Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

Make EvalML compatible with Woodwork changes #4066

Merged
merged 23 commits into from
Mar 14, 2023
Merged

Make EvalML compatible with Woodwork changes #4066

Show file tree
Hide file tree
Changes from 21 commits
Commits
Show all changes
23 commits
Select commit Hold shift + click to select a range
4e76e5d
update file
bchen1116 Mar 9, 2023
4ed1c64
update release notes
bchen1116 Mar 9, 2023
0889ad3
lint
ParthivNaresh Mar 9, 2023
080a4cc
update latest dep
ParthivNaresh Mar 9, 2023
dc20f23
update min
ParthivNaresh Mar 9, 2023
0141795
update test
bchen1116 Mar 10, 2023
5996b78
Merge branch 'main' into bc_target_leakage_boolean
bchen1116 Mar 10, 2023
06b87dd
revert woodwork versions
ParthivNaresh Mar 13, 2023
e2d9401
Merge branch 'main' into make_compat_with_object_inference
ParthivNaresh Mar 13, 2023
02f5f7f
set ww min to 0.22.0
ParthivNaresh Mar 13, 2023
a7f8ef2
Merge branch 'make_compat_with_object_inference' of https://github.co…
ParthivNaresh Mar 13, 2023
a9c365c
empty
ParthivNaresh Mar 13, 2023
818c838
Merge remote-tracking branch 'origin/bc_target_leakage_boolean' into …
ParthivNaresh Mar 14, 2023
7ecf770
merge target leakage update
ParthivNaresh Mar 14, 2023
d0803e2
update wodwork for meta.yaml
ParthivNaresh Mar 14, 2023
9a0977c
remove print
ParthivNaresh Mar 14, 2023
92b979d
lint
ParthivNaresh Mar 14, 2023
c17717e
Update timeseries.ipynb
ParthivNaresh Mar 14, 2023
6c8e2b5
update release notes
ParthivNaresh Mar 14, 2023
032e1dd
Merge branch 'make_compat_with_object_inference' of https://github.co…
ParthivNaresh Mar 14, 2023
f0b820c
retrigger codecov
ParthivNaresh Mar 14, 2023
623c4d2
lint
ParthivNaresh Mar 14, 2023
632b124
trigger
ParthivNaresh Mar 14, 2023
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 1 addition & 1 deletion .github/meta.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -36,7 +36,7 @@ outputs:
- click>=8.0.0
- shap >=0.40.0
- texttable >=1.6.2
- woodwork >=0.21.1
- woodwork >=0.22.0
- featuretools>=1.16.0
- nlp-primitives>=2.9.0
- python >=3.8.*
Expand Down
2 changes: 2 additions & 0 deletions docs/source/release_notes.rst
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -23,6 +23,8 @@ Release Notes
* Testing Changes
* Use ``release.yaml`` for performance tests on merge to main :pr:`4007`
* Pin ``github-action-check-linked-issues`` at v1.4.5 :pr:`4042`
* Updated tests to support Woodwork's object dtype inference for numeric columns :pr:`4066`
* Updated ``TargetLeakageDataCheck`` tests to handle boolean targets properly :pr:`4061`
ParthivNaresh marked this conversation as resolved.
Show resolved Hide resolved

.. warning::

Expand Down
1 change: 1 addition & 0 deletions docs/source/user_guide/timeseries.ipynb
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -131,6 +131,7 @@
"outputs": [],
"source": [
"X[\"Categorical\"] = [str(i % 4) for i in range(len(X))]\n",
"X[\"Categorical\"] = X[\"Categorical\"].astype(\"category\")\n",
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

do you want to add a comment here and below explaining why we're setting it as category instead of leaving it as object? Might be good to remember.

