Each year, pneumonia affects about 450 million people globally and results in about 4 million deaths. In developing countries, pneumonia remains a leading cause of death in chronic patients and older adults. Chest X-rays are currently the best available method for diagnosing pneumonia, which is very critical for the treatment of patients. Deep learning-based image recognition technology can significantly improve the efficiency of pneumonia detection. Data Augmentation is a technology that artificially expands the training data set by using limited data to generate more comparable data. Excellent data augmentation methods can effectively improve the performance of neural networks and are currently widely used in various fields of deep learning. We propose a novel data augmentation method called Attention Mask in this paper, which provides accurate predictions and a more explainable attention focus comparing with many traditional data augmentation methods, such as random erasing and hide-and-seek. We guide the weight and focal point of the model with the attention mechanism to avoid the model from relying too much on superficial features. After data augmentation, the self-ensemble of different stage models also makes the entire system more stable. The experiments show that the attention is reasonably diverted, which is extremely helpful in classifying the target that is fallibility correctly. On Chest X-Ray Images (Pneumonia) dataset for classification, our method notably improves performance over baselines.