A node fusion method and device for a computational graph

Abstract

The embodiment of the application discloses a kind of node fusion methods and equipment of calculation graph, it can be applied to artificial intelligence field, specifically can be applied to deep learning framework, method includes: converting neural network into calculation graph, and one or more parallel branch groups are extracted from calculation graph based on the dependency relationship between nodes in calculation graph, dependency relationship is used to indicate at least one of the following relationships: parallel branch group has common parent node, parallel branch group has common child node, parallel branch group has no parent node, parallel branch group has no child node, finally, the nodes of any parallel branch group from different sub-branches are fused respectively, to obtain new calculation graph.The application considers the possibility of other parallel branches, so as to find the parallel branch combination different from the rule defined in the prior art, i.e., the nodes in these branch combinations can be fused in the node fusion of the calculation graph, thereby expanding the range of fusible nodes that can be obtained.

Images