Codebook quantization model training method, search data quantization method and device

The invention provides a codebook quantization model training method and a search data quantization method and device. The training method comprises the following steps: mapping each search sample into B orthogonal spaces by utilizing a processor based on B orthogonal sub-matrixes to obtain B orthogonal vectors; the processor is utilized to generate a distribution probability matrix according to each orthogonal vector and B center point codebooks of an orthogonal space corresponding to each orthogonal vector, and a center point codebook set comprises B center point codebooks; determining an index corresponding to a maximum value according to the maximum value in the distribution probability matrix, and generating a one-hot coding vector of an orthogonal vector in an orthogonal space; generating an output training vector according to the search sample and the plurality of one-hot coding vectors; inputting the search sample, the central point codebook set, the output training vector and the B orthogonal sub-mat