Documentation

Single-Operator Comparison

Single-Operator Comparison Page

The single-operator comparison function is used to select an operator based on Comparison Results and specify parameters as required for comparison. The tensor comparison result page is displayed as areas 6, 7, and 8, as shown in Figure 1. For details, see Table 1.

Figure 1 Single-operator comparison
Table 1 Single-operator comparison result

Area

Description

6

Operator Detail: single-operator comparison, which delivers the single-operator comparison command. For details, see Comparison Procedure.

7

Top N single-operator comparison results. For details about the fields in each column, see Table 2.

8

Cumulative Error: cumulative error distribution line chart for top N tensor elements. For details, see Table 2.

Comparison Procedure

The single-operator comparison operations are displayed in area 6 of Figure 1, as shown in Figure 2. The procedure is as follows:

Figure 2 Single-operator comparison
  1. Based on the read tensor information, select the operator to be compared and the corresponding tensor (output/input) from the Operator drop-down list.
  2. Set Metric to Absolute Error or Relative Error.
  3. Set top N elements to output for the metric. The value range is [1, 10000].
  4. Click Compare.

    For details about Top N result display, see Comparison Results.

Comparison Results

The single-operator comparison results are displayed in areas 7 and 8 of Figure 1, as shown in Figure 3. For details, see Table 2.

Figure 3 Single-operator comparison result
Table 2 Comparison result field description

Field

Description

Area 7

Index

Number of items to be compared.

N C H W

Data format.

NPUDump

Dump value of the NPU Dump operator.

GroundTruth

Dump value of the Ground Truth operator.

Absolute Error

Absolute error. The value is the difference between the dump value of the NPU Dump operator and that of the Ground Truth operator. A maximum of 6 decimal places are allowed.

Relative Error

Relative error. The value is obtained by dividing the Absolute Error value by the dump value of the Ground Truth operator. If the dump value of the Ground Truth operator is 0, a hyphen (-) is displayed. A maximum of 6 decimal places are allowed.

Area 8

A % of tensor elements (B elements) have an absolute error greater than C.

The absolute error of all compared tensor elements of the current operator is A %. That is, the absolute error of B elements exceeds C. The absolute error can be relative error based on the metric value configured in area 5. The slider on the right controls the value of C. The value range is determined by the maximum and minimum values of AbsoluteError or RelativeError in area 6.

Cumulative error distribution line chart

The horizontal coordinate indicates AbsoluteError or RelativeError in area 6. The value range depends on the maximum and minimum values of AbsoluteError or RelativeError. The vertical coordinate indicates the accumulated percentage, that is, the percentage of all tensor elements that are less than or equal to the threshold when AbsoluteError or RelativeError reaches the threshold in all tensor elements.

Note: When you move the cursor to a tensor point, the tensor information is displayed, including Index (number of operator items to be compared), Absolute Error/Relative Error, and Cumulative Percentage.
Parent topic: Tensor Comparison