Device for determining network configurations of a neural network under a number of ancillary conditions

摘要

Device for the determination of a Pareto set of network configurations, identifying the network configurations for a neural network for a given application and determining the specified application in the form of training data provided, including a machine-readable storage medium, stored on the command which, when running through a computer, caused the computer to perform a procedure with the following steps:(a) provide (S1) a majority of randomly selected network configurations (a);(b) Determine hyperparameters (A) for each of the network configurations provided (a) by means of an optimizer;(c) training of neural networks configured with the provided network configurations (a) and depending on the corresponding determined hyperparameters (A) on the provided training data;(d) establish a performance (accuracy) for each of the trained neural networks on at least part of the training data provided;(e) Evaluation of the trained neural networks against specified optimisation targets (n-pars);(f) Adding the network configurations (a), the determined hyperparameters (A) and the measured performance (accuracy) and the evaluated optimization targets (n) pars) to a further training data set;(g) adapting the optimizer for the further training data set, in particular depending on network configurations (a) and the accuracy obtained to predict suitable hyperparameters (A);(h) Repeat the following steps:and. Select (S3) network configurations (a) from the next training data set;Hey. random mutation (S4) of the selected network configurations; andiii. Execute steps (b) to (g) with the modified network configurations;(i) Repeated execution of the following steps:and. Select (S5) network configurations from the next training data set;Hey. Compress (angel. Pruning (S6) of the selected network configurations;iii. Execute steps (b) to (g) with the compressed network configurations;j) Select (S7) of the network configurations (a) from the further training data set, which corresponds to a Pareto set in terms of performance and at least one of the other optimization objectives.