How to extract and reuse weights from tf.layers abstractions?

Question

Python

I want to use TensorFlow abstractions like tf.layers.dense() whenever possible instead of variable weights and offsets, their initialization, relu() and matmul().
I rewrite the model, where first there are 100500 lines of code with weights-byas, then the order of the layers is determined in the function:

part of the model code

def model( data, infere = False):
        keep_prob = 1 if infere else 0.5
        L2_regularizer = tf.contrib.layers.l2_regularizer(scale = 1e-4)
        TNI = tf.truncated_normal_initializer(stddev = 0.2, seed = seed)

        # Layers
        conv = tf.nn.conv2d(data, layer1_weights, [1,1,1,1], padding='SAME')
        hidden = tf.nn.relu(conv + layer1_biases)
        maxpool = tf.nn.max_pool(hidden, [1,2,2,1], [1,2,2,1], 'SAME')

        conv = tf.nn.conv2d(maxpool, layer2_weights, [1,1,1,1], padding='SAME')
        hidden = tf.nn.relu(conv + layer2_biases)
        maxpool = tf.nn.max_pool(hidden, [1,3,3,1], [1,2,2,1], 'SAME')

        shape = maxpool.get_shape().as_list()
        reshape = tf.reshape(maxpool, [shape[0], shape[1] * shape[2] * shape[3]])

        dense = tf.layers.dense( inputs = reshape, 
                                    units = 64, 
                                    activation = tf.nn.relu,
                                    kernel_initializer = TNI,
                                    bias_initializer = tf.ones_initializer(),
                                    kernel_regularizer = L2_regularizer,
                                    name = "Dense_1"
                                 )

        dropout = tf.nn.dropout( dense, keep_prob = keep_prob, seed = seed)

        dense = tf.layers.dense( inputs = dropout,
                                    units = 10, 
                                    activation = tf.nn.relu,
                                    kernel_initializer = TNI,
                                    bias_initializer = tf.ones_initializer(),
                                    kernel_regularizer = L2_regularizer,
                                    name = "Dense_2"
                                 )
        return dense

How to deal with weights for dense layers at the end so that their weights are common between the training, validation and test models? Somehow pull them out and declare them at the beginning, above the function? I didn’t understand what the reuse = True parameter does - maybe it will work with it?