文档

Accuracy Comparison Workflow

The overall accuracy comparison workflow is as follows:
Figure 1 Overall accuracy comparison workflow
  1. Set up the environment. For details, see Environment Setup.
  2. Prepare data for comparison. For details, see Data Preparation.
    1. Inference scenario: Prepare the .npy data files of an original model trained under a third-party framework and the dump data files of the corresponding offline model.
    2. Training scenario: Prepare the .npy data files of an original model trained under a third-party framework running on the GPUs and the dump data and computational graph files of a training network running on the Ascend AI Processors.
  3. Perform tensor comparison. For details, see Tensor Comparison.
    1. Perform network-wide comparison.

      Enable the Ascend > Model Accuracy Analyzer function of MindStudio, configure the prepared comparison data files to the corresponding parameters, and set the comparison parameters.

    2. MindStudio performs the comparison and outputs the comparison result.
    3. (Optional) MindStudio Advisor analyzes the comparison result and provides expert suggestions. For details, see Expert Suggestions on Comparison Results.
    4. Locate the faulty operator based on the analysis result.
    5. Perform single-operator comparison.
    6. Analyze the single-operator fault.
Parent topic: Accuracy Comparison