Automatic modulation classification (AMC) is a key technique for desiging non-cooperative communication systems, and deep learning (DL) is applied effectively into AMC for improving the classification accuracy.However, most of the DL-based AMC methods have a large number of parameters and high computational complexity, and they cannot be directly applied into scenarios with limited computing power and storage space.In this paper, we propose a lightweight and low-complexity AMC method using ultra lite convolutional neural network (ULCNN), which is based on multiple tricks, including data augmentation, complex-valued convolution, separable convolution, channel attention, channel shuffle. Simulation results demonstrate that our proposed ULCNN-based AMC method achieves the average accuracy of 62.47\% on RML2016.10a and only 9,751 parameters. Moreover, ULCNN is verified on a typical edge device (Raspberry Pi), where interference time per sample is about 0.775 ms. The reproducible code can be download from GitHub at https://github.com/BeechburgPieStar/Ultra-Lite-Convolutional-Neural-Network-for-Automatic-Modulation-Classification .