variational_autoencoder.py 文件源码

def _create_network(self):
        # Initialize autoencode network weights and biases
        network_weights = self._initialize_weights(**self.network_architecture)

        # Use recognition network to determine mean and 
        # (log) variance of Gaussian distribution in latent
        # space
        self.z_mean, self.z_log_sigma_sq = \
            self._recognition_network(network_weights["weights_recog"], 
                                      network_weights["biases_recog"])

        # Draw one sample z from Gaussian distribution
        n_z = self.network_architecture["n_z"]
        eps = tf.random_normal((self.batch_size, n_z), 0, 1, 
                               dtype=tf.float32)
        # z = mu + sigma*epsilon
        self.z = tf.add(self.z_mean, 
                        tf.mul(tf.sqrt(tf.exp(self.z_log_sigma_sq)), eps))

        # Use generator to determine mean of
        # Bernoulli distribution of reconstructed input
        self.x_reconstr_mean = \
            self._generator_network(network_weights["weights_gener"],
                                    network_weights["biases_gener"])