Enabling Profiling in a Training Project

Perform the following operations in MindStudio.

1. Choose File > Open... on the menu bar, select the migrated training project, and click OK to import the project. See Figure 1.
   Figure 1 Training project import
   
2. Modify the training script.
   • Edit the train.py file to set the device for training and add the Profiling enabling code.
     • Set the device and add the following code:

       NPU_DEVICE='npu:1'
       torch.npu.set_device(NPU_DEVICE)

       Figure 2 Device setting
       
     • Add the Profiling enabling code and configure the flush path of profile data. See Figure 3.

       with torch.npu.profile(profiler_result_path="/tmp/pytorch/profiling"):

       Figure 3 Profiling enabling
       
   • To accelerate training and profile data collection, edit the global_settings.py file in the conf directory to set EPOCH to 1.
     Figure 4 EPOCH setting
     
3. Run the training script. For details, see Training After Migration.