Partially revert #519 due to performance regression & other issues (#521)

test/test_transforms.py

@@ -652,7 +652,7 @@ class Tester(unittest.TestCase):
         # Checking if RandomHorizontalFlip can be printed as string
         transforms.RandomHorizontalFlip().__repr__()
 
-    @unittest.skipIf(stats is None, 'scipt.stats is not available')
+    @unittest.skipIf(stats is None, 'scipy.stats is not available')
     def test_normalize(self):
         def samples_from_standard_normal(tensor):
             p_value = stats.kstest(list(tensor.view(-1)), 'norm', args=(0, 1)).pvalue

torchvision/transforms/functional.py

@@ -167,9 +167,10 @@ def normalize(tensor, mean, std):
     if not _is_tensor_image(tensor):
         raise TypeError('tensor is not a torch image.')
 
-    mean = torch.Tensor(mean).view((tensor.shape[0], 1, 1))
-    std = torch.Tensor(std).view((tensor.shape[0], 1, 1))
-    return tensor.sub_(mean).div_(std)
+    # This is faster than using broadcasting, don't change without benchmarking
+    for t, m, s in zip(tensor, mean, std):
+        t.sub_(m).div_(s)
+    return tensor
 
 
 def resize(img, size, interpolation=Image.BILINEAR):