Introduction

Compared with industry benchmark operators, the computation results of Ascend-proprietary operators on the AI processor may exhibit discrepancies:

• Model migration: When a source model (GPU-based) is migrated to the Ascend environment for training or online inference, discrepancies may arise between Ascend-proprietary operators and benchmark operators.
• Model conversion: During conversion, the ATC tool performs optimizations such as operator elimination, fusion, and splitting. These operations may cause divergence between Ascend-proprietary and benchmark operator computation results.
• Model compatibility: For offline models converted through ATC, factors such as CANN or model version iterations, model optimization, hardware upgrades, or changes in operator fusion settings may lead to accuracy degradation in the optimized model.

To assist developers in resolving operator accuracy issues, the accuracy debugging tool provides a comparison function between Ascend-proprietary and benchmark operator computation results.

For details about ATC, see ATC.

Overall Accuracy Comparison Workflow

The overall accuracy comparison workflow is as follows:

The following flowchart describes the three supported scenarios:

• GPU vs. Ascend NPU (training): Comparison of accuracy data between the source model (GPU-based) and the model migrated to the Ascend NPU environment. For details, see GPU vs. NPU (TensorFlow 1.15 Training/Online Inference).
• GPU/CPU vs. Ascend NPU (inference): Comparison of accuracy data between the source model (GPU/CPU-based) and the NPU model after conversion. For details, see Comparison Between GPU and NPU (TensorFlow Offline Inference), Comparison Between GPU and NPU (ONNX Offline Inference), and Comparison Between GPU/CPU and NPU (Caffe Offline Inference).
• Ascend NPU vs. Ascend NPU (inference): Comparison of accuracy data between two versions of ATC-converted offline models. This comparison is required in cases of CANN version iterations, model version iterations, model optimization, hardware upgrades, or inconsistent operator fusion settings before ATC conversion. For details, see Comparison Between NPU and NPU (Offline Inference).

Figure 1 Overall accuracy comparison workflow