Automatic Facial Expression Recognition has always been an interesting and challenging problem. In this project, we experimented with two different convolutional neural networks (CNNs) models (shallow model and deep model) on different facial expression datasets, compared the performance of the same model on image datasets of different resolutions, and tested the performance of one CNN model on the FER2013 dataset. The result showed that the deep model achieves a better validation and test accuracy than the shallow model, and training with higher resolution images will lead to a better performance.