search

pytorch / captum

Model interpretability and understanding for PyTorch
https://captum.ai
BSD 3-Clause "New" or "Revised" License
4.91k stars 494 forks source link

Captum for regression problem #206

Closed Hailey-Liu closed 4 years ago

Hailey-Liu commented 4 years ago

Hi all,

I am wondering if there are examples that I could learn to use Captum for regression problem as well as using volume data. My problem setting is feeding volume data with WxHxD (64x64x64) to a 3D convnet which has only one neuron in the top layer that output a real number. Thanks.

NarineK commented 4 years ago

Hi @Hailey-Liu , if you leave target=None, you should be able to attribute the output, a real value, to the inputs and layers( in case of layer algorithms) to the network. We don't have tutorials but we have test models which aren't classification models: https://github.com/pytorch/captum/blob/master/tests/attr/helpers/basic_models.py#L16

Have you tried to do attribution and encountered any problems ?

Hailey-Liu commented 4 years ago

I see. Thanks for your reply @NarineK . I haven't tried the attribution, still working on familiarizing myself with the Captum.