Light sheet fluorescence microscopy (LSFM) has been proposed as a promising tool for biological research due to its ability to observe the dynamics of living cells for hours and days. Nonetheless, this raises the strict requirement for the overlapping between the light sheet and the detection focal plane to get optimal image quality. Here, we describe a fast and accurate deep learning-based autofocusing method to overcome the challenge of unstable focusing in LSFM with a single shot. This method is compatible with any light sheet imaging setup with a spatial light modulator for light sheet generation. A predicted root-mean-square error of 0.0942 µm across a range of ±0.7 µm in a light sheet microscope using a 1.1 numerical aperture detection objective can be obtained. The neural network architecture we proposed shows advantages in small memory size, few training data set requirements, and good generalization to untrained sample types.