How to use the mlagents.tf_utils.tf.layers.dense function in mlagents

hidden_2,
                    self.z_size,
                    reuse=reuse,
                    name="gail_z_mean",
                    kernel_initializer=ModelUtils.scaled_init(0.01),
                )

                self.noise = tf.random_normal(tf.shape(z_mean), dtype=tf.float32)

                # Sampled latent code
                self.z = z_mean + self.z_sigma * self.noise * self.use_noise
                estimate_input = self.z
            else:
                estimate_input = hidden_2

            estimate = tf.layers.dense(
                estimate_input,
                1,
                activation=tf.nn.sigmoid,
                name="gail_d_estimate",
                reuse=reuse,
            )
            return estimate, z_mean, concat_input

name="gail_d_hidden_1",
                reuse=reuse,
            )

            hidden_2 = tf.layers.dense(
                hidden_1,
                self.h_size,
                activation=ModelUtils.swish,
                name="gail_d_hidden_2",
                reuse=reuse,
            )

            z_mean = None
            if self.use_vail:
                # Latent representation
                z_mean = tf.layers.dense(
                    hidden_2,
                    self.z_size,
                    reuse=reuse,
                    name="gail_z_mean",
                    kernel_initializer=ModelUtils.scaled_init(0.01),
                )

                self.noise = tf.random_normal(tf.shape(z_mean), dtype=tf.float32)

                # Sampled latent code
                self.z = z_mean + self.z_sigma * self.noise * self.use_noise
                estimate_input = self.z
            else:
                estimate_input = hidden_2

            estimate = tf.layers.dense(

axis=1,
            )
            hidden = tf.concat([hidden, prev_action_oh], axis=1)

            self.memory_in = tf.placeholder(
                shape=[None, self.m_size], dtype=tf.float32, name="recurrent_in"
            )
            hidden, memory_out = self.create_recurrent_encoder(
                hidden, self.memory_in, self.sequence_length
            )
            self.memory_out = tf.identity(memory_out, name="recurrent_out")

        policy_branches = []
        for size in self.act_size:
            policy_branches.append(
                tf.layers.dense(
                    hidden,
                    size,
                    activation=None,
                    use_bias=False,
                    kernel_initializer=LearningModel.scaled_init(0.01),
                )
            )

        self.all_log_probs = tf.concat(policy_branches, axis=1, name="action_probs")

        self.action_masks = tf.placeholder(
            shape=[None, sum(self.act_size)], dtype=tf.float32, name="action_masks"
        )
        output, _, normalized_logits = self.create_discrete_action_masking_layer(
            self.all_log_probs, self.action_masks, self.act_size
        )

self.policy_memory_in,
                self.sequence_length,
                name="lstm_policy",
            )
            self.policy_memory_out = memory_out
        with tf.variable_scope(scope):
            mu = tf.layers.dense(
                hidden_policy,
                self.act_size[0],
                activation=None,
                name="mu",
                kernel_initializer=LearningModel.scaled_init(0.01),
            )

            # Policy-dependent log_sigma_sq
            log_sigma_sq = tf.layers.dense(
                hidden_policy,
                self.act_size[0],
                activation=None,
                name="log_std",
                kernel_initializer=LearningModel.scaled_init(0.01),
            )

            self.log_sigma_sq = tf.clip_by_value(log_sigma_sq, LOG_STD_MIN, LOG_STD_MAX)

            sigma_sq = tf.exp(self.log_sigma_sq)

            # Do the reparameterization trick
            policy_ = mu + tf.random_normal(tf.shape(mu)) * sigma_sq

            _gauss_pre = -0.5 * (
                ((policy_ - mu) / (tf.exp(self.log_sigma_sq) + EPSILON)) ** 2

)

            hidden_value, memory_value_out = self.create_recurrent_encoder(
                hidden_streams[1],
                self.memory_in[:, _half_point:],
                self.sequence_length,
                name="lstm_value",
            )
            self.memory_out = tf.concat(
                [memory_policy_out, memory_value_out], axis=1, name="recurrent_out"
            )
        else:
            hidden_policy = hidden_streams[0]
            hidden_value = hidden_streams[1]

        mu = tf.layers.dense(
            hidden_policy,
            self.act_size[0],
            activation=None,
            kernel_initializer=LearningModel.scaled_init(0.01),
            reuse=tf.AUTO_REUSE,
        )

        self.log_sigma_sq = tf.get_variable(
            "log_sigma_squared",
            [self.act_size[0]],
            dtype=tf.float32,
            initializer=tf.zeros_initializer(),
        )

