We address the problem of facial expression analysis. The proposed approach predicts both basic emotion labels and valence/arousal values as a continuous measure for the emotional state. We train our system on the AffectNet dataset, which shows a high variation of faces, facial expressions and other conditions like illumination and occlusions. Evaluation on the AffectNet dataset and cross-database evaluation on the Aff-Wild dataset shows that our approach predicts emotion categories and valence and arousal values with high accuracies.