == 0.0 which we divide out to debias
            self._debiasing_term = tf.get_variable(
                dtype=tf.float32,
                shape=shape,
                initializer=tf.constant_initializer(0.0),
                name="debiasing_term",
                trainable=False)
            self.shape = shape

            # the stored mean and square are biased due to setting ema[-1] = 0.0,
            # we correct for this by dividing by the debiasing term:
            self.mean = self._biased_mean / tf.maximum(self._debiasing_term, epsilon)
            self.std = _std_from_mean_and_square(mean=self.mean, square=self._biased_sq / tf.maximum(self._debiasing_term, epsilon))