In video surveillance scene, pedestrian attributes recognition is a complex task. There are some challenges in pedestrian attribute recognition, such as background noise and data imbalance. The problem is still open. This paper combines several effective methods and proposes a simpler and more effective method. Firstly, we restrain the weight of labels with more samples and improve the weight of the labels with less samples to relieve the effect of data imbalance. Secondly, we apply attention mechanisms which are effective in extracting the discriminative features. Also, we use bag of tricks to achieve a strong baseline, such as cosine learning rate, warm-up, data augmentation, etc. Finally, we evaluate these effective tricks in Pedestrian Attribute Recognition task and our model achieves state-of-the-art results in several important metrics in datasets PETA, RAP and PA-100k.