Overview

Model Accuracy Analyzer only compares the computation results on Ascend AI Processors of Huawei proprietary operators with those of third-party equivalents. You need to analyze the comparison results and locate faults. However, result analysis is difficult. This section describes how to use the Advisor system to analyze accuracy comparison results, which effectively reduces the troubleshooting time. You only need to select Advisor during task configuration in Comparison Procedure, and the system automatically analyzes the result file after the comparison is complete and provides tuning suggestions.

Currently, the following detection analysis types are supported: FP16 overflow/underflow detection, input inconsistency check, and network-wide consistency check (faulty node detection, single-point error detection, and consistency check).