Implementation of Person Re-Identification Algorithm Based on Huawei Mindspore and Pytorch Hybrid Architecture

Based on Huawei's Mindspore and Pytorch hybrid architecture, this paper carries out localization operation of the current mainstream pedestrian re-recognition algorithm - AGW model, and makes it successfully run correctly on the domestic hardware platform. In this paper, the construction mode of Mindspore map and the technical characteristics of Mindspore branch structure in the process of localization of pedestrian rerecognition model are deeply analyzed, and the difficulties of localization of AGW model are explored. The effects of the implementation of data processing logic in different frameworks on the running performance of AGW localization model are studied. Develop a cross-framework parameter mapping module to map the parameters in Pytorch to the Mindspore framework in order to maintain the performance of the localized AGW model. Finally, the visualization module proposed in this paper can track the feature map changes in the domestic AGW model. In this paper, the domestic AGW model was tested on SYSU-MM01 and RegDB data sets, and its performance in RANK1 index and mAP index reached more than 80% of the original model. The experimental results show that the localized AGW model based on Huawei's Mindspore and Pytorch hybrid architecture runs successfully and correctly on the localized platform.