Data processing method and device

Abstract

Data processing method and apparatus are disclosed. The method includes: processing input data based on the data processing network, to obtain intermediate data; acquiring a fixed attention feature output by the attention network, which is obtained after training the model with at least an initialized attention feature, respective attention weights included in the initialized attention feature are not all the same; processing the intermediate data based on the data aggregation network and according to the fixed attention feature, to obtain output data. The fixed attention network does not receive any input previously, that is, attention feature output by the attention network is irrelevant to input data. The fixed attention feature can not only allow a more important parameter of the model to play a greater role to improve accuracy of the model and determine importance of the parameter, but also be conducive to further compression of the model.