CustomOpBuilder-Based Custom Operators

Overview

In dynamic graph mode, network workflows are easier to debug, supporting operations like single-operator execution, normal functions/networks, and standalone gradient computations.

While Custom Primitive-Based Custom Operators support both static and dynamic graphs, they require extensive definitions. To simplify the integration of custom operators in dynamic graphs, MindSpore has introduced a new Python API, CustomOpBuilder, which not only improves usability but also enhances the execution performance of custom operators in dynamic graphs.

When developing operators using C++ interfaces, users need to define the operator function body, including deriving and constructing output tensors, calling and executing device operators, and more. Once the function body is defined, the pybind11 component can be used to register C++ functions as Python module interfaces.

Usage Scenarios

• Developing Forward and Backward Operators Using the Function Interface: Introduces the method of defining custom operator forward and backward propagation functions.