Update svhn.py to be consistent with other datasets

Make the labels be 1d tensors of type `np.int64`.
Assign label `0`  to data samples of digit 0. The original dataset assigns them label `10`.

torchvision/datasets/svhn.py

@@ -66,7 +66,16 @@ class SVHN(data.Dataset):
         loaded_mat = sio.loadmat(os.path.join(root, self.filename))
 
         self.data = loaded_mat['X']
-        self.labels = loaded_mat['y']
+        # loading from the .mat file gives an np array of type np.uint8
+        # converting to np.int64, so that we have a LongTensor after 
+        # the conversion from the numpy array
+        # the squeeze is needed to obtain a 1D tensor
+        self.labels = loaded_mat['y'].astype(np.int64).squeeze()
+        
+        # the svhn dataset assigns the class label "10" to the digit 0
+        # this makes it inconsistent with several loss functions 
+        # which expect the class labels to be in the range [0, C-1]
+        np.place(self.labels, self.labels==10, 0)
         self.data = np.transpose(self.data, (3, 2, 0, 1))
 
     def __getitem__(self, index):