Overview

The computation results of Ascend proprietary operators on Ascend AI Processors may be different from those of third-party equivalents (industrial benchmarks).

• Original network models can be migrated to Ascend training environments for training, but in this case, the computation results of Ascend proprietary operators may be different from those of third-party equivalents.
• Original network models can be migrated to Ascend environments for online inference, but in this case, the computation results of proprietary operators may be different from those of third-party equivalents.
• ATC optimizes models during conversion through operator elimination, fusion, and partition. As a result, the computation results of Ascend proprietary operators may be different from those of third-party equivalents.

  For details about ATC, see the ATC Instructions.

• There maybe a decrease in accuracy of an upgraded or tuned offline model generated through ATC-based conversion due to the CANN version iteration, model version iteration, model tuning, hardware upgrade, or operator fusion being enabled or disabled before model conversion.

To quickly resolve the operator accuracy issues, accuracy comparison tools must be designed to compare the computation results of Ascend proprietary operators and third-party equivalents.