Model similarity comparison method, system and device and medium

Abstract

The invention discloses a model similarity comparison method, which comprises the following steps of: obtaining parameters of a plurality of to-be-compared models so as to construct a directed graph corresponding to each model according to the parameters; calculating a transformation vector of each directed graph from a first layer to a last layer; expanding the dimensions of the rest transformation vectors according to the transformation vector of the maximum dimension, so as to enable the dimensions of the transformation vectors to be the same; and calculating a cosine distance between everytwo of the plurality of transformation vectors, and determining similarities of the to-be-compared models according to the cosine distances. The invention further discloses a system, computer equipment and a readable storage medium. According to the scheme provided by the invention, the similarities between the models are measured through a vector distance-based calculation method. According to the scheme, starting from the structure and semantics of the model, the similarity information of the given model is rapidly described without depending on a specific data set, and the invention has wide application in practice.

Description

technical field [0001] The invention relates to the field of similarity comparison, in particular to a model similarity comparison method, system, device and storage medium. Background technique [0002] In order to quickly obtain a suitable model, a model database is usually built for different types of networks that have been trained in history. In the process of building the model database, in order to ensure the diversity of the models, it is required that the new model is not highly correlated with the models in the database. Therefore, the research model similarity becomes the key. [0003] Such as figure 1 As shown, the current common model similarity calculation method is, given two deep learning models A and B, for the sample set S, perform the same task, and analyze the similarity of the models from the results. For example, for binary classification tasks, common similarity comparison methods are defined as follows: [0004] [0005] Among them, the numerat...

AI Technical Summary

Problems solved by technology

[0006] Existing model similarity calculation methods measure the similarity value between models through the test results of real data sets, and the execution process requires time and hardware resources. Easy to use

Moreover, the existing model similarity calculation methods depend on specific test data sets, and different data sets will have different test results, resulting in different model similarity values.

For a given model, this method makes the result of model similarity uncertain, and has no theoretical guiding significance for the definition of model similarity

Images