Neural network structure search method, device, electronic device and medium

Abstract

The embodiments of the present application disclose a neural network structure search method, device, electronic device and medium, which relate to the technical field of neural networks. The specific implementation scheme is: the controller sends the candidate network structure information obtained by the search to the trainer according to the preset communication protocol; the trainer trains the neural network model according to the candidate network structure information and sample data, and according to the preset communication protocol, will be based on The index information generated by the neural network model is fed back to the controller; the controller searches again according to the index information. If the candidate network structure information obtained by the search converges, the candidate network structure with the convergent structure information is determined as the target neural network structure obtained by the final search. . Based on the preset communication mechanism, the information exchange between the controller and the trainer based on the preset communication protocol is realized, which overcomes the technical problem of strong coupling caused by the controller and the trainer must be based on the same development framework. Decoupling from the trainer.

Description

technical field [0001] The embodiments of the present application relate to the field of computer technologies, in particular to the field of neural network technologies, and in particular to a method, apparatus, electronic device, and medium for searching a neural network structure. Background technique [0002] Through Neural Architecture Search (NAS), developers such as model tuning can automatically search for the optimal neural network structure. [0003] At present, in the process of automatic neural network structure search, the controller and the trainer must be based on the same development framework, so that the optimal network structure can be obtained through the cycle of structure search, model construction and training. . Therefore, the coupling between the controller and the trainer in the prior art is too strong, and not only cannot be independently debugged as an independent process, but also requires secondary development and index alignment of the trainer...

AI Technical Summary

Problems solved by technology

Therefore, the coupling between the controller and the trainer in the prior art is too strong, not only cannot be independently debugged as an independent process, but also in the case of different development frameworks, the trainer needs to be redeveloped and index aligned. It increases the development difficulty and time consumption, and requires high technical ability of developers

Images