I am using the torch.autograd.profiler and I am not clear how to best work with the data it produces.
My best solution has been adding a build_df() method to the profiler, which returns a pandas dataframe representation of the data, as opposed to an EventList or string, which is returned by key_averages() and table respectively. This allows me to perform the common statistical tests and data manipulation I need in a well documented manor.
I am curious what solutions others have used to work with the data that the profiler provides, perhaps there is an API which I missed.
Thanks!