Methods and systems for converting weights of a deep neural network from a first number format to a second number format

摘要

Converting a plurality of weights of a filter of a Deep Neural Network (DNN) from a first to a second number format to enable the DNN to be implemented in hardware logic, the second format having less precision than the first. The conversion comprising determining, for each of the plurality of weights, a quantisation error associated with quantising that weight to the second format in accordance with a first quantisation method 402. The total quantisation error is determined for the plurality of weights based on said quantisation errors 404. A subset of the weights is identified to be quantised to the second format in accordance with a second quantisation method based on the total quantisation error 406. A set of quantised weights is generated, each weight in the subset based on quantising that weight to the second format in accordance with the second quantisation method and each of the remaining weights based on quantising that weight to the second format in accordance with the first quantisation method 408.