The effect.item() objects there are the actual signed activation diffs, not the absolute diffs used earlier at the branchings stage. So I need a t.abs() call to measure unsigned effect sizes.
It may also turn out that positive-valued diffs have larger magnitudes than negative-valued diffs, and that negative-valued diffs are the objects of interest in the graphs. In that case, I'll have to write in code to ignore positive-valued diffs while flagging that.
The
effect.item()
objects there are the actual signed activation diffs, not the absolute diffs used earlier at the branchings stage. So I need at.abs()
call to measure unsigned effect sizes.It may also turn out that positive-valued diffs have larger magnitudes than negative-valued diffs, and that negative-valued diffs are the objects of interest in the graphs. In that case, I'll have to write in code to ignore positive-valued diffs while flagging that.