Model-based prediction method and device

Abstract

The embodiment of the invention provides a model-based prediction method and device, and the method comprises the steps: enabling a model operation environment to receive an input tensor of a machinelearning model; enabling the model operation environment to send a table look-up request to the embedded operation environment, wherein the table look-up request comprises the input tensor so as to request to perform low-dimensional conversion on the input tensor; enabling the model operation environment to receive a table look-up result returned by the embedded operation environment, wherein thetable look-up result is obtained by performing embedded query and processing on the embedded operation environment according to the input tensor; and enabling the model operation environment to inputthe table look-up result into a machine learning model, and operate the machine learning model to complete model-based prediction.

Description

technical field [0001] The present disclosure relates to the technical field of machine learning, and in particular to a model-based prediction method and device. Background technique [0002] In artificial intelligence, feature engineering is a very important link. Feature engineering is the process of converting the input data of a machine learning model into a feature vector, which is an important initial step in machine learning. Embedding is a data conversion method in feature engineering. The input data of machine learning may be a high-dimensional tensor, and the high-dimensional tensor can be mapped to a low-dimensional continuous space through the embedding layer to achieve feature extraction. Purpose. Machine learning on large inputs (typically high-dimensional sparse tensors) is made easier by mapping high-dimensional spaces to low-dimensional spaces such as embeddings. [0003] However, the amount of parameters in the embedding layer is usually very large. In ...

AI Technical Summary

Problems solved by technology

This situation poses a great challenge to the deployment and prediction of real-time online models: a single machine cannot provide enough shared memory to store the entire model, so it is necessary to purchase a customized machine to provide services, which increases the cost of equipment purchase, operation and maintenance ; Moreover, the capacity of a single machine is occupied by a large embedding layer, which also makes the model run unable to fully utilize the computing power of the underlying hardware resources, resulting in waste of resources

Images