Overview

With open AscendCL APIs, users can construct graphs based on operator prototypes. This eliminates the need for complex and time-consuming framework adaptation. Alternatively, Ascend provides framework parsers to import existing models from popular frameworks (Caffe, TensorFlow, or ONNX) into graphs supported by Compute Architecture for Neural Networks (CANN), introducing an easy way to abstract away framework details.

After a graph is constructed using operator prototypes or parsed from an existing model, you can modify, build, and execute the graph to implement model expression and high-level build and optimization.