"X[\"Numeric\"] = [i for i in range(len(X))]\n",
"\n",
"# Re-split the data since we modified X\n",
Expand Down
4 changes: 2 additions & 2 deletions evalml/tests/component_tests/test_column_selector_transformers.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -71,7 +71,7 @@ def test_select_by_type_empty_X():
[
lambda X, X_t: X_t.empty,
lambda X, X_t: X_t.empty,
lambda X, X_t: X_t.equals(X[["three"]].astype("int64")),
lambda X, X_t: X_t.equals(X[["one", "three"]].astype("int64")),
lambda X, X_t: X_t.astype(str).equals(X.astype(str)),
],
),
Expand Down Expand Up @@ -135,7 +135,7 @@ def test_column_transformer_transform(class_to_test, checking_functions):
SelectByType,
[
lambda X, X_t: X_t.empty,
lambda X, X_t: X_t.equals(X[["three"]].astype("int64")),
lambda X, X_t: X_t.equals(X[["one", "three"]].astype("int64")),
lambda X, X_t: X_t.astype(str).equals(X.astype(str)),
],
),
Expand Down
15 changes: 9 additions & 6 deletions evalml/tests/component_tests/test_target_encoder.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -121,29 +121,32 @@ def test_cols():
{
"col_1": [1, 2, 1, 1, 2] * 2,
"col_2": ["2", "1", "1", "1", "1"] * 2,
"col_3": ["a", "a", "a", "a", "a"] * 2,
"col_3": ["a", "a", "a", "a", "b"] * 2,
},
)
X_expected = X.astype({"col_1": "int64", "col_2": "category", "col_3": "category"})
X_expected = X.astype({"col_1": "int64", "col_2": "int64", "col_3": "category"})
y = pd.Series([0, 1, 1, 1, 0] * 2)
encoder = TargetEncoder(cols=[])
encoder.fit(X, y)
X_t = encoder.transform(X)
assert_frame_equal(X_expected, X_t)

encoder = TargetEncoder(cols=["col_2"])
encoder = TargetEncoder(cols=["col_3"])
encoder.fit(X, y)
X_t = encoder.transform(X)
X_expected = pd.DataFrame(
{
"col_1": pd.Series([1, 2, 1, 1, 2] * 2, dtype="int64"),
"col_2": [0.161365, 0.749863, 0.749863, 0.749863, 0.749863] * 2,
"col_3": pd.Series(["a", "a", "a", "a", "a"] * 2, dtype="category"),
"col_2": [2, 1, 1, 1, 1] * 2,
"col_3": pd.Series(
[0.749863, 0.749863, 0.749863, 0.749863, 0.161365] * 2,
dtype="float64",
),
},
)
assert_frame_equal(X_expected, X_t, check_less_precise=True)

encoder = TargetEncoder(cols=["col_2", "col_3"])
encoder = TargetEncoder(cols=["col_3"])
encoder.fit(X, y)
X_t = encoder.transform(X)
encoder2 = TargetEncoder()
Expand Down
19 changes: 13 additions & 6 deletions evalml/tests/data_checks_tests/test_id_columns_data_check.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -127,15 +127,15 @@ def test_id_columns_strings():
id_cols_check = IDColumnsDataCheck(id_threshold=1.0)
assert id_cols_check.validate(X) == [
DataCheckWarning(
message="Columns 'Id' are 100.0% or more likely to be an ID column",
message="Columns 'Id', 'col_3_id' are 100.0% or more likely to be an ID column",
data_check_name=id_data_check_name,
message_code=DataCheckMessageCode.HAS_ID_COLUMN,
details={"columns": ["Id"]},
details={"columns": ["Id", "col_3_id"]},
action_options=[
DataCheckActionOption(
DataCheckActionCode.DROP_COL,
data_check_name=id_data_check_name,
metadata={"columns": ["Id"]},
metadata={"columns": ["Id", "col_3_id"]},
),
],
).to_dict(),
Expand Down Expand Up @@ -293,17 +293,24 @@ def test_unidentified_first_col_primary_key(
)

id_cols_check = IDColumnsDataCheck(id_threshold=0.95)
if input_type == "string":
order = ["col_2", "col_3_id", "col_1_id"]
order_msg = "Columns 'col_2', 'col_3_id', 'col_1_id' are 95.0% or more likely to be an ID column"
else:
order = ["col_2", "col_1_id", "col_3_id"]
order_msg = "Columns 'col_2', 'col_1_id', 'col_3_id' are 95.0% or more likely to be an ID column"

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

What change caused this? It's a surprising distinction to make

Also, a nitpicky clarity change - we can replace the separate order messages here with order_msg = f"Columns '{order[0]}', '{order[1]}', '{order[2]}' are 95% or more..." just to make it clearer what the difference is!

