The following TensorFlow code will build a convolutional autoencoder model for the MNIST dataset. This first part of the code will construct the graph of your model, the encoder and the decoder. In the code, we highlight the part of the model whose output will be our latent vector:

class CAE_CNN(object): 
   def __init__(self, img_size = 28, latent_size=20): 
       self.__x = tf.placeholder(tf.float32, shape=[None, img_size * img_size], name='IMAGE_IN') 
       self.__x_image = tf.reshape(self.__x, [-1, img_size, img_size, 1]) 
 
       with tf.name_scope('ENCODER'): 
           ##### ENCODER 
           # CONV1: Input 28x28x1 after CONV 5x5 P:2 S:2 H_out: 1 + (28+4-5)/2 = 14, W_out= 1 + (28+4-5)/2 = 14 
           self.__conv1_act = tf.layers.conv2d(inputs=self.__x_image, strides=(2, 2), 
                                               filters=16, kernel_size=[5, 5], padding="same", activation=tf.nn.relu) 
 
           # CONV2: Input 14x14x16 after CONV 5x5 P:0 S:2 H_out: 1 + (14+4-5)/2 = 7, W_out= 1 + (14+4-5)/2 = 7 
           self.__conv2_act = tf.layers.conv2d(inputs=self.__conv1_act, strides=(2, 2), 
                                               filters=32, kernel_size=[5, 5], padding="same", activation=tf.nn.relu) 

           with tf.name_scope('LATENT'): 
           # Reshape: Input 7x7x32 after [7x7x32] 
           self.__enc_out = tf.reshape(self.__conv2_act, [tf.shape(self.__x)[0], 7 * 7 * 32]) 
           self.__guessed_z = tf.layers.dense(inputs=self.__enc_out, 
                                              units=latent_size, activation=None, name="latent_var") 
           tf.summary.histogram("latent", self.__guessed_z) 
 
 
       with tf.name_scope('DECODER'): 
           ##### DECODER (At this point we have 1x18x64 
           self.__z_develop = tf.layers.dense(inputs=self.__guessed_z, 
                                              units=7 * 7 * 32, activation=None, name="z_matrix") 
           self.__z_develop_act = tf.nn.relu(tf.reshape(self.__z_develop, [tf.shape(self.__x)[0], 7, 7, 32])) 
 
           # DECONV1 
           self.__conv_t2_out_act = tf.layers.conv2d_transpose(inputs=self.__z_develop_act, 
                                                               strides=(2, 2), kernel_size=[5, 5], filters=16, 
                                                               padding="same", activation=tf.nn.relu) 
 
           # DECONV2 
           # Model output 
           self.__y = tf.layers.conv2d_transpose(inputs=self.__conv_t2_out_act, 
                                                               strides=(2, 2), kernel_size=[5, 5], filters=1, 
                                                               padding="same", activation=tf.nn.sigmoid) 
 
           # We want the output flat for using on the loss 
           self.__y_flat = tf.reshape(self.__y, [tf.shape(self.__x)[0], 28 * 28]) 

The code snippet related to the convolutional autoencoder loss is as follows:

with tf.name_scope("CAE_LOSS"): 
   # L2 loss
   loss = tf.losses.mean_squared_error(labels=model_in, predictions=model_out_flat)
 
# Solver configuration 
with tf.name_scope("Solver"): 
   train_step = tf.train.AdamOptimizer(0.0001).minimize(loss)