Update ComponentGraph to enforce needing .x and .y for each component specified in the graph #2563
Conversation
…hs_make_pipelines
…eryx/evalml into 2493_component_graphs_make_pipelines
…eryx/evalml into 2493_component_graphs_make_pipelines
| y_train=y, | ||
| problem_type="binary", | ||
| max_iterations=1, | ||
| allowed_component_graphs=dummy_classifier_linear_component_graph, |
There was a problem hiding this comment.
allowed_component_graphs plays no role in the validity of this test, so removing :)
|
|
||
|
|
||
| @pytest.fixture | ||
| def dummy_classifier_dict_component_graph(dummy_classifier_estimator_class): |
There was a problem hiding this comment.
Removed these in favor of example_graph and example_regression_graph since they were confusing; they used dummy classifier but were labeled as Random Forest. Didn't provide much, if any, value over the example_graph fixture
| return TransformerA, TransformerB, TransformerC, EstimatorA, EstimatorB, EstimatorC | ||
|
|
||
|
|
||
| @pytest.fixture |
There was a problem hiding this comment.
Moved to conftest so that it could be used in other files
angela97lin
requested review from
ParthivNaresh,
bchen1116,
chukarsten,
dsherry,
freddyaboulton and
jeremyliweishih
|
@chukarsten I know you keep hoping for a large PR but hopefully, after splitting out the original PR, this is as big as it gets 😂 |
chukarsten
left a comment
chukarsten
left a comment
There was a problem hiding this comment.
Awesome, this is a really great PR and I really appreciate your cleanup as you go, philosophy. Looks really good. Thank you.
| self._i = 0 | ||
| self._compute_order = self.generate_order(self.component_dict) | ||
|
|
||
| def _validate_component_dict(self): |
There was a problem hiding this comment.
Like this function - super readable.
evalml/pipelines/component_graph.py
Outdated
| import pdb | ||
|
|
||
| pdb.set_trace() |
There was a problem hiding this comment.
NoooooooOOOooooooOOoooo
| "Imputer": {"numeric_impute_strategy": Categorical(["most_frequent", "mean"])}, | ||
| "Imputer_1": {"numeric_impute_strategy": Categorical(["median", "mean"])}, | ||
| "Random Forest Classifier": {"n_estimators": Categorical([50, 100])}, |
There was a problem hiding this comment.
How'd you get this past black without it exploding into 20+ lines?
There was a problem hiding this comment.
LOL I just double checked: my line is 88 chars long, and black's limit is 88 chars per line--so I just bareeeely made it 😅
| if is_linear | ||
| else dummy_classifier_dict_component_graph | ||
| ) | ||
| component_graph = {"CG": example_graph} |
Closes #2481, based on #2490 :)
The output of estimators (and all other components) must be accessed via
.xor.y; simply using the component name will no longer suffice. The convention to defaulting to.xif not specified was chosen when we did not have many (if any) components that outputted an appropriate .y value. Now that we do have more target transformers, I think enforcing explicit edges is helpful in clarity.This PR also deletes some fixtures and does some cleanup for our tests.