Added 'SelectFirstTransformer' transformation type

sklearn2pmml/ensemble/__init__.py

@@ -179,6 +179,19 @@ def predict_proba(self, X, **predict_proba_params):
 				result.append(proba[i - 1] - proba[i])
 		return numpy.asarray(result).T
 
+def _to_sparse(X, step_mask, step_result):
+	# Make array
+	if len(step_result.shape) == 1:
+		result = numpy.empty((X.shape[0], ), dtype = object)
+	else:
+		result = numpy.empty((X.shape[0], step_result.shape[1]), dtype = object)
+	# Fill array
+	if len(step_result.shape) == 1:
+		result[step_mask.ravel()] = step_result
+	else:
+		result[step_mask.ravel(), :] = step_result
+	return result
+
 class _BaseEnsemble(_BaseComposition):
 
 	def __init__(self, steps, controller):
@@ -216,19 +229,6 @@ def _to_evaluation_dataset(self, X):
 			return self.controller.transform(X)
 		return X
 
-def _to_sparse(X, step_mask, step_result):
-	# Make array
-	if len(step_result.shape) == 1:
-		result = numpy.empty((X.shape[0], ), dtype = object)
-	else:
-		result = numpy.empty((X.shape[0], step_result.shape[1]), dtype = object)
-	# Fill array
-	if len(step_result.shape) == 1:
-		result[step_mask.ravel()] = step_result
-	else:
-		result[step_mask.ravel(), :] = step_result
-	return result
-
 class Link(BaseEstimator):
 
 	def __init__(self, estimator, augment_funcs, prefit = False):

sklearn2pmml/preprocessing/__init__.py

@@ -10,7 +10,7 @@
 from scipy.sparse import lil_matrix
 from sklearn.base import BaseEstimator, TransformerMixin
 from sklearn.pipeline import Pipeline
-from sklearn2pmml.util import cast, dt_transform, ensure_1d, ensure_def, eval_rows, to_expr_func, Expression
+from sklearn2pmml.util import cast, dt_transform, ensure_1d, ensure_def, eval_rows, eval_expr_rows, to_expr_func, Expression, Predicate
 
 import numpy
 import pandas
@@ -606,3 +606,65 @@ def transform(self, X):
 		if self.trim_blanks:
 			Xt = numpy.char.strip(Xt)
 		return Xt
+
+def _to_sparse(X, step_mask, step_result):
+	# Make array
+	result = numpy.empty((X.shape[0], step_result.shape[1]), dtype = object)
+	# Fill array
+	result[step_mask.ravel(), :] = step_result
+	return result
+
+class SelectFirstTransformer(BaseEstimator, TransformerMixin):
+
+	def __init__(self, steps, controller = None):
+		for step in steps:
+			if type(step) is not tuple:
+				raise TypeError("Step is not a tuple")
+			if len(step) != 3:
+				raise TypeError("Step is not a three-element (name, transformer, predicate) tuple")
+			name, transformer, predicate = step
+			if not isinstance(predicate, (str, Predicate)):
+				raise TypeError()
+		self.steps = steps
+		if controller:
+			if not hasattr(controller, "transform"):
+				raise TypeError()
+		self.controller = controller
+
+	def _to_evaluation_dataset(self, X):
+		if self.controller is not None:
+			return self.controller.transform(X)
+		return X
+
+	def fit(self, X, y = None):
+		X_eval = self._to_evaluation_dataset(X)
+		mask = numpy.zeros(X.shape[0], dtype = bool)
+		for name, transformer, predicate in self.steps:
+			step_mask = eval_expr_rows(X_eval, predicate, dtype = bool)
+			step_mask[mask] = False
+			if numpy.sum(step_mask) < 1:
+				raise ValueError(predicate)
+			step_X = X[step_mask]
+			step_y = y[step_mask] if y is not None else None
+			transformer.fit(step_X, step_y)
+			mask = numpy.logical_or(mask, step_mask)
+		return self
+
+	def transform(self, X):
+		result = None
+		X_eval = self._to_evaluation_dataset(X)
+		mask = numpy.zeros(X.shape[0], dtype = bool)
+		for name, transformer, predicate in self.steps:
+			step_mask = eval_expr_rows(X_eval, predicate, dtype = bool)
+			step_mask[mask] = False
+			if numpy.sum(step_mask) < 1:
+				continue
+			step_X = X[step_mask]
+			step_result = transformer.transform(step_X)
+			step_result = _to_sparse(X, step_mask, step_result)
+			if result is None:
+				result = step_result
+			else:
+				result[step_mask] = step_result[step_mask]
+			mask = numpy.logical_or(mask, step_mask)
+		return result