3 results about "Forward recursion" patented technology

The application discloses a kind of sea surface flow field inversion and reconstruction method based on stationary satellite high frequency image, belong to marine remote sensing technical field.It includes the following steps: satellite image is preprocessed and locally enhanced;Initial inversion is carried out using dense optical flow algorithm for single time phase flow field;The flow field data cube of space-time grid uniformization is constructed;Time series reconstruction is carried out using two-way Kalman filtering, through forward recursion, backward recursion and two-way fusion, using time series information to correct abnormal vector and fill in missing data;Finally, based on the physical consistency principle of seawater approximate incompressibility, the reconstructed flow field is subjected to iterative divergence constraint, and the non-physical disturbance is weakened.The application effectively solves the problem that the traditional optical remote sensing flow measurement method is easy to produce abnormal vector and poor spatiotemporal consistency under complex sea conditions such as cloud cover and weak texture by introducing time series correlation and physical constraint, significantly improves the accuracy, stability and physical reasonableness of sea surface flow field inversion.

The invention discloses an assembly line optimization method and system for accelerating multi-mode large model lightweight training, and the method comprises the steps: obtaining a data set, determining the batch size, setting the assembly line stage number, the micro-batch number and the required assembly line stage number according to the GPU number, and enabling the micro-batch number to be an integral multiple of the GPU number; dividing the to-be-trained model into parts with the number of assembly line stages according to the structure of the to-be-trained model, wherein the difference between the calculated amounts of different parts does not exceed 10%; deploying the models of different stages to a GPU according to a predetermined rule; and performing a training process on each GPU according to a pipeline parallel rule. According to the method, the characteristic that forward recursion and back propagation are close in time consumption in lightweight training of the multi-modal large model is utilized, bubbles in pipeline parallelization are reduced while low GPU video memory occupation is used, and the training efficiency is improved.

The application provides a CVT error calculation method based on an improved Kalman smoothing method and a medium, aiming to improve the error estimation accuracy and stability of the CVT system. The method combines specific CVT accuracy level calibration parameters, optimizes the real-time data of multiple in-phase CVTs through Kalman filtering and Kalman smoothing technology, and improves the measurement accuracy. In the method, the algorithm effectively reduces the influence of noise on CVT error calculation through iterative optimization of multi-time data, and improves the measurement accuracy and stability of the CVT system. Through the double optimization process of forward recursion and backward correction, the application realizes accurate correction of the CVT error in the dynamic changing working environment. The Kalman smoothing method can combine historical data and real-time data, not only improving the accuracy of single estimation, but also enhancing the stability and reliability of error calculation in multi-time data.