The invention discloses an efficient
federated learning large model training method based on dynamic layer selection, and relates to the technical field of efficient training of
federated learning. According to the method, the
server side selects part of the most important
layers and transmits the most important
layers to the
client side for training, namely, part of the calculated amount is distributed to the
server side, and the
client side only needs to update the parameter amount of a very small part of the
layers, so that the calculation pressure and the uploading pressure of the
client side are greatly reduced; the computing pressure and the communication pressure of the client are obviously reduced; the method for selecting part of the most important layers specifically comprises the following steps: a
server performs preliminary training on a to-be-trained model based on a small amount of local training data, obtains a gradient value of each
linear layer of the to-be-trained model, and calculates an importance
score of each
linear layer based on the gradient value; the server selects a
linear layer corresponding to a plurality of previous highest importance scores and a plurality of later lowest importance scores as a current important layer; and the client only trains part of the most important layer according to the local training result, so that the calculation demand of training and the parameter quantity of uploaded data are greatly reduced. And meanwhile, the model does not need to be additionally changed, so that the expandability is very high.