[fbsync] add segmentation reference consistency tests (#6591)

Summary:
* add segmentation reference consistency tests

* fall back to smoke tests for resize

* add test for RandomCrop

Reviewed By: YosuaMichael

Differential Revision: D39885419

fbshipit-source-id: 12692442af350ca9b97e8764ed429d2d56e0682a

Co-authored-by: Vasilis Vryniotis <[email protected]>

Author: datumbox

test/prototype_common_utils.py

@@ -69,7 +69,8 @@ def __init__(
 
     def _process_inputs(self, actual, expected, *, id, allow_subclasses):
         actual, expected = [
-            to_image_tensor(input) if not isinstance(input, torch.Tensor) else input for input in [actual, expected]
+            to_image_tensor(input) if not isinstance(input, torch.Tensor) else features.Image(input)
+            for input in [actual, expected]
         ]
         # This broadcast is needed, because `features.Mask`'s can have a 2D shape, but converting the equivalent PIL
         # image to a tensor adds a singleton leading dimension.

test/test_prototype_transforms_consistency.py

@@ -1,5 +1,7 @@
 import enum
 import inspect
+import random
+from collections import defaultdict
 from importlib.machinery import SourceFileLoader
 from pathlib import Path
 
@@ -16,13 +18,15 @@
     make_image,
     make_images,
     make_label,
+    make_segmentation_mask,
 )
 from torchvision import transforms as legacy_transforms
 from torchvision._utils import sequence_to_str
 from torchvision.prototype import features, transforms as prototype_transforms
+from torchvision.prototype.transforms import functional as F
+from torchvision.prototype.transforms._utils import query_chw
 from torchvision.prototype.transforms.functional import to_image_pil
 
-
 DEFAULT_MAKE_IMAGES_KWARGS = dict(color_spaces=[features.ColorSpace.RGB], extra_dims=[(4,)])
 
 
@@ -852,10 +856,12 @@ def test_aa(self, inpt, interpolation):
         assert_equal(expected_output, output)
 
 
-# Import reference detection transforms here for consistency checks
-# torchvision/references/detection/transforms.py
-ref_det_filepath = Path(__file__).parent.parent / "references" / "detection" / "transforms.py"
-det_transforms = SourceFileLoader(ref_det_filepath.stem, ref_det_filepath.as_posix()).load_module()
+def import_transforms_from_references(reference):
+    ref_det_filepath = Path(__file__).parent.parent / "references" / reference / "transforms.py"
+    return SourceFileLoader(ref_det_filepath.stem, ref_det_filepath.as_posix()).load_module()
+
+
+det_transforms = import_transforms_from_references("detection")
 
 
 class TestRefDetTransforms:
@@ -873,7 +879,7 @@ def make_datapoints(self, with_mask=True):
 
         yield (pil_image, target)
 
-        tensor_image = torch.randint(0, 256, size=(3, *size), dtype=torch.uint8)
+        tensor_image = torch.Tensor(make_image(size=size, color_space=features.ColorSpace.RGB))
         target = {
             "boxes": make_bounding_box(image_size=size, format="XYXY", extra_dims=(num_objects,), dtype=torch.float),
             "labels": make_label(extra_dims=(num_objects,), categories=80),
@@ -883,7 +889,7 @@ def make_datapoints(self, with_mask=True):
 
         yield (tensor_image, target)
 
-        feature_image = features.Image(torch.randint(0, 256, size=(3, *size), dtype=torch.uint8))
+        feature_image = make_image(size=size, color_space=features.ColorSpace.RGB)
         target = {
             "boxes": make_bounding_box(image_size=size, format="XYXY", extra_dims=(num_objects,), dtype=torch.float),
             "labels": make_label(extra_dims=(num_objects,), categories=80),
@@ -927,3 +933,165 @@ def test_transform(self, t_ref, t, data_kwargs):
             expected_output = t_ref(*dp)
 