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

In this case, col_3_id is a random series of string-represented numbers, which under the new inference will become the Integer ltype. Following this logic in validate, since they're all unique values and since the column ends in id, it gets identified as a possible ID column.

Previously however, since this column wasn't a numeric logical type, the logic here didn't catch it, but this did.

This just ends up resulting in a different ordering by which these columns are caught as possible ID columns by the data check

assert id_cols_check.validate(X) == [
DataCheckWarning(
message="Columns 'col_2', 'col_1_id', 'col_3_id' are 95.0% or more likely to be an ID column",
message=order_msg,
data_check_name=id_data_check_name,
message_code=DataCheckMessageCode.HAS_ID_COLUMN,
details={"columns": ["col_2", "col_1_id", "col_3_id"]},
details={"columns": order},
action_options=[
DataCheckActionOption(
DataCheckActionCode.DROP_COL,
data_check_name=id_data_check_name,
metadata={"columns": ["col_2", "col_1_id", "col_3_id"]},
metadata={"columns": order},
),
],
).to_dict(),
Expand Down
26 changes: 18 additions & 8 deletions evalml/tests/data_checks_tests/test_target_leakage_data_check.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -193,15 +193,15 @@ def test_target_leakage_types():

expected = [
DataCheckWarning(
message="Columns 'a', 'b' are 80.0% or more correlated with the target",
message="Columns 'a', 'b', 'c' are 80.0% or more correlated with the target",
data_check_name=target_leakage_data_check_name,
message_code=DataCheckMessageCode.TARGET_LEAKAGE,
details={"columns": ["a", "b"]},
details={"columns": ["a", "b", "c"]},
action_options=[
DataCheckActionOption(
DataCheckActionCode.DROP_COL,
data_check_name=target_leakage_data_check_name,
metadata={"columns": ["a", "b"]},
metadata={"columns": ["a", "b", "c"]},
),
],
).to_dict(),
Expand Down Expand Up @@ -356,8 +356,21 @@ def test_target_leakage_data_check_warnings_pearson():
y = y.astype(bool)

leakage_check = TargetLeakageDataCheck(pct_corr_threshold=0.5, method="pearson")
# pearsons does not support boolean columns
assert leakage_check.validate(X, y) == []
assert leakage_check.validate(X, y) == [
DataCheckWarning(
message="Columns 'a', 'b', 'c', 'd' are 50.0% or more correlated with the target",
data_check_name=target_leakage_data_check_name,
message_code=DataCheckMessageCode.TARGET_LEAKAGE,
details={"columns": ["a", "b", "c", "d"]},
action_options=[
DataCheckActionOption(
DataCheckActionCode.DROP_COL,
data_check_name=target_leakage_data_check_name,
metadata={"columns": ["a", "b", "c", "d"]},
),
],
).to_dict(),
]

y = y.astype(int)
assert leakage_check.validate(X, y) == [
Expand Down Expand Up @@ -447,9 +460,6 @@ def test_target_leakage_none_measures(measures):
X["b"] = y
y = y.astype(bool)

if measures in ["pearson", "spearman"]:
assert leakage_check.validate(X, y) == []
return
assert len(leakage_check.validate(X, y))


