Avoiding cycles in neural networks

A computer device comprising:

• a hardware processor;
• and a non-transitory storage medium storing instructions executable on the hardware processor to:
• obtain a source code representation of a neural network program, the source code representation being according to a domain specific language (DSL);
• parse and process the source code representation to create a graph structure for use in execution of the neural network program;
• create an instance of a proxy class to store a pointer to a first proxy instance of a first tensor;
• detect, in the source code representation, a compound operation including the first tensor, the compound operation in the source code representation modeled in the graph structure as a cycle in the graph structure;
• in response to detecting the compound operation:
• provide a second proxy instance of the first tensor, and
• modify the instance of the proxy class to store a pointer to the second proxy instance of the first tensor;
• and
• use the instance of the proxy class to facilitate implementation of a level of indirection to replace the cycle in the graph structure with an acyclic portion to create an acyclic graph structure, wherein the acyclic graph structure indicates assignment of a result of the compound operation to the second proxy instance of the first tensor.

A non-transitory computer readable medium comprising instructions that when executed by one or more hardware processors cause the one or more hardware processors to:

• obtain a source code representation of a neural network program, the source code representation being according to a domain specific language (DSL);
• parse and process the source code representation to create a graph structure for use in execution of the neural network program;
• create an instance of a proxy class to store a pointer to a first proxy instance of the first tensor;
• detect, in the source code representation, a compound operation including a first tensor, the compound operation in the source code representation modeled in the graph structure as a cycle in the graph structure;
• in response to detecting the compound operation:
• provide a second proxy instance of the first tensor, and
• modify the instance of the proxy class to store a pointer to the second proxy instance of the first tensor;
• and
• use the instance of the proxy class to facilitate implementation of a level of indirection to replace the cycle in the graph structure with an acyclic portion to create an acyclic graph structure, wherein the acyclic graph structure indicates assignment of a result of the compound operation to the second proxy instance of the first tensor.