Distributed graph neural network training method and system based on model migration
The invention provides a distributed graph neural network training method and system based on model migration, and the method comprises the steps: sequentially migrating each graph neural network model to a corresponding server according to the position where the graph neural network model is locate...
Gespeichert in:
Hauptverfasser: | , , , |
---|---|
Format: | Patent |
Sprache: | chi ; eng |
Schlagworte: | |
Online-Zugang: | Volltext bestellen |
Tags: |
Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
|
Zusammenfassung: | The invention provides a distributed graph neural network training method and system based on model migration, and the method comprises the steps: sequentially migrating each graph neural network model to a corresponding server according to the position where the graph neural network model is located after the graph neural network model is redistributed according to a corresponding batch of training vertexes in each iteration of a training process; performing training and gradient accumulation by using the vertex features of the micrograph stored in the corresponding server, and remotely acquiring the vertex features of the micrograph not stored in the corresponding server; when all the graph neural network models finish training of the last micro-graph, accumulated gradients among all the graph neural network models are synchronized, and finally model parameters are updated; according to the method disclosed by the invention, when the feature vector data is in the remote state, the feature data is not transm |
---|