Model parameter conversion method and device

Abstract

The invention discloses a model parameter conversion method and device, and the method can obtain model parameters under a parameter type required by a to-be-constructed model, transmits the model parameters into the to-be-constructed model, carries out the format conversion through a pre-deployed parameter conversion relation in the to-be-constructed model based on the model parameters, and obtains the model parameters of the to-be-constructed model. The method comprises the following steps: acquiring a first parameter in a standard format according to a to-be-constructed model to obtain the first parameter in the standard format, further converting the first parameter in the standard format through a preset conversion function in the to-be-constructed model to obtain a first parameter in a parameter type required by a user-defined calculation layer, and transmitting the first parameter into the user-defined calculation layer of the to-be-constructed model; according to the method and the device, the to-be-constructed model containing the user-defined computing layer is constructed, the target service corresponding to the to-be-constructed model is executed through the constructed to-be-constructed model, and the user-defined computing layer is a network layer additionally configured for the target service, so that the model deployment efficiency is improved.

Description

technical field [0001] The present specification relates to the field of computer technology, and in particular, to a method and device for model parameter conversion. Background technique [0002] At present, with the continuous development of computer technology, machine learning models can gradually be widely used in all walks of life, and solve technical problems in many industries. [0003] In practical applications, developers can use the Torch framework when developing and training machine learning models, that is, Torch models are used in the development and training stages. Computationally efficient, Torch models can be converted to TensorRT models for deployment. [0004] Inside the machine learning model, developers can also write custom computing layers that need to be added to the machine learning model under the CUDA kernel framework to achieve special business needs. Because they need to be in the Torch framework and the model under the TensorRT framework res...

AI Technical Summary

Problems solved by technology

[0004] Inside the machine learning model, developers can also write a custom computing layer that needs to be added to the machine learning model under the CUDA kernel framework to achieve special business requirements, because it needs to be in the model under the Torch framework and the TensorRT framework Pass model parameters (such as convolution kernel parameters, network weights, etc.) In the custom computing layer inside the framework, every time developers build a model, they often need to pass parameters into the custom computing layer in this way, which reduces the efficiency of deploying the machine learning model

Images