Hardware-friendly weight compression coding method oriented to visual Transform

The invention discloses a hardware-friendly weight compression coding method oriented to a visual Transform, and belongs to the technical field of calculation, reckoning or counting. The storage space occupied by the weight parameters is reduced by performing adaptive precision coding on the weight data. The method comprises the steps of type-by-type coding of weight data, self-adaptive precision selection and a coding evaluation mechanism. Firstly, weight data of different types of network layers are initialized and coded, then the optimal weight coding precision is found by utilizing a Bayesian information criterion, and finally iterative coding is carried out and an optimal coding result is reserved according to a score obtained by a coding evaluation mechanism. According to the method, the weight storage overhead can be effectively compressed while the precision of the visual Transform is kept, and the decoding unit is friendly in hardware and suitable for edge device deployment. 本发明公开一种面向视觉Transformer的硬件