Make check_is_fitted work again

`check_is_fitted` was changed (see #570) in a way that we won't raise
a `NotInitializedError` anymore despite the net not being
initialized. This is now solved by basically porting the old
`check_is_fitted` behavior to skorch.

skorch/utils.py

@@ -15,8 +15,6 @@
 import numpy as np
 from scipy import sparse
 from sklearn.utils import safe_indexing
-from sklearn.exceptions import NotFittedError
-from sklearn.utils.validation import check_is_fitted as sklearn_check_is_fitted
 import torch
 from torch.nn.utils.rnn import PackedSequence
 from torch.utils.data.dataset import Subset
@@ -494,15 +492,13 @@ def check_is_fitted(estimator, attributes, msg=None, all_or_any=all):
         msg = ("This %(name)s instance is not initialized yet. Call "
                "'initialize' or 'fit' with appropriate arguments "
                "before using this method.")
-    try:
-        sklearn_check_is_fitted(
-            estimator=estimator,
-            attributes=attributes,
-            msg=msg,
-            all_or_any=all_or_any,
-        )
-    except NotFittedError as e:
-        raise NotInitializedError(str(e))
+
+
+    if not isinstance(attributes, (list, tuple)):
+        attributes = [attributes]
+
+    if not all_or_any([hasattr(estimator, attr) for attr in attributes]):
+        raise NotInitializedError(msg % {'name': type(estimator).__name__})
 
 
 class TeeGenerator: