Accuracy Measurement Tool

The open-source tool Octave is introduced to analyze the relative error in a visualized manner.

Make the following preparations:

• Construct data: approximation and actual value. The approximation is typically the NumPy data, and the actual value is the operator output. Both values can be generated using ST cases.
• Install Octave: Download the installation package from https://www.gnu.org/software/octave/download.html.

The following describes the procedure of analyzing the relative error with Octave.

The following description involves only basic commands. For details about the tool instructions, visit https://octave.org/doc/interpreter/.

Figure 13-1 Example data.csv file

1. Load data.

   m = dlmread('data.csv', ','); 

   Load the data in the file to matrix m, where, data.csv is the name of the data file, and ',' is the file separator.

2. Read data.

   input = m(4:8940, 7);                      // Read all data in column 7.
   gpu_output = m(4:8940, 13);         // Read all data in column 13.
   model_output = m(4:8940, 19);      // Read all data in column 19.

3. Calculate the relative error and absolute error.

   gpu_diff = abs(model_output - gpu_output);                      // Calculate the relative error.
   gpu_error = diff ./ gpu_output;                                   // Calculate the absolute error.

   In the preceding command, gpu_output is the approximation, and model_output is the actual value.

4. Visualize the result.

   plot(input, gpu_error, 'o');
   xlabel('x')                  //x-axis label
   ylabel('value');            //y-axis label
   title(''Diff of model and cpu or gpu); //title
   hold on;                   //Draw again.
   plot (x, cpu_error, 'o');     //The process of cpu_error is omitted here, which is similar to that of gpu_error.
   hold off;
   Axis([0 1 -0.002 0.002]);             //x-axis and y-axis ranges
   legend('gpu_error','cpu_error');      //Legend

   The following figure shows a visualized result example.

   Figure 13-2 Visualized result using Octave