Fix undefined behaviour: preprocess_input copying/not copying the input arrays (#10153)

* Add copy option for image preprocessing

* Fix unnecessary import

* Fix style

* fix test error

* Make modifications in-place instead

keras/applications/imagenet_utils.py

@@ -38,7 +38,8 @@ def _preprocess_numpy_input(x, data_format, mode):
     # Returns
         Preprocessed Numpy array.
     """
-    x = x.astype(K.floatx())
+    if not issubclass(x.dtype.type, np.floating):
+        x = x.astype(K.floatx(), copy=False)
 
     if mode == 'tf':
         x /= 127.5
@@ -153,6 +154,9 @@ def preprocess_input(x, data_format=None, mode='caffe'):
 
     # Arguments
         x: Input Numpy or symbolic tensor, 3D or 4D.
+            The preprocessed data is written over the input data
+            if the data types are compatible. To avoid this
+            behaviour, `numpy.copy(x)` can be used.
         data_format: Data format of the image tensor/array.
         mode: One of "caffe", "tf" or "torch".
             - caffe: will convert the images from RGB to BGR,

tests/keras/applications/imagenet_utils_test.py

@@ -38,6 +38,22 @@ def test_preprocess_input():
     assert_allclose(out1, out2.transpose(1, 2, 0))
     assert_allclose(out1int, out2int.transpose(1, 2, 0))
 
+    # Test that writing over the input data works predictably
+    for mode in ['torch', 'tf']:
+        x = np.random.uniform(0, 255, (2, 10, 10, 3))
+        xint = x.astype('int')
+        x2 = utils.preprocess_input(x, mode=mode)
+        xint2 = utils.preprocess_input(xint)
+        assert_allclose(x, x2)
+        assert xint.astype('float').max() != xint2.max()
+    # Caffe mode works differently from the others
+    x = np.random.uniform(0, 255, (2, 10, 10, 3))
+    xint = x.astype('int')
+    x2 = utils.preprocess_input(x, data_format='channels_last', mode='caffe')
+    xint2 = utils.preprocess_input(xint)
+    assert_allclose(x, x2[..., ::-1])
+    assert xint.astype('float').max() != xint2.max()
+
 
 def test_preprocess_input_symbolic():
     # Test image batch