Accuracy Optimization

This section provides step-by-step instructions for optimizing model inference accuracy with typical configurations and examples. Before optimizing inference accuracy of your model, first debug its inference functionality.

There are two possible causes for inference accuracy drop or even wrong inference results:

• Issues caused by the cascading of model inference and other functions, such as DVPP+AIPP+model inference. In this case, the parameter configurations of the APIs or the AIPP configuration during model conversion may be improper.

  If your service has a DVPP+AIPP+model inference cascade, this is the most likely cause.

  This document also provides configuration suggestions based on positive and negative examples for reference.

• In other cases, network accuracy drop is caused by improper configuration or implementation of particular operators. Use the Model Accuracy Analyzer to dump your model and analyze the dump data according to the provided procedure. This section details how to compare and analyze dump data with specific cases.

Figure 1 Inference inaccuracy issue

Unless otherwise specified, "inference" in this section refers to inference on an offline model (.om format).

This section focuses on network-level accuracy optimization. For details about accuracy optimization of custom operators, see Ascend C Operator Development Guide .