Expand Down
2 changes: 1 addition & 1 deletion evalml/tests/dependency_update_check/latest_dependency_versions.txt
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -33,5 +33,5 @@ statsmodels==0.13.5
texttable==1.6.7
tomli==2.0.1
vowpalwabbit==9.7.0
woodwork==0.21.2
woodwork==0.22.0
xgboost==1.7.4
2 changes: 1 addition & 1 deletion evalml/tests/dependency_update_check/minimum_requirements.txt
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -32,5 +32,5 @@ statsmodels==0.12.2
texttable==1.6.2
tomli==2.0.1
vowpalwabbit==8.11.0
woodwork==0.21.1
woodwork==0.22.0
xgboost==1.7.0
2 changes: 1 addition & 1 deletion evalml/tests/dependency_update_check/minimum_test_requirements.txt
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -41,5 +41,5 @@ statsmodels==0.12.2
texttable==1.6.2
tomli==2.0.1
vowpalwabbit==8.11.0
woodwork==0.21.1
woodwork==0.22.0
xgboost==1.7.0
15 changes: 14 additions & 1 deletion evalml/tests/model_understanding_tests/prediction_explanations_tests/test_explainers.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1148,7 +1148,20 @@ def test_json_serialization(
num_to_explain=1,
output_format="dict",
)
assert json.loads(json.dumps(best_worst)) == best_worst

def str_to_num_decoder(dct):
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

can you explain why we need this now? do we need to add support directly into prediction explanations?

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I'm concerned by this change as well. I think json.loads(json.dumps(best_worst)) == best_worst is an axiom we should maintain, rather than patching it over for the sake of the upgrade. What change predicated this, and how can we account for it in prediction explanations instead?

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This is because the output of explain_predictions_best_worst results in the keys of predicted_values being integers, since they're numeric values being represented as strings. Woodwork's new inference will infer strings as their appropriate numeric logical type if their incoming dtype is object and no other logical type is specified.

'predicted_values': {'error_name': 'Cross Entropy',
                                   'error_value': 9.967001210947085e-07,
                                   'index_id': 33,
                                   'predicted_value': 1,
                                   'probabilities': {0: 0.0, 1: 1.0},
                                   'target_value': 1},

but JSON doesn't allow non-string keys, so it replaces them with

'predicted_values': {'error_name': 'Cross Entropy',
                                   'error_value': 9.967001210947085e-07,
                                   'index_id': 33,
                                   'predicted_value': 1,
                                   'probabilities': {'0': 0.0, '1': 1.0}, # String keys
                                   'target_value': 1},

What I added here was an object_hook to be called instead of the standard dict which converts the string key back to an int just so they can be compared for the test.

I think what's happening is that this is creating keys and assigning them to the probabilities key even if they're integers. I can modify this to stringify all keys since JSON will end up converting them anyways. I just thought it would be better for the output of explain_predictions_best_worst to represent the actual values even if JSON would end up stringifying them. @eccabay @jeremyliweishih does that make sense?

Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

wait @ParthivNaresh why does woodwork inference matter when we're testing if dumping and loading the same json will result in the same dictionary?

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

@jeremyliweishih Because JSON only allows key names to be strings, so if the values themselves aren't strings, then they get dumped as integers and read as strings. I think somewhere in the prediction explanation process, the original dtype of y is lost and reinferred, which means if the original dtype was a string, it would be reinferred as an integer.

I tried replacing

elif problem_type == problem_type.BINARY:
        X, y = X_y_binary
        y = pd.Series(y).astype("str")
        pipeline = logistic_regression_binary_pipeline

with

elif problem_type == problem_type.BINARY:
        X, y = X_y_binary
        y = ww.init_series(pd.Series(y), logical_type="Unknown")
        pipeline = logistic_regression_binary_pipeline

or explicitly setting it as string dtype (string instead of str), but it still writes out the dictionary in best_worst as an integer

Copy link
Contributor

@eccabay eccabay Mar 14, 2023 •
edited
Loading

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I just thought it would be better for the output of explain_predictions_best_worst to represent the actual values even if JSON would end up stringifying them

Doesn't your second comment contradict this, so explain_predictions_best_worst is still modifying the actual values, but in the wrong direction?

My preference remains that we keep the assertion that json.loads(json.dumps(best_worst)) == best_worst, and we update the code to maintain this rather than the test.

