Pytorch profiler `add_metadata` api


I want to add some extra information when using the PyTorch profiler, and I found the add_metadata_json API in the official documentation of pytorch. But the doc did not explain how this function works and whether it’s possible to draw some self-defined charts on the TensorBoard. Could anyone give me some suggestions about how to use this API?

Based on this small unit test it seems that this metadata will be directly stored in the trace and you should be able to read it afterwards.

Hi, ptrblck! Thank you for your reply! I am developing a new tensorboard plugin for our project, and I really want to get it compatible with PyTorch Profiler. From your reply, I guess I can append some new JSON snippets and parse them in my plugin?

Yes, I think this is the case based on the posted unit test, but let us know if this approach is not working.