fix aggregating sparse multiclass explanations to global in global-only streaming with batches mode #519
Conversation
imatiach-msft
force-pushed
the
ilmat/fix-sparse-multi
branch
3 times, most recently
from
882caad
to
fd029e5
Compare
…ly streaming with batches mode
imatiach-msft
force-pushed
the
ilmat/fix-sparse-multi
branch
from
fd029e5
to
b980d0b
Compare
Codecov Report
@@ Coverage Diff @@
## master #519 +/- ##
==========================================
+ Coverage 86.66% 86.70% +0.04%
==========================================
Files 62 62
Lines 4049 4063 +14
==========================================
+ Hits 3509 3523 +14
Misses 540 540
Flags with carried forward coverage won't be shown. Click here to find out more.
Continue to review full report at Codecov.
|
| @@ -1661,17 +1661,17 @@ def _get_local_explanation_row(explainer, evaluation_examples, i, batch_size): | |||
| return explainer.explain_local(rows) | |||
|
|
|||
|
|
|||
| def _aggregate_streamed_local_explanations(explainer, evaluation_examples, classification, features, | |||
| def _aggregate_streamed_local_explanations(explainer, evaluation_examples, task, features, | |||
There was a problem hiding this comment.
this seems a breaking change, do you mind check if any of our notebook or reference code in rai or aml does not have named parameter?
There was a problem hiding this comment.
It's not a breaking change, we are still passing the same parameter value, just the variable name that is passed was changed. Regardless, this method is private because it begins with an underscore anyway, so we can change it however we like.
There was a problem hiding this comment.
It's not an optional keyword argument, right?
Fix aggregating sparse multiclass explanations to global in global-only streaming with batches mode.
In the case that the explanations are
1.) computed on sparse multiclass data, and
2.) we set parameters to only compute the global explanation from local importance values, and
3.) we are using specifically the streaming aggregation from batches mode
(notice a lot of specific cases there), the logic currently fails as it is not designed to handle a list of sparse matrices. The logic for this scenario is implemented in this PR and a test is added.