Image recognition method for efficient GPU training model based on wide model sparse data set

Abstract

The invention discloses an image recognition method for an efficient GPU training model based on a wide model sparse data set, and the method comprises the steps: converting pixels of an image into vectors, and carrying out the training and classification of the vectors; establishing a machine learning training method CuWide; enabling the cuWide to adopt a flow pipeline with a replication strategy, an importance cache, column-oriented storage and multi-flow technology. GPU is utilized and a large number of sparse data sets are used for efficiently training the image recognition prediction widemodel, the image recognition prediction wide model is deployed and trained on the GPU, the method is excellent in performance in the aspect of training large-scale wide models, and the image recognition efficiency can be greatly improved.

Description

technical field [0001] The invention belongs to the technical field of machine learning and image processing, and relates to an image classification and recognition method, in particular to an image recognition method based on a high-efficiency GPU training model based on a wide model sparse data set. Background technique [0002] Machine learning and data mining problems such as recommender systems, ad click-through rate prediction, and image recognition have become increasingly popular due to their success in many practical services, such as image classification tasks that combine computer vision and machine learning algorithms to extract meaning from images , which can assign a label to an image or a high-level human-readable sentence explaining the content of the image, is a hot topic both in academia and industry. In the field of image recognition, the accuracy of classification directly affects commercial profits. In the case of a huge amount of data, not only the accu...

AI Technical Summary

Problems solved by technology

Therefore, phase-based strategies are difficult to start from the advantages of shared memory

[0005] 2. The gap between sparse irregular data and GPU SIMT architecture

However, since GPUs are designed with SIMT architecture to handle data-intensive tasks, without careful tuning, sparse data can easily lead to many redundant memory accesses

The feature distribution approximates a power-law distribution, which is highly skewed. This feature distribution causes non-uniform access to model parameters, leading to thread conflicts for high-frequency features when using GPUs.

Images