[fbsync] Transforms without dispatcher (#5421)

Summary:
* add prototype transforms that don't need dispatchers

* cleanup

* remove legacy_transform decorator

* remove legacy classes

* remove explicit param passing

* streamline extra_repr

* remove obsolete ._supports() method

* cleanup

* remove Query

* cleanup

* fix tests

* kernels -> functional

* move image size and num channels extraction to functional

* extend legacy function to extract image size and num channels

* implement dispatching for auto augment

* fix auto augment dispatch

* revert some naming changes

* remove ability to pass params to autoaugment

* fix legacy image size extraction

* align prototype.transforms.functional with transforms.functional

* cleanup

* fix image size and channels extraction

* fix affine and rotate

* revert image size to (width, height)

* Minor corrections

Reviewed By: datumbox

Differential Revision: D34579512

fbshipit-source-id: 2044269d771b3488010b62ff611cf4f75ef75ed4

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

Author: 2 people

test/test_prototype_transforms.py

@@ -1,9 +1,8 @@
 import itertools
 
-import PIL.Image
 import pytest
 import torch
-from test_prototype_transforms_kernels import make_images, make_bounding_boxes, make_one_hot_labels
+from test_prototype_transforms_functional import make_images, make_bounding_boxes, make_one_hot_labels
 from torchvision.prototype import transforms, features
 from torchvision.transforms.functional import to_pil_image
 
@@ -25,15 +24,6 @@ def make_vanilla_tensor_bounding_boxes(*args, **kwargs):
         yield bounding_box.data
 
 
-INPUT_CREATIONS_FNS = {
-    features.Image: make_images,
-    features.BoundingBox: make_bounding_boxes,
-    features.OneHotLabel: make_one_hot_labels,
-    torch.Tensor: make_vanilla_tensor_images,
-    PIL.Image.Image: make_pil_images,
-}
-
-
 def parametrize(transforms_with_inputs):
     return pytest.mark.parametrize(
         ("transform", "input"),
@@ -52,15 +42,21 @@ def parametrize(transforms_with_inputs):
 def parametrize_from_transforms(*transforms):
     transforms_with_inputs = []
     for transform in transforms:
-        dispatcher = transform._DISPATCHER
-        if dispatcher is None:
-            continue
-
-        for type_ in dispatcher._kernels:
+        for creation_fn in [
+            make_images,
+            make_bounding_boxes,
+            make_one_hot_labels,
+            make_vanilla_tensor_images,
+            make_pil_images,
+        ]:
+            inputs = list(creation_fn())
             try:
-                inputs = INPUT_CREATIONS_FNS[type_]()
-            except KeyError:
+                output = transform(inputs[0])
+            except Exception:
                 continue
+            else:
+                if output is inputs[0]:
+                    continue
 
             transforms_with_inputs.append((transform, inputs))
 
@@ -69,7 +65,7 @@ def parametrize_from_transforms(*transforms):
 
 class TestSmoke:
     @parametrize_from_transforms(
-        transforms.RandomErasing(),
+        transforms.RandomErasing(p=1.0),
         transforms.HorizontalFlip(),
         transforms.Resize([16, 16]),
         transforms.CenterCrop([16, 16]),
@@ -141,35 +137,6 @@ def test_auto_augment(self, transform, input):
     def test_normalize(self, transform, input):
         transform(input)
 
-    @parametrize(
-        [
-            (
-                transforms.ConvertColorSpace("grayscale"),
-                itertools.chain(
-                    make_images(),
-                    make_vanilla_tensor_images(color_spaces=["rgb"]),
-                    make_pil_images(color_spaces=["rgb"]),
-                ),
-            )
-        ]
-    )
-    def test_convert_bounding_color_space(self, transform, input):
-        transform(input)
-
-    @parametrize(
-        [
-            (
-                transforms.ConvertBoundingBoxFormat("xyxy", old_format="xywh"),
-                itertools.chain(
-                    make_bounding_boxes(),
-                    make_vanilla_tensor_bounding_boxes(formats=["xywh"]),
-                ),
-            )
-        ]
-    )
-    def test_convert_bounding_box_format(self, transform, input):
-        transform(input)
-
     @parametrize(
         [
             (

test/test_prototype_transforms_kernels.py renamed to test/test_prototype_transforms_functional.py

@@ -3,7 +3,7 @@
 
 import pytest
 import torch.testing
-import torchvision.prototype.transforms.kernels as K
+import torchvision.prototype.transforms.functional as F
 from torch import jit
 from torch.nn.functional import one_hot
 from torchvision.prototype import features
@@ -134,10 +134,10 @@ def __init__(self, *args, **kwargs):
         self.kwargs = kwargs
 
