Negative padding for functional_pil #2381 (#2744)

* Negative padding for functional_pil #2381

* Tests for PIL negative padding #2381

* Move PIL vs tensor test inside test_pad

* Adapt test_pad from test_transforms_tensor.py
Co-authored-by: vfdev <vfdev.5@gmail.com>

test/test_transforms.py

@@ -380,6 +380,16 @@ class Tester(unittest.TestCase):
         self.assertTrue(np.all(symmetric_middle_slice == np.asarray([0, 1, 200, 200, 1, 0])))
         self.assertEqual(transforms.ToTensor()(symmetric_padded_img).size(), (3, 32, 34))
 
+        # Check negative padding explicitly for symmetric case, since it is not
+        # implemented for tensor case to compare to
+        # Crop 1 to left, pad 2 to top, pad 3 to right, crop 3 to bottom
+        symmetric_padded_img_neg = F.pad(img, (-1, 2, 3, -3), padding_mode='symmetric')
+        symmetric_neg_middle_left = np.asarray(symmetric_padded_img_neg).transpose(2, 0, 1)[0][17][:3]
+        symmetric_neg_middle_right = np.asarray(symmetric_padded_img_neg).transpose(2, 0, 1)[0][17][-4:]
+        self.assertTrue(np.all(symmetric_neg_middle_left == np.asarray([1, 0, 0])))
+        self.assertTrue(np.all(symmetric_neg_middle_right == np.asarray([200, 200, 0, 0])))
+        self.assertEqual(transforms.ToTensor()(symmetric_padded_img_neg).size(), (3, 28, 31))
+
     def test_pad_raises_with_invalid_pad_sequence_len(self):
         with self.assertRaises(ValueError):
             transforms.Pad(())

test/test_transforms_tensor.py

@@ -121,23 +121,27 @@ class Tester(TransformsTester):
         )
 
     def test_pad(self):
-
+        for m in ["constant", "edge", "reflect", "symmetric"]:
+            fill = 127 if m == "constant" else 0
+            # Negative pad currently unsupported for Tensor and symmetric
+            multipliers = [1] if m == "symmetric" else [1, -1]
+            for mul in multipliers:
                 # Test functional.pad (PIL and Tensor) with padding as single int
                 self._test_functional_op(
-            "pad", fn_kwargs={"padding": 2, "fill": 0, "padding_mode": "constant"}
+                    "pad", fn_kwargs={"padding": mul * 2, "fill": fill, "padding_mode": m}
                 )
                 # Test functional.pad and transforms.Pad with padding as [int, ]
-        fn_kwargs = meth_kwargs = {"padding": [2, ], "fill": 0, "padding_mode": "constant"}
+                fn_kwargs = meth_kwargs = {"padding": [mul * 2, ], "fill": fill, "padding_mode": m}
                 self._test_op(
                     "pad", "Pad", fn_kwargs=fn_kwargs, meth_kwargs=meth_kwargs
                 )
                 # Test functional.pad and transforms.Pad with padding as list
-        fn_kwargs = meth_kwargs = {"padding": [4, 4], "fill": 0, "padding_mode": "constant"}
+                fn_kwargs = meth_kwargs = {"padding": [mul * 4, 4], "fill": fill, "padding_mode": m}
                 self._test_op(
                     "pad", "Pad", fn_kwargs=fn_kwargs, meth_kwargs=meth_kwargs
                 )
                 # Test functional.pad and transforms.Pad with padding as tuple
-        fn_kwargs = meth_kwargs = {"padding": (2, 2, 2, 2), "fill": 127, "padding_mode": "constant"}
+                fn_kwargs = meth_kwargs = {"padding": (mul * 2, 2, 2, mul * 2), "fill": fill, "padding_mode": m}
                 self._test_op(
                     "pad", "Pad", fn_kwargs=fn_kwargs, meth_kwargs=meth_kwargs
                 )

torchvision/transforms/functional_pil.py

@@ -330,6 +330,15 @@ def pad(img, padding, fill=0, padding_mode="constant"):
             pad_right = padding[2]
             pad_bottom = padding[3]
 
+        p = [pad_left, pad_top, pad_right, pad_bottom]
+        cropping = -np.minimum(p, 0)
+
+        if cropping.any():
+            crop_left, crop_top, crop_right, crop_bottom = cropping
+            img = img.crop((crop_left, crop_top, img.width - crop_right, img.height - crop_bottom))
+
+        pad_left, pad_top, pad_right, pad_bottom = np.maximum(p, 0)
+
         if img.mode == 'P':
             palette = img.getpalette()
             img = np.asarray(img)