        sigma_sq = tf.exp(self.log_sigma_sq)

axis=1,
            )
            hidden_policy = tf.concat([hidden_policy, prev_action_oh], axis=1)

            hidden_policy, memory_out = self.create_recurrent_encoder(
                hidden_policy,
                self.policy_memory_in,
                self.sequence_length,
                name="lstm_policy",
            )
            self.policy_memory_out = memory_out
        with tf.variable_scope(scope):
            policy_branches = []
            for size in self.act_size:
                policy_branches.append(
                    tf.layers.dense(
                        hidden_policy,
                        size,
                        activation=None,
                        use_bias=False,
                        kernel_initializer=tf.initializers.variance_scaling(0.01),
                    )
                )
            all_logits = tf.concat(policy_branches, axis=1, name="action_probs")

            output, normalized_probs, normalized_logprobs = self.create_discrete_action_masking_layer(
                all_logits, self.action_masks, self.act_size
            )

            self.action_probs = normalized_probs

            # Really, this is entropy, but it has an analogous purpose to the log probs in the

:param reuse: Whether or not to reuse variables. Useful for creating Q of policy.
        :param num_outputs: Number of outputs of each Q function. If discrete, equal to number of actions.
        """
        with tf.variable_scope(self.join_scopes(scope, "q1_encoding"), reuse=reuse):
            q1_hidden = self.create_vector_observation_encoder(
                hidden_input, h_size, self.activ_fn, num_layers, "q1_encoder", reuse
            )
            if self.use_recurrent:
                q1_hidden, memory_out = self.create_recurrent_encoder(
                    q1_hidden, self.q1_memory_in, self.sequence_length, name="lstm_q1"
                )
                self.q1_memory_out = memory_out

            q1_heads = {}
            for name in stream_names:
                _q1 = tf.layers.dense(q1_hidden, num_outputs, name="{}_q1".format(name))
                q1_heads[name] = _q1

            q1 = tf.reduce_mean(list(q1_heads.values()), axis=0)
        with tf.variable_scope(self.join_scopes(scope, "q2_encoding"), reuse=reuse):
            q2_hidden = self.create_vector_observation_encoder(
                hidden_input, h_size, self.activ_fn, num_layers, "q2_encoder", reuse
            )
            if self.use_recurrent:
                q2_hidden, memory_out = self.create_recurrent_encoder(
                    q2_hidden, self.q2_memory_in, self.sequence_length, name="lstm_q2"
                )
                self.q2_memory_out = memory_out

            q2_heads = {}
            for name in stream_names:
                _q2 = tf.layers.dense(q2_hidden, num_outputs, name="{}_q2".format(name))

q1_heads[name] = _q1

            q1 = tf.reduce_mean(list(q1_heads.values()), axis=0)
        with tf.variable_scope(self.join_scopes(scope, "q2_encoding"), reuse=reuse):
            q2_hidden = self.create_vector_observation_encoder(
                hidden_input, h_size, self.activ_fn, num_layers, "q2_encoder", reuse
            )
            if self.use_recurrent:
                q2_hidden, memory_out = self.create_recurrent_encoder(
                    q2_hidden, self.q2_memory_in, self.sequence_length, name="lstm_q2"
                )
                self.q2_memory_out = memory_out

            q2_heads = {}
            for name in stream_names:
                _q2 = tf.layers.dense(q2_hidden, num_outputs, name="{}_q2".format(name))
                q2_heads[name] = _q2

            q2 = tf.reduce_mean(list(q2_heads.values()), axis=0)

        return q1_heads, q2_heads, q1, q2

self, state_in: tf.Tensor, action_in: tf.Tensor, done_in: tf.Tensor, reuse: bool
    ) -> Tuple[tf.Tensor, tf.Tensor, tf.Tensor]:
        """
        Creates the encoder for the discriminator
        :param state_in: The encoded observation input
        :param action_in: The action input
        :param done_in: The done flags input
        :param reuse: If true, the weights will be shared with the previous encoder created
        """
        with tf.variable_scope("GAIL_model"):
            if self.use_actions:
                concat_input = tf.concat([state_in, action_in, done_in], axis=1)
            else:
                concat_input = state_in

            hidden_1 = tf.layers.dense(
                concat_input,
                self.h_size,
                activation=ModelUtils.swish,
                name="gail_d_hidden_1",
                reuse=reuse,
            )

            hidden_2 = tf.layers.dense(
                hidden_1,
                self.h_size,
                activation=ModelUtils.swish,
                name="gail_d_hidden_2",
                reuse=reuse,
            )

            z_mean = None