fix n_features_in setting

Author: jameslamb

python-package/lightgbm/compat.py

@@ -48,7 +48,7 @@ def validate_data(
             # NOTE: check_X_y() calls check_array() internally, so only need to call one or the other of them here
             if isinstance(y, str) and y == "no_validation":
                 X = check_array(X, accept_sparse=accept_sparse, force_all_finite=ensure_all_finite)
-                return X
+                return X  # noqa: RET504
 
             # if we reach here, we're validating features and labels
             X, y = check_X_y(

python-package/lightgbm/sklearn.py

@@ -147,6 +147,32 @@ def _get_weight_from_constructed_dataset(dataset: Dataset) -> Optional[np.ndarra
     return weight
 
 
+def _num_features_for_raw_input(X: _LGBM_ScikitMatrixLike) -> int:
+    """Get number of features from raw input data.
+
+    Some ``scikit-learn`` versions accomplish this with a private function
+    ``sklearn.utils.validation._num_features()``. This is here to avoid depending on that.
+    """
+    # if a list, assume a list of lists where each item is one row of training data,
+    # and all rows have the same number of features
+    if isinstance(X, list):
+        return len(X[0])
+
+    # scikit-learn estimator checks error out if .shape is accessed unconditionally...
+    # but allow hard-coding an assumption that anything that isn't a list and doesn't have .shape
+    # must be convertible to something following the Array API via a __array__() method
+    if hasattr(X, "shape"):
+        X_shape = X.shape
+    else:
+        X_shape = X.__array__().shape
+
+    # this condition accounts for training on a 1-dimensional input
+    if len(X_shape) == 1:
+        return 1
+    else:
+        return X_shape[1]
+
+
 class _ObjectiveFunctionWrapper:
     """Proxy class for objective function."""
 
@@ -906,13 +932,10 @@ def fit(
         params["metric"] = [e for e in eval_metrics_builtin if e not in params["metric"]] + params["metric"]
         params["metric"] = [metric for metric in params["metric"] if metric is not None]
 
-        # this needs to be set before calling _LGBMValidateData(), as that function uses
-        # self.n_features_in_ (the corresponding public attribute) to check that the input has
-        # the expected number of features
-        if isinstance(X, list):
-            self._n_features_in = len(X[0])
-        else:
-            self._n_features_in = X.shape[1]
+        # `sklearn.utils.validation.validate_data()` expects self.n_features_in_ to already be set by the
+        # time it's called (if you call it with reset=True like LightGBM does), and the scikit-learn
+        # estimator checks complain if X.shape is accessed unconditionally
+        self._n_features_in = _num_features_for_raw_input(X)
 
         if not isinstance(X, (pd_DataFrame, dt_DataTable)):
             _X, _y = _LGBMValidateData(