Skip to content
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
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
14 changes: 10 additions & 4 deletions deeptab/models/base.py
Original file line number Diff line number Diff line change
Expand Up @@ -126,7 +126,11 @@ def _validate_fit_inputs(X, y, regression: bool, family: str | None = None, X_va
"""Validate training and paired validation inputs before estimator mutation.

Targets must be finite and valid for the task. Explicit validation features
must match training width and, for named DataFrames, ordered columns.
must match training width and, when the training frame carries string column
names, the same ordered columns. Unnamed validation input (an array or a
frame with positional labels) is rejected for a named training frame here,
before any fitted state is replaced, rather than failing inside the fitted
preprocessor once the model has been rebuilt.
"""
from deeptab.core.sklearn_compat import ensure_dataframe

Expand All @@ -152,10 +156,12 @@ def _validate_fit_inputs(X, y, regression: bool, family: str | None = None, X_va
)
if validation_features.shape[1] != features.shape[1]:
raise column_count_error(features.shape[1], validation_features.shape[1])
if hasattr(X, "columns") and hasattr(X_val, "columns") and list(X.columns) != list(X_val.columns):
training_columns = list(features.columns)
validation_columns = list(validation_features.columns)
if all(isinstance(name, str) for name in training_columns) and validation_columns != training_columns:
raise column_name_error(
missing=[str(name) for name in X.columns if name not in X_val.columns],
extra=[str(name) for name in X_val.columns if name not in X.columns],
missing=[name for name in training_columns if name not in validation_columns],
extra=[str(name) for name in validation_columns if name not in training_columns],
)
if not regression or family == "categorical":
classes = np.unique(targets)
Expand Down
7 changes: 6 additions & 1 deletion docs/core_concepts/sklearn_api.md
Original file line number Diff line number Diff line change
Expand Up @@ -109,7 +109,12 @@ point regressors accept one numeric target column. Distributional models retain
their task-specific target dimensions, including Dirichlet component columns.
When supplying an explicit validation set, provide both `X_val` and `y_val`.
Feature widths, ordered DataFrame columns, and target widths must match training
data; classifier validation labels must occur in the training classes.
data; classifier validation labels must occur in the training classes. When `X`
is a DataFrame with string column names, `X_val` must carry the same names in
the same order: an unnamed array, a list of rows, or a frame with positional
labels is rejected with a `ColumnNameError` before any fitted state changes.
Training on an array and validating on a named frame remains possible, because
array columns are used by position.

Preflight input rejection leaves a previously fitted estimator unchanged,
including classes, preprocessing, weights, and predictions. A failure during
Expand Down
30 changes: 30 additions & 0 deletions tests/integration/test_estimator_input_validation.py
Original file line number Diff line number Diff line change
Expand Up @@ -120,6 +120,36 @@ def test_all_missing_column_fits_and_predicts(self):
preds = model.predict(X)
assert preds.shape == (40,)

@pytest.mark.parametrize("estimator_name", ["MLPClassifier", "MLPRegressor", "MLPLSS"])
def test_unnamed_validation_features_for_named_frame_are_rejected_before_building(self, estimator_name):
"""An array X_val for a DataFrame X used to pass preflight and crash inside the
fitted preprocessor's column transformer after the model had been rebuilt."""
import deeptab.models as models
from deeptab.core.exceptions import ColumnNameError

rng = np.random.default_rng(0)
X = pd.DataFrame(rng.standard_normal((48, 3)), columns=pd.Index(["a", "b", "c"]))
y = rng.integers(0, 2, 48) if estimator_name == "MLPClassifier" else rng.standard_normal(48)
options = {"family": "normal"} if estimator_name == "MLPLSS" else {}

model = getattr(models, estimator_name)()
with pytest.raises(ColumnNameError, match=r"Missing : \['a', 'b', 'c'\]"):
model.fit(X.iloc[:32], y[:32], X_val=X.iloc[32:].to_numpy(), y_val=y[32:], max_epochs=1, **options)
assert model._data_module is None
assert model._task_model is None
assert not model._built

def test_named_validation_frame_for_array_training_data_fits(self):
from deeptab.models import MLPRegressor

rng = np.random.default_rng(0)
X = pd.DataFrame(rng.standard_normal((48, 3)), columns=pd.Index(["a", "b", "c"]))
y = rng.standard_normal(48)

model = MLPRegressor()
model.fit(X.iloc[:32].to_numpy(), y[:32], X_val=X.iloc[32:], y_val=y[32:], max_epochs=1, batch_size=16)
assert model.predict(X.iloc[:5].to_numpy()).shape == (5,)


class TestIntegerColumnLabels:
"""Integer column labels that are not 0..n-1 used to be read by position for
Expand Down
20 changes: 20 additions & 0 deletions tests/lifecycle/test_failed_refit.py
Original file line number Diff line number Diff line change
Expand Up @@ -286,6 +286,26 @@ def test_named_validation_schema_preserves_state(tmp_path):
_assert_preserved(model, validation_features, snapshot)


@pytest.mark.parametrize("container", ["ndarray", "positional_frame", "list"])
def test_unnamed_validation_features_preserve_state(tmp_path, container):
"""An unnamed X_val for a named training frame is rejected before the fitted
model is replaced, instead of failing inside the fitted preprocessor afterwards."""
features, labels, validation_features, validation_labels = _data()
features = pd.DataFrame(features, columns=pd.Index(["first", "second", "third"]))
validation_features = pd.DataFrame(validation_features, columns=features.columns)
model = _model(MLPClassifier, tmp_path)
model.fit(features, labels, X_val=validation_features, y_val=validation_labels, **_fit_options())
snapshot = _snapshot(model, validation_features)
unnamed = {
"ndarray": validation_features.to_numpy(),
"positional_frame": pd.DataFrame(validation_features.to_numpy()),
"list": validation_features.to_numpy().tolist(),
}[container]
with pytest.raises(DataError, match="first"):
model.fit(features, labels, X_val=unnamed, y_val=validation_labels, **_fit_options())
_assert_preserved(model, validation_features, snapshot)


def test_dirichlet_refit_preserves_target_dimensions_and_selected_family(tmp_path):
features, _, validation_features, _ = _data()
targets = np.random.default_rng(19).dirichlet([2, 3, 4], size=len(features))
Expand Down
Loading