This paper proposes a parallel adaptive network named PA-Net for polyp segmentation, which enables to dynamically adjust convolution kernels and attention weights according to the input. It enhances the focus on key regions, thereby ameliorating the challenge of hard identification due to the complex morphology and the blurring of the border between polyp and mucosa. Furthermore, as a lightweight network, PA-Net does not lead to excessive computational overhead and real-time efficiency of 34 FPS achieved. The experimental results show that PA-Net achieves excellent segmentation performance on public colonoscopy dataset.