Fix issue 231, per-sample weights incorrect when using valid_size > 0.0

sknn/backend/lasagne/mlp.py

@@ -246,10 +246,19 @@ def _initialize_impl(self, X, y=None, w=None):
 
         if self.valid_size > 0.0:
             assert self.valid_set is None, "Can't specify valid_size and valid_set together."
-            X, X_v, y, y_v = sklearn.cross_validation.train_test_split(
-                                X, y,
-                                test_size=self.valid_size,
-                                random_state=self.random_state)
+            indices = numpy.arange(X.shape[0])
+            indices, indices_v = sklearn.cross_validation.train_test_split(
+                                    indices,
+                                    test_size=self.valid_size,
+                                    random_state=self.random_state)
+            X_v = X[indices_v]
+            y_v = y[indices_v]
+
+            X = X[indices]
+            y = y[indices]
+            if w is not None:
+                w = w[indices]
+
             self.valid_set = X_v, y_v
 
         if self.valid_set and self.is_convolution():
@@ -263,7 +272,7 @@ def _initialize_impl(self, X, y=None, w=None):
             params.extend(mlp_layer.get_params())
 
         self.trainer, self.validator = self._create_mlp_trainer(params)
-        return X, y
+        return X, y, w
 
     def _predict_impl(self, X):
         if self.is_convolution():

sknn/mlp.py

@@ -210,7 +210,7 @@ def _fit(self, X, y, w=None):
         X, y = self._reshape(X, y)
 
         if not self.is_initialized:
-            X, y = self._initialize(X, y, w)
+            X, y, w = self._initialize(X, y, w)
 
         log.info("Training on dataset of {:,} samples with {} total size.".format(data_shape[0], data_size))
         if data_shape[1:] != X.shape[1:]: