Closed oliveradk closed 4 months ago
Thanks for the feedback! I mostly implemented your suggestions, but instead of adding a dimension to patch_mask by default, only add the batch dimension if set_mask_batch_size
is called - this minimizes the chance of any downstream bugs being introduced, and I don't think any of the edge indexing functionality is critical if the main use case is to collect attribution scores over batches (please let me know if I'm missing something crucial there)
I will just do a couple of checks locally and then merge in next hour or so
@oliveradk Could you please fix merge conflicts and final comment? Then I will merge.
Fixed merged conflicts and addressed comment (I had been running the test incorrectly and had to make some more tweaks)
Adds support for computing input-wise mask gradients. Useful for e.g. doing anomaly detection using edge attribution scores