Closed sangkeun00 closed 4 months ago
For more detailed influence analysis, one may want to look into the influence score of each module (or a group of modules). Users can specify their own computation groups as below
influence_groups = ["att", "mlp"] run.influence.compute_influence_all(..., influence_groups=influence_groups) run.influence.compute_influence(..., influence_groups=influence_groups) run.influence.compute_self_influence(..., influence_groups=influence_groups)
For more detailed influence analysis, one may want to look into the influence score of each module (or a group of modules). Users can specify their own computation groups as below