CPC G06N 3/105 (2013.01) [G06F 8/34 (2013.01); G06F 18/214 (2023.01); G06F 18/2137 (2023.01); G06F 18/2178 (2023.01); G06N 3/044 (2023.01); G06N 3/047 (2023.01); G06N 3/08 (2013.01)] | 20 Claims |
1. A method comprising:
maintaining, by a computational graph system that manages execution of computational graphs representing neural network operations for users of the computational graph system, data specifying a plurality of pre-trained neural networks, wherein each of the pre-trained neural networks is a neural network that has been trained on training data to determine trained values of respective parameters of the neural network;
obtaining data specifying a user computational graph representing neural network operations, the user computational graph comprising a plurality of nodes connected by edges;
modifying the user computational graph by modifying a subgraph of the user computational graph using a particular pre-trained neural network from the plurality of pre-trained neural networks, comprising:
identifying the subgraph of the user computational graph, wherein the subgraph of the user computational graph receives a subgraph input, and generates a subgraph output after processing the subgraph input, wherein the subgraph output is provided as an input to a succeeding node in the user computational graph;
selecting the particular pre-trained neural network from the plurality of pre-trained neural networks, wherein the particular pre-trained neural network receives an input conforming to the subgraph input, and generates an output conforming to the subgraph output; and
replacing the subgraph with the particular pre-trained neural network by inserting a remote call node into the user computational graph to modify the user computational graph, wherein the remote call node is configured to, during execution of the modified user computational graph:
receive data representing the subgraph input,
provide, using a remote call, the subgraph input as an input to the particular pre-trained neural network,
obtain, in response to the remote call, a remote output generated by the particular pre-trained neural network for the subgraph input, and
provide, as the subgraph output, the remote output generated by the particular pre-trained neural network to the succeeding node; and
storing the modified user computational graph in a memory device.
|