1results about How to "Extensive knowledge" patented technology

This application relates to a method, apparatus, computer device, storage medium, and computer program product for training a knowledge distillation model. The method includes: adding a first weight matrix and a second weight matrix to the linear layer of the student model to be trained, based on the original weight matrix of the linear layer; obtaining a first predicted label for sample data using a teacher model, and obtaining a second predicted label for the sample data using the student model to be trained; the sample data carrying sample labels; determining the model loss value of the student model to be trained using the first predicted label, the second predicted label, and the sample labels; freezing the original weight matrix, and adjusting the first and second weight matrices using the model loss value to train the model, thereby obtaining the trained student model. This method can improve the training efficiency of the student model and reduce processing resource consumption.