PyTorch Visualization for Comparison

Prerequisites

• You have performed operations in Environment Setup.
• Install tb-graph-ascend.

  pip install tb-graph-ascend

• You have performed operations in Accuracy Data Collection.

Procedure

1. Create a configuration file for comparison.
   For example, create a compare.json configuration file in the directory where the training script is located and copy the following content to the file:

   1 2 3 4 5

   { "npu_path": "./dump_data_npu", "bench_path": "./dump_data_gpu", "is_print_compare_log": true }

   The paths specified by npu_path and bench_path must be in the same environment.

2. Perform graph building for comparison.

   msprobe -f pytorch graph -i ./compare.json -o ./output

   After the comparison is complete, a .vis file is generated in ./output.

3. Start TensorBoard.

   tensorboard --logdir ./output --bind_all

   The path specified by --logdir is the ./output path in 2.

   After the preceding command is executed, the following log is displayed:

   1	TensorBoard 2.19.0 at http://ubuntu:6008/ (Press CTRL+C to quit)

   Open a browser on Windows and navigate to http://ubuntu:6008/. Replace ubuntu with the server's IP address, for example, http://192.168.1.10:6008/.

   If the access is successful, the TensorBoard page is displayed, as shown in the figure below.

   Figure 1 PyTorch Visualization for comparison
   

   In this example, level is set to L1 during data dumping. As a result, no model structure data is collected, and no data is available when performing PyTorch Visualization. The data shown in the preceding figure was collected when level was not set to L1.