search

pytorch / captum

Model interpretability and understanding for PyTorch
https://captum.ai
BSD 3-Clause "New" or "Revised" License
4.8k stars 485 forks source link

No module named 'captum.influence' #929

Open jlubao opened 2 years ago

jlubao commented 2 years ago

📚 Documentation

The tutorial from https://captum.ai/tutorials/TracInCP_Tutorial throws an error when importing the needed libraries.

No module named 'captum.influence'

my library versions:

captum == 0.4.1 pytorch == 1.9.0

felixmeyjr commented 2 years ago

Hello. The tutorial is working with the versions: captum == 0.5.0 pytorch == 1.11.0

Maybe updating your libraries might help. BR

aobo-y commented 2 years ago

hi @jlubao , as @felixmeyjr said, captum.influence is a new function introduced in the latest version 0.5.0. You may refer to our change logs for details https://github.com/pytorch/captum/releases

felixmeyjr commented 2 years ago

Hi @aobo-y, are there plans to implement any other methods, i.e., such as Representer Points, Fisher Kernels or even Influence functions? Or does this not make sense due to current papers suggesting otherwise? Thanks

99warriors commented 2 years ago

Hi @felixmeyjr we are open to implementing other methods if there is an unmet need - would be interested to learn about your possible use cases.

felixmeyjr commented 2 years ago

Hi @99warriors, I'm currently looking into the Representer Points paper, to check if there are any additional use cases, where it would make sense to implement this. I'll let you know as soon as I know. Thanks. BR