[CVPR 2021] Official PyTorch implementation for Transformer Interpretability Beyond Attention Visualization, a novel method to visualize classifications by Transformer based networks.
Existing code used the lrp instance instead of orig_lrp, which led to the same attribution as the transformer attribution. Following this change the produced attributions are different as expected.
Also updated the transformer visualization to use method="transformer_attribution" instead of the legacy method="grad". This change has not effect, and was made to align the visualization code with the rest of the repo for consistency.
Existing code used the lrp instance instead of orig_lrp, which led to the same attribution as the transformer attribution. Following this change the produced attributions are different as expected.
Also updated the transformer visualization to use method="transformer_attribution" instead of the legacy method="grad". This change has not effect, and was made to align the visualization code with the rest of the repo for consistency.