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

explain_predictions_best_worst is modifying this somewhere but that's not the expectation. If I pass in a string dtype (which Woodwork won't infer as numeric, but Categorical in this case) then that would be the expected output.

I'm taking a look into explain_predictions_best_worst to see where this unexpected reinference is happening

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

@eccabay I think the issue was that ww.init_series was being used to determine the unique classes that existed in a classification pipeline. If the column names were strings then they got inferred as integers because of this.

new_dct = {}
for k, v in dct.items():
if isinstance(k, str) and k.isdigit():
k = int(k)
if isinstance(v, dict):
v = str_to_num_decoder(v)
new_dct[k] = v
return new_dct

assert (
json.loads(json.dumps(best_worst), object_hook=str_to_num_decoder) == best_worst
)

report = explain_predictions(
pipeline,
Expand Down
23 changes: 17 additions & 6 deletions evalml/tests/model_understanding_tests/test_partial_dependence.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1425,10 +1425,14 @@ def test_graph_partial_dependence_regression_and_binary_categorical(

X = pd.DataFrame(X)
X.columns = [str(i) for i in range(X.shape[1])]
X["categorical_column"] = pd.Series([i % 3 for i in range(X.shape[0])]).astype(
X["categorical_column"] = pd.Series(
[f"cat_{i % 3}" for i in range(X.shape[0])],
).astype(
"str",
)
X["categorical_column_2"] = pd.Series([i % 6 for i in range(X.shape[0])]).astype(
X["categorical_column_2"] = pd.Series(
[f"cat_{i % 6}" for i in range(X.shape[0])],
).astype(
"str",
)

Expand All @@ -1442,7 +1446,7 @@ def test_graph_partial_dependence_regression_and_binary_categorical(
)
plot_data = fig.to_dict()["data"][0]
assert plot_data["type"] == "bar"
assert list(plot_data["x"]) == ["0", "1", "2"]
assert list(plot_data["x"]) == ["cat_0", "cat_1", "cat_2"]

fig = graph_partial_dependence(
pipeline,
Expand All @@ -1453,7 +1457,7 @@ def test_graph_partial_dependence_regression_and_binary_categorical(
fig_dict = fig.to_dict()
plot_data = fig_dict["data"][0]
assert plot_data["type"] == "contour"
assert fig_dict["layout"]["yaxis"]["ticktext"] == ["0", "1", "2"]
assert fig_dict["layout"]["yaxis"]["ticktext"] == ["cat_0", "cat_1", "cat_2"]
assert (
fig_dict["layout"]["title"]["text"]
== "Partial Dependence of 'categorical_column' vs. '0'"
Expand All @@ -1468,8 +1472,15 @@ def test_graph_partial_dependence_regression_and_binary_categorical(
fig_dict = fig.to_dict()
plot_data = fig_dict["data"][0]
assert plot_data["type"] == "contour"
assert fig_dict["layout"]["xaxis"]["ticktext"] == ["0", "1", "2"]
assert fig_dict["layout"]["yaxis"]["ticktext"] == ["0", "1", "2", "3", "4", "5"]
assert fig_dict["layout"]["xaxis"]["ticktext"] == ["cat_0", "cat_1", "cat_2"]
assert fig_dict["layout"]["yaxis"]["ticktext"] == [
"cat_0",
"cat_1",
"cat_2",
"cat_3",
"cat_4",
"cat_5",
]
assert (
fig_dict["layout"]["title"]["text"]
== "Partial Dependence of 'categorical_column_2' vs. 'categorical_column'"
Expand Down
2 changes: 1 addition & 1 deletion pyproject.toml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -41,7 +41,7 @@ dependencies = [
"shap >= 0.40.0",
"statsmodels >= 0.12.2",
"texttable >= 1.6.2",
"woodwork >= 0.21.1",
"woodwork >= 0.22.0",
"dask >= 2022.2.0, != 2022.10.1",
"featuretools >= 1.16.0",
"nlp-primitives >= 2.9.0",
Expand Down