Better error message when data conversion fails.

Specifically, when conversion to torch tensor or to numpy array fails,
users would often only get "AttributeError: bla has no attribute
.cuda". Now a more specific TypeError is shown.

skorch/dataset.py

@@ -289,7 +289,7 @@ def check_cv(self, y):
             # doesn't work, still try.
             try:
                 y_arr = to_numpy(y)
-            except AttributeError:
+            except (AttributeError, TypeError):
                 y_arr = y
 
         if self._is_float(self.cv):

skorch/utils.py

@@ -23,6 +23,11 @@ class Ansi(Enum):
     ENDC = '\033[0m'
 
 
+def is_torch_data_type(x):
+    # pylint: disable=protected-access
+    return isinstance(x, (torch.tensor._TensorBase, Variable))
+
+
 def to_var(X, use_cuda):
     """Generic function to convert a input data to pytorch Variables.
 
@@ -73,6 +78,9 @@ def to_tensor(X, use_cuda):
     elif np.isscalar(X):
         X = torch.from_numpy(np.array([X]))
 
+    if not is_torch_data_type(X):
+        raise TypeError("Cannot convert this data type to a torch tensor.")
+
     if use_cuda:
         X = X.cuda()
     return X
@@ -91,6 +99,9 @@ def to_numpy(X):
     if is_pandas_ndframe(X):
         return X.values
 
+    if not is_torch_data_type(X):
+        raise TypeError("Cannot convert this data type to a numpy array.")
+
     if X.is_cuda:
         X = X.cpu()