How to use the mlagents.tf_utils.tf.identity function in mlagents

prev_action_oh = tf.concat(
                [
                    tf.one_hot(self.prev_action[:, i], self.act_size[i])
                    for i in range(len(self.act_size))
                ],
                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(

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.output = tf.identity(output)
        self.normalized_logits = tf.identity(normalized_logits, name="action")

        self.create_value_heads(self.stream_names, hidden)

        self.action_holder = tf.placeholder(
            shape=[None, len(policy_branches)], dtype=tf.int32, name="action_holder"
        )
        self.action_oh = tf.concat(
            [
                tf.one_hot(self.action_holder[:, i], self.act_size[i])
                for i in range(len(self.act_size))
            ],
            axis=1,
        )
        self.selected_actions = tf.stop_gradient(self.action_oh)

        self.all_old_log_probs = tf.placeholder(

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.output = tf.identity(output)
        self.normalized_logits = tf.identity(normalized_logits, name="action")

        self.create_value_heads(self.stream_names, hidden)

        self.action_holder = tf.placeholder(
            shape=[None, len(policy_branches)], dtype=tf.int32, name="action_holder"
        )
        self.action_oh = tf.concat(
            [
                tf.one_hot(self.action_holder[:, i], self.act_size[i])
                for i in range(len(self.act_size))
            ],
            axis=1,
        )
        self.selected_actions = tf.stop_gradient(self.action_oh)

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.output = tf.identity(output)
        self.normalized_logits = tf.identity(normalized_logits, name="action")

        self.create_value_heads(self.stream_names, hidden)

        self.action_holder = tf.placeholder(
            shape=[None, len(policy_branches)], dtype=tf.int32, name="action_holder"
        )
        self.action_oh = tf.concat(
            [
                tf.one_hot(self.action_holder[:, i], self.act_size[i])
                for i in range(len(self.act_size))
            ],
            axis=1,
        )
        self.selected_actions = tf.stop_gradient(self.action_oh)

self.memory_out = tf.identity(memory_policy_out, "recurrent_out")
        else:
            hidden_policy = encoded

        self.action_masks = tf.placeholder(
            shape=[None, sum(self.act_size)], dtype=tf.float32, name="action_masks"
        )

        with tf.variable_scope("policy"):
            distribution = MultiCategoricalDistribution(
                hidden_policy, self.act_size, self.action_masks
            )
        # It's important that we are able to feed_dict a value into this tensor to get the
        # right one-hot encoding, so we can't do identity on it.
        self.output = distribution.sample
        self.all_log_probs = tf.identity(distribution.log_probs, name="action")
        self.selected_actions = tf.stop_gradient(
            distribution.sample_onehot
        )  # In discrete, these are onehot
        self.entropy = distribution.entropy
        self.total_log_probs = distribution.total_log_probs

)
            self.memory_out = memory_value_out
        else:
            hidden_value = hidden_stream

        self.value_heads, self.value = ModelUtils.create_value_heads(
            self.stream_names, hidden_value
        )
        self.all_old_log_probs = tf.placeholder(
            shape=[None, sum(self.policy.act_size)],
            dtype=tf.float32,
            name="old_probabilities",
        )

        self.old_log_probs = tf.reduce_sum(
            (tf.identity(self.all_old_log_probs)), axis=1, keepdims=True
        )

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

            self.memory_in = tf.placeholder(
                shape=[None, self.m_size], dtype=tf.float32, name="recurrent_in"
            )
            hidden_policy, memory_policy_out = ModelUtils.create_recurrent_encoder(
                hidden_policy,
                self.memory_in,
                self.sequence_length_ph,
                name="lstm_policy",
            )

            self.memory_out = tf.identity(memory_policy_out, "recurrent_out")
        else:
            hidden_policy = encoded

        self.action_masks = tf.placeholder(
            shape=[None, sum(self.act_size)], dtype=tf.float32, name="action_masks"
        )

        with tf.variable_scope("policy"):
            distribution = MultiCategoricalDistribution(
                hidden_policy, self.act_size, self.action_masks
            )
        # It's important that we are able to feed_dict a value into this tensor to get the
        # right one-hot encoding, so we can't do identity on it.
        self.output = distribution.sample
        self.all_log_probs = tf.identity(distribution.log_probs, name="action")
        self.selected_actions = tf.stop_gradient(

def make_inputs(self) -> None:
        """
        Creates the input layers for the discriminator
        """
        self.done_expert_holder = tf.placeholder(shape=[None], dtype=tf.float32)
        self.done_policy_holder = tf.placeholder(shape=[None], dtype=tf.float32)
        self.done_expert = tf.expand_dims(self.done_expert_holder, -1)
        self.done_policy = tf.expand_dims(self.done_policy_holder, -1)

        if self.policy.brain.vector_action_space_type == "continuous":
            action_length = self.policy.act_size[0]
            self.action_in_expert = tf.placeholder(
                shape=[None, action_length], dtype=tf.float32
            )
            self.expert_action = tf.identity(self.action_in_expert)
        else:
            action_length = len(self.policy.act_size)
            self.action_in_expert = tf.placeholder(
                shape=[None, action_length], dtype=tf.int32
            )
            self.expert_action = tf.concat(
                [
                    tf.one_hot(self.action_in_expert[:, i], act_size)
                    for i, act_size in enumerate(self.policy.act_size)
                ],
                axis=1,
            )

        encoded_policy_list = []
        encoded_expert_list = []

tanh_squash=tanh_squash,
                condition_sigma=condition_sigma_on_obs,
            )

        if tanh_squash:
            self.output_pre = distribution.sample
            self.output = tf.identity(self.output_pre, name="action")
        else:
            self.output_pre = distribution.sample
            # Clip and scale output to ensure actions are always within [-1, 1] range.
            output_post = tf.clip_by_value(self.output_pre, -3, 3) / 3
            self.output = tf.identity(output_post, name="action")

        self.selected_actions = tf.stop_gradient(self.output)

        self.all_log_probs = tf.identity(distribution.log_probs, name="action_probs")
        self.entropy = distribution.entropy

        # We keep these tensors the same name, but use new nodes to keep code parallelism with discrete control.
        self.total_log_probs = distribution.total_log_probs

def make_inputs(self) -> None:
        """
        Creates the input layers for the discriminator
        """
        self.done_expert = tf.placeholder(shape=[None, 1], dtype=tf.float32)
        self.done_policy = tf.placeholder(shape=[None, 1], dtype=tf.float32)

        if self.policy.brain.vector_action_space_type == "continuous":
            action_length = self.policy.act_size[0]
            self.action_in_expert = tf.placeholder(
                shape=[None, action_length], dtype=tf.float32
            )
            self.expert_action = tf.identity(self.action_in_expert)
        else:
            action_length = len(self.policy.act_size)
            self.action_in_expert = tf.placeholder(
                shape=[None, action_length], dtype=tf.int32
            )
            self.expert_action = tf.concat(
                [
                    tf.one_hot(self.action_in_expert[:, i], act_size)
                    for i, act_size in enumerate(self.policy.act_size)
                ],
                axis=1,
            )