             assert_equal(expected_output, output)
+
+
+seg_transforms = import_transforms_from_references("segmentation")
+
+
+# We need this transform for two reasons:
+# 1. transforms.RandomCrop uses a different scheme to pad images and masks of insufficient size than its name
+#    counterpart in the detection references. Thus, we cannot use it with `pad_if_needed=True`
+# 2. transforms.Pad only supports a fixed padding, but the segmentation datasets don't have a fixed image size.
+class PadIfSmaller(prototype_transforms.Transform):
+    def __init__(self, size, fill=0):
+        super().__init__()
+        self.size = size
+        self.fill = prototype_transforms._geometry._setup_fill_arg(fill)
+
+    def _get_params(self, sample):
+        _, height, width = query_chw(sample)
+        padding = [0, 0, max(self.size - width, 0), max(self.size - height, 0)]
+        needs_padding = any(padding)
+        return dict(padding=padding, needs_padding=needs_padding)
+
+    def _transform(self, inpt, params):
+        if not params["needs_padding"]:
+            return inpt
+
+        fill = self.fill[type(inpt)]
+        fill = F._geometry._convert_fill_arg(fill)
+
+        return F.pad(inpt, padding=params["padding"], fill=fill)
+
+
+class TestRefSegTransforms:
+    def make_datapoints(self, supports_pil=True, image_dtype=torch.uint8):
+        size = (256, 640)
+        num_categories = 21
+
+        conv_fns = []
+        if supports_pil:
+            conv_fns.append(to_image_pil)
+        conv_fns.extend([torch.Tensor, lambda x: x])
+
+        for conv_fn in conv_fns:
+            feature_image = make_image(size=size, color_space=features.ColorSpace.RGB, dtype=image_dtype)
+            feature_mask = make_segmentation_mask(size=size, num_categories=num_categories, dtype=torch.uint8)
+
+            dp = (conv_fn(feature_image), feature_mask)
+            dp_ref = (
+                to_image_pil(feature_image) if supports_pil else torch.Tensor(feature_image),
+                to_image_pil(feature_mask),
+            )
+
+            yield dp, dp_ref
+
+    def set_seed(self, seed=12):
+        torch.manual_seed(seed)
+        random.seed(seed)
+
+    def check(self, t, t_ref, data_kwargs=None):
+        for dp, dp_ref in self.make_datapoints(**data_kwargs or dict()):
+
+            self.set_seed()
+            output = t(dp)
+
+            self.set_seed()
+            expected_output = t_ref(*dp_ref)
+
+            assert_equal(output, expected_output)
+
+    @pytest.mark.parametrize(
+        ("t_ref", "t", "data_kwargs"),
+        [
+            (
+                seg_transforms.RandomHorizontalFlip(flip_prob=1.0),
+                prototype_transforms.RandomHorizontalFlip(p=1.0),
+                dict(),
+            ),
+            (
+                seg_transforms.RandomHorizontalFlip(flip_prob=0.0),
+                prototype_transforms.RandomHorizontalFlip(p=0.0),
+                dict(),
+            ),
+            (
+                seg_transforms.RandomCrop(size=480),
+                prototype_transforms.Compose(
+                    [
+                        PadIfSmaller(size=480, fill=defaultdict(lambda: 0, {features.Mask: 255})),
+                        prototype_transforms.RandomCrop(size=480),
+                    ]
+                ),
+                dict(),
+            ),
+            (
+                seg_transforms.Normalize(mean=(0.485, 0.456, 0.406), std=(0.229, 0.224, 0.225)),
+                prototype_transforms.Normalize(mean=(0.485, 0.456, 0.406), std=(0.229, 0.224, 0.225)),
+                dict(supports_pil=False, image_dtype=torch.float),
+            ),
+        ],
+    )
+    def test_common(self, t_ref, t, data_kwargs):
+        self.check(t, t_ref, data_kwargs)
+
+    def check_resize(self, mocker, t_ref, t):
+        mock = mocker.patch("torchvision.prototype.transforms._geometry.F.resize")
+        mock_ref = mocker.patch("torchvision.transforms.functional.resize")
+
+        for dp, dp_ref in self.make_datapoints():
+            mock.reset_mock()
+            mock_ref.reset_mock()
+
+            self.set_seed()
+            t(dp)
+            assert mock.call_count == 2
+            assert all(
+                actual is expected
+                for actual, expected in zip([call_args[0][0] for call_args in mock.call_args_list], dp)
+            )
+
+            self.set_seed()
+            t_ref(*dp_ref)
+            assert mock_ref.call_count == 2
+            assert all(
+                actual is expected
+                for actual, expected in zip([call_args[0][0] for call_args in mock_ref.call_args_list], dp_ref)
+            )
+
+            for args_kwargs, args_kwargs_ref in zip(mock.call_args_list, mock_ref.call_args_list):
+                assert args_kwargs[0][1] == [args_kwargs_ref[0][1]]
+
+    def test_random_resize_train(self, mocker):
+        base_size = 520
+        min_size = base_size // 2
+        max_size = base_size * 2
+
+        randint = torch.randint
+
+        def patched_randint(a, b, *other_args, **kwargs):
+            if kwargs or len(other_args) > 1 or other_args[0] != ():
+                return randint(a, b, *other_args, **kwargs)
+
+            return random.randint(a, b)
+
+        # We are patching torch.randint -> random.randint here, because we can't patch the modules that are not imported
+        # normally
+        t = prototype_transforms.RandomResize(min_size=min_size, max_size=max_size, antialias=True)
+        mocker.patch(
+            "torchvision.prototype.transforms._geometry.torch.randint",
+            new=patched_randint,
+        )
+
+        t_ref = seg_transforms.RandomResize(min_size=min_size, max_size=max_size)
+
+        self.check_resize(mocker, t_ref, t)
+
+    def test_random_resize_eval(self, mocker):
+        torch.manual_seed(0)
+        base_size = 520
+
+        t = prototype_transforms.Resize(size=base_size, antialias=True)
+
+        t_ref = seg_transforms.RandomResize(min_size=base_size, max_size=base_size)
+
+        self.check_resize(mocker, t_ref, t)