Added antialias option to transforms.functional.resize (#3761)

* WIP Added antialias option to transforms.functional.resize

* Updates according to the review

* Excluded these C++ files for iOS build

* Added support for mixed downsampling/upsampling

* Fixed heap overflow caused by explicit loop unrolling

* Applied PR review suggestions
- used pytest parametrize instead unittest
- cast to scalar_t ptr
- removed interpolate aa files for ios/android keeping original cmake version

Author: vfdev-5

android/ops/CMakeLists.txt

@@ -14,6 +14,13 @@ file(GLOB VISION_SRCS
     ../../torchvision/csrc/ops/*.h
     ../../torchvision/csrc/ops/*.cpp)
 
+# Remove interpolate_aa sources as they are temporary code
+# see https://github.com/pytorch/vision/pull/3761
+# and IndexingUtils.h is unavailable on Android build
+list(REMOVE_ITEM VISION_SRCS "${CMAKE_CURRENT_LIST_DIR}/../../torchvision/csrc/ops/cpu/interpolate_aa_kernels.cpp")
+list(REMOVE_ITEM VISION_SRCS "${CMAKE_CURRENT_LIST_DIR}/../../torchvision/csrc/ops/interpolate_aa.cpp")
+list(REMOVE_ITEM VISION_SRCS "${CMAKE_CURRENT_LIST_DIR}/../../torchvision/csrc/ops/interpolate_aa.h")
+
 add_library(${TARGET} SHARED
   ${VISION_SRCS}
 )

ios/CMakeLists.txt

@@ -11,6 +11,13 @@ file(GLOB VISION_SRCS
     ../torchvision/csrc/ops/*.h
     ../torchvision/csrc/ops/*.cpp)
 
+# Remove interpolate_aa sources as they are temporary code
+# see https://github.com/pytorch/vision/pull/3761
+# and using TensorIterator unavailable with iOS
+list(REMOVE_ITEM VISION_SRCS "${CMAKE_CURRENT_LIST_DIR}/../torchvision/csrc/ops/cpu/interpolate_aa_kernels.cpp")
+list(REMOVE_ITEM VISION_SRCS "${CMAKE_CURRENT_LIST_DIR}/../torchvision/csrc/ops/interpolate_aa.cpp")
+list(REMOVE_ITEM VISION_SRCS "${CMAKE_CURRENT_LIST_DIR}/../torchvision/csrc/ops/interpolate_aa.h")
+
 add_library(${TARGET} STATIC
   ${VISION_SRCS}
 )

test/test_functional_tensor.py

@@ -1018,5 +1018,52 @@ def test_perspective_interpolation_warning(tester):
         tester.assertTrue(res1.equal(res2))
 
 
+@pytest.mark.parametrize('device', ["cpu", ])
+@pytest.mark.parametrize('dt', [None, torch.float32, torch.float64, torch.float16])
+@pytest.mark.parametrize('size', [[96, 72], [96, 420], [420, 72]])
+@pytest.mark.parametrize('interpolation', [BILINEAR, ])
+def test_resize_antialias(device, dt, size, interpolation, tester):
+
+    if dt == torch.float16 and device == "cpu":
+        # skip float16 on CPU case
+        return
+
+    script_fn = torch.jit.script(F.resize)
+    tensor, pil_img = tester._create_data(320, 290, device=device)
+
+    if dt is not None:
+        # This is a trivial cast to float of uint8 data to test all cases
+        tensor = tensor.to(dt)
+
+    resized_tensor = F.resize(tensor, size=size, interpolation=interpolation, antialias=True)
+    resized_pil_img = F.resize(pil_img, size=size, interpolation=interpolation)
+
+    tester.assertEqual(
+        resized_tensor.size()[1:], resized_pil_img.size[::-1],
+        msg=f"{size}, {interpolation}, {dt}"
+    )
+
+    resized_tensor_f = resized_tensor
+    # we need to cast to uint8 to compare with PIL image
+    if resized_tensor_f.dtype == torch.uint8:
+        resized_tensor_f = resized_tensor_f.to(torch.float)
+
+    tester.approxEqualTensorToPIL(
+        resized_tensor_f, resized_pil_img, tol=0.5, msg=f"{size}, {interpolation}, {dt}"
+    )
+    tester.approxEqualTensorToPIL(
+        resized_tensor_f, resized_pil_img, tol=1.0 + 1e-5, agg_method="max",
+        msg=f"{size}, {interpolation}, {dt}"
+    )
+
+    if isinstance(size, int):
+        script_size = [size, ]
+    else:
+        script_size = size
+
+    resize_result = script_fn(tensor, size=script_size, interpolation=interpolation, antialias=True)
+    tester.assertTrue(resized_tensor.equal(resize_result), msg=f"{size}, {interpolation}, {dt}")
+
+
 if __name__ == '__main__':
     unittest.main()

test/test_transforms.py

@@ -348,6 +348,10 @@ def test_resize(self):
 
                 self.assertEqual((owidth, oheight), result.size)
 
+        with self.assertWarnsRegex(UserWarning, r"Anti-alias option is always applied for PIL Image input"):
+            t = transforms.Resize(osize, antialias=False)
+            t(img)
+
     def test_random_crop(self):
         height = random.randint(10, 32) * 2
         width = random.randint(10, 32) * 2