The relprop method of the Linear layer

hila-chefer / Transformer-Explainability

[CVPR 2021] Official PyTorch implementation for Transformer Interpretability Beyond Attention Visualization, a novel method to visualize classifications by Transformer based networks.

MIT License

1.75k stars 232 forks source link

The relprop method of the Linear layer #64

Open Hongbo-Z opened 8 months ago

Hongbo-Z commented 8 months ago

HI Chefer,

Is there a typo with Lines 218 to 219 https://github.com/hila-chefer/Transformer-Explainability/blob/c3e578f76b954e8528afeaaee26de3f07e3fe559/modules/layers_ours.py#L218-L219

which should be the below? S1 = safe_divide(R, Z1) S2 = safe_divide(R, Z2) according to https://github.com/wjNam/Relative_Attributing_Propagation/blob/7fa96822740591b605712f251e556ec8487d1eea/modules/layers.py#L268-L269C36

Thanks,

Hongbo

michaelkiper commented 3 months ago

I want to follow this issue as well. It's something I noticed as well. Curious why you add by both the Z1 and Z2 terms to project the Relevance