 
-class KernelInfo:
+class FunctionalInfo:
     def __init__(self, name, *, sample_inputs_fn):
         self.name = name
-        self.kernel = getattr(K, name)
+        self.functional = getattr(F, name)
         self._sample_inputs_fn = sample_inputs_fn
 
     def sample_inputs(self):
@@ -146,21 +146,21 @@ def sample_inputs(self):
     def __call__(self, *args, **kwargs):
         if len(args) == 1 and not kwargs and isinstance(args[0], SampleInput):
             sample_input = args[0]
-            return self.kernel(*sample_input.args, **sample_input.kwargs)
+            return self.functional(*sample_input.args, **sample_input.kwargs)
 
-        return self.kernel(*args, **kwargs)
+        return self.functional(*args, **kwargs)
 
 
-KERNEL_INFOS = []
+FUNCTIONAL_INFOS = []
 
 
 def register_kernel_info_from_sample_inputs_fn(sample_inputs_fn):
-    KERNEL_INFOS.append(KernelInfo(sample_inputs_fn.__name__, sample_inputs_fn=sample_inputs_fn))
+    FUNCTIONAL_INFOS.append(FunctionalInfo(sample_inputs_fn.__name__, sample_inputs_fn=sample_inputs_fn))
     return sample_inputs_fn
 
 
 @register_kernel_info_from_sample_inputs_fn
-def horizontal_flip_image():
+def horizontal_flip_image_tensor():
     for image in make_images():
         yield SampleInput(image)
 
@@ -172,12 +172,12 @@ def horizontal_flip_bounding_box():
 
 
 @register_kernel_info_from_sample_inputs_fn
-def resize_image():
+def resize_image_tensor():
     for image, interpolation in itertools.product(
         make_images(),
         [
-            K.InterpolationMode.BILINEAR,
-            K.InterpolationMode.NEAREST,
+            F.InterpolationMode.BILINEAR,
+            F.InterpolationMode.NEAREST,
         ],
     ):
         height, width = image.shape[-2:]
@@ -200,20 +200,20 @@ def resize_bounding_box():
 
 
 class TestKernelsCommon:
-    @pytest.mark.parametrize("kernel_info", KERNEL_INFOS, ids=lambda kernel_info: kernel_info.name)
-    def test_scriptable(self, kernel_info):
-        jit.script(kernel_info.kernel)
+    @pytest.mark.parametrize("functional_info", FUNCTIONAL_INFOS, ids=lambda functional_info: functional_info.name)
+    def test_scriptable(self, functional_info):
+        jit.script(functional_info.functional)
 
     @pytest.mark.parametrize(
-        ("kernel_info", "sample_input"),
+        ("functional_info", "sample_input"),
         [
-            pytest.param(kernel_info, sample_input, id=f"{kernel_info.name}-{idx}")
-            for kernel_info in KERNEL_INFOS
-            for idx, sample_input in enumerate(kernel_info.sample_inputs())
+            pytest.param(functional_info, sample_input, id=f"{functional_info.name}-{idx}")
+            for functional_info in FUNCTIONAL_INFOS
+            for idx, sample_input in enumerate(functional_info.sample_inputs())
         ],
     )
-    def test_eager_vs_scripted(self, kernel_info, sample_input):
-        eager = kernel_info(sample_input)
-        scripted = jit.script(kernel_info.kernel)(*sample_input.args, **sample_input.kwargs)
+    def test_eager_vs_scripted(self, functional_info, sample_input):
+        eager = functional_info(sample_input)
+        scripted = jit.script(functional_info.functional)(*sample_input.args, **sample_input.kwargs)
 
         torch.testing.assert_close(eager, scripted)

torchvision/prototype/features/_bounding_box.py

@@ -41,7 +41,7 @@ def to_format(self, format: Union[str, BoundingBoxFormat]) -> BoundingBox:
         #  promote this out of the prototype state
 
         # import at runtime to avoid cyclic imports
-        from torchvision.prototype.transforms.kernels import convert_bounding_box_format
+        from torchvision.prototype.transforms.functional import convert_bounding_box_format
 
         if isinstance(format, str):
             format = BoundingBoxFormat[format]

torchvision/prototype/features/_encoded.py

@@ -43,7 +43,7 @@ def decode(self) -> Image:
         #  promote this out of the prototype state
 
         # import at runtime to avoid cyclic imports
-        from torchvision.prototype.transforms.kernels import decode_image_with_pil
+        from torchvision.prototype.transforms.functional import decode_image_with_pil
 
         return Image(decode_image_with_pil(self))
 

torchvision/prototype/transforms/__init__.py

@@ -1,13 +1,14 @@
-from torchvision.transforms import AutoAugmentPolicy, InterpolationMode  # usort: skip
-from . import kernels  # usort: skip
+from torchvision.transforms import InterpolationMode, AutoAugmentPolicy  # usort: skip
+
 from . import functional  # usort: skip
+
 from ._transform import Transform  # usort: skip
 
 from ._augment import RandomErasing, RandomMixup, RandomCutmix
 from ._auto_augment import RandAugment, TrivialAugmentWide, AutoAugment
 from ._container import Compose, RandomApply, RandomChoice, RandomOrder
 