recommenders/models/deeprec/models/sequential/sum_cells.py [193:228]:
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
        n_a, n_b = tf.nn.l2_normalize(last, 1), tf.nn.l2_normalize(inputs, 1)
        dist = tf.expand_dims(tf.reduce_sum(input_tensor=n_a * n_b, axis=1), 1)
        dist = tf.math.pow(self._alpha, dist)

        att_weights = att_weights * tf.expand_dims(dist, 1)

        reset = tf.sigmoid(
            tf.compat.v1.nn.xw_plus_b(
                tf.concat([inputs, h_hat], axis=-1), self._reset_W, self._reset_b
            )
        )
        erase = tf.sigmoid(
            tf.compat.v1.nn.xw_plus_b(
                tf.concat([inputs, h_hat], axis=-1), self._erase_W, self._erase_b
            )
        )
        add = tf.tanh(
            tf.compat.v1.nn.xw_plus_b(
                tf.concat([inputs, reset * h_hat], axis=-1), self._add_W, self._add_b
            )
        )

        start_part01 = state[:, : self._slots, :]
        state01 = start_part01 * (
            tf.ones_like(start_part01) - att_weights * tf.expand_dims(erase, 1)
        )
        state01 = state01 + att_weights * tf.expand_dims(erase, 1) * tf.expand_dims(
            add, 1
        )
        state01 = tf.reshape(state01, [-1, self._slots * self._real_units])

        start_part02 = state[:, self._slots, :]
        state02 = start_part02 * (tf.ones_like(start_part02) - dist * erase)
        state02 = state02 + dist * erase * add
        state = tf.concat([state01, state02, inputs], axis=-1)
        return state, state
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -



recommenders/models/deeprec/models/sequential/sum_cells.py [331:366]:
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
        n_a, n_b = tf.nn.l2_normalize(last, 1), tf.nn.l2_normalize(inputs, 1)
        dist = tf.expand_dims(tf.reduce_sum(input_tensor=n_a * n_b, axis=1), 1)
        dist = tf.math.pow(self._alpha, dist)

        att_weights = att_weights * tf.expand_dims(dist, 1)

        reset = tf.sigmoid(
            tf.compat.v1.nn.xw_plus_b(
                tf.concat([inputs, h_hat], axis=-1), self._reset_W, self._reset_b
            )
        )
        erase = tf.sigmoid(
            tf.compat.v1.nn.xw_plus_b(
                tf.concat([inputs, h_hat], axis=-1), self._erase_W, self._erase_b
            )
        )
        add = tf.tanh(
            tf.compat.v1.nn.xw_plus_b(
                tf.concat([inputs, reset * h_hat], axis=-1), self._add_W, self._add_b
            )
        )

        start_part01 = state[:, : self._slots, :]
        state01 = start_part01 * (
            tf.ones_like(start_part01) - att_weights * tf.expand_dims(erase, 1)
        )
        state01 = state01 + att_weights * tf.expand_dims(erase, 1) * tf.expand_dims(
            add, 1
        )
        state01 = tf.reshape(state01, [-1, self._slots * self._real_units])

        start_part02 = state[:, self._slots, :]
        state02 = start_part02 * (tf.ones_like(start_part02) - dist * erase)
        state02 = state02 + dist * erase * add
        state = tf.concat([state01, state02, inputs], axis=-1)
        return state, state
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -