def z_est(self):
print "Layer ", self.layer_number, ": ", self.output_nodes, " -> ", self.input_nodes, ", denoising cost: ", self._denoising_cost
u = unlabeled(self.input_layer.activation_train)
u = LadderLayer.batch_normalization(
u, self.input_layer.mean_clean_unlabeled,
self.input_layer.variance_clean_unlabeled)
return self._g_gauss(unlabeled(self.input_layer.z_corrupted), u)
def unsupervised_cost_train(self):
cost = tf.reduce_mean(
tf.reduce_sum(
tf.square(self.z_est_bn - unlabeled(self.input_layer.z_clean)),
1))
# TODO: input_nodes may change...
return (cost / self.input_nodes) * self._denoising_cost
def unsupervised_cost_train(self):
cost = tf.reduce_mean(tf.reduce_sum(tf.square(self.z_est_bn - unlabeled(self.input_layer.z_clean)), 1))
# TODO: input_nodes may change...
return (cost / self.input_nodes) * self._denoising_cost
def z_est(self):
print "Layer ", self.layer_number, ": ", self.output_nodes, " -> ", self.input_nodes, ", denoising cost: ", self._denoising_cost
u = unlabeled(self.input_layer.activation_train)
u = LadderLayer.batch_normalization(u, self.input_layer.mean_clean_unlabeled, self.input_layer.variance_clean_unlabeled)
return self._g_gauss(unlabeled(self.input_layer.z_corrupted), u)
