S2_Fc_landmark = tf.layers.dense(S2_Fc1, n_landmark * 2)
S2_Ret = LandmarkTransformLayer(S2_Fc_landmark + S2_InputLandmark, S2_AffineParam, Inverse=True)

S2_Fc_emotion = tf.layers.dense(S2_Fc1, nb_emotions)

S2_Cost_landm = tf.reduce_mean(NormRmse(GroundTruth, S2_Ret)) # landmarks loss
S2_Cost_emotion = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(labels=one_hot_labels, logits = S2_Fc_emotion)) # emotion loss

Joint_Cost = alpha*S2_Cost_landm + beta*S2_Cost_emotion

S2_Optimizer = tf.train.GradientDescentOptimizer(learning_rate=learning_rate).minimize(Joint_Cost,\
                                                                                       global_step=global_step)