Comparison Result Analysis Guide
- The asterisk (*) indicates that a new unique operator on the NPU does not have a corresponding standard operator. In this case, comparison is unavailable and NaN is displayed.
- If the cosine similarity and KL divergence results are NaN while other algorithms yield valid results, it indicates that either the data to be compared or the benchmark data of the operator is zero. If the KL divergence result is Inf, it indicates that one of the benchmark data values is zero.
- For common accuracy issues, output expert suggestions provided by Advisor to quickly locate faulty operators.
- For complex scenarios that cannot be covered by expert suggestions, you can view different algorithm indicators in the network-wide comparison results and locate operators with accuracy issues based on Computation Accuracy Evaluation Indicators.
- In scenarios with multiple comparisons or fuzzy issue demarcation, you can configure the -r or -s option to achieve accuracy comparison within any selected range of operators, especially for large-scale networks, enabling rapid location of accuracy issues. For details about the -r or -s option, see Command Syntax.
- For suspicious comparison results, check the accuracy issues according to Accuracy Comparison Between .npy Files.
- For operators with large accuracy differences, you can use the Single-Operator Comparison function to further analyze the detailed accuracy differences of corresponding tensors.
- You can optimize faulty operators by modifying, replacing, or merging them based on specific scenarios.
Parent topic: Comparison Result Analysis