Parsing and Displaying Data
The Merge Reports function parses the raw profile data collected by Profiling and displays the parsed data in the MindStudio GUI. The configuration procedure is as follows:
- Choose on the menu bar.Figure 1 Merge Reports entry
- In the Merge Reports dialog box displayed, set the data source input directory (profile data flush path configured in the training script), data output directory, and output JSON file name.Figure 2 Merge Reports configuration
- Click OK to start execution.Figure 3 Merge Reports execution
- After the execution is successful, the Timeline View is automatically displayed, showing the AscendCL API calling and operator time consumption.
For details about the Timeline View page, see "Profiling > Profile Data Display > Timeline View" in the MindStudio User Guide.
Figure 4 Timeline view
If the size of the .json file exceeds the maximum cache for automatic result display of Merge Reports, the file may fail to be automatically displayed. In this case, manually open it.
Zoom in the view to check the time consumption of AscendCL APIs and operators. Switch to Statistics, select OP Info, and click Task Duration(us) to sort the operators. Then, find the operators with top time consumption and view and analyze their information.
Figure 5 Statistics tab page of Timeline View
Parent topic: Profiling (PyTorch Training)