Dual tensor parallel method for mixed word frequency embedding

The invention discloses a mixed word frequency embedded double tensor parallel method, which specifically comprises the following steps of: S1, scanning a data set used for training once through a task distributor, counting the occurrence frequency of a word id of each query, and then, a greedy algorithm (minimax: enabling the maximum difference of word frequencies between the embedded tables after cutting to be as small as possible) is utilized to uniformly cut rows of the embedded tables to parallel equipment according to the total number of the word frequencies, so that the number of the word frequencies on each piece of equipment is basically consistent, and the invention relates to the technical field of deep learning. According to the mixed word frequency embedded double tensor parallel method, through word frequency distribution information of an embedded table, uniform transverse cutting according to the page view is achieved during tensor parallel, and uniform spreading of the training amount is guar