Training method and device for zero-shot learning model based on reinforcement learning

Abstract

The invention discloses a training method and device of a zero-sample learning model based on reinforcement learning, and a computer-readable storage medium. The method includes the following steps: obtaining a trained classification model; fixing the parameters of the trained classification model, and training Input the data into the trained classification model; obtain the feature data obtained from the training data after feature extraction from the trained classification model, and convert the feature data into a state vector; input the state vector into the action prediction model based on reinforcement learning, according to the current reward The value is predicted to obtain the action prediction result and execute the corresponding action to adjust the input training data; the parameters of the action prediction model based on reinforcement learning are optimized through the loss function to obtain the trained action prediction model based on reinforcement learning, and then combined with The trained classification model composes the trained reinforcement learning based zero-shot learning model. The present invention achieves the effect of further improving the performance of the zero-sample learning model.

Description

technical field [0001] The present application relates to the technical field of image recognition, and in particular to a training method and device for a zero-shot learning model based on reinforcement learning, and a computer-readable storage medium. Background technique [0002] Zero Shot Learning (ZSL) is an image classification technique in which the training set and the test set have no intersection in the data category. The traditional image classification technology belongs to supervised learning, which needs to label each picture to facilitate the training of the model. However, there are many kinds of creatures in the real world, and labeling pictures of all species will consume a lot of labor costs, and some species even require expert knowledge for labeling. This limits the generation of data sets, which in turn makes it difficult for traditional image classification techniques to be widely promoted. However, zero-shot learning does not need to label pictures ...

AI Technical Summary

Problems solved by technology

[0003] The embodiment of the present application provides a training method and device for a zero-shot learning model based on reinforcement learning, and a computer-readable storage medium, which solves the problem that the traditional technology has a poor effect when performing target search on an image, and realizes image-based The serialization operation improves the accuracy and effect of target search for images, and achieves the effect of further improving the performance of the zero-shot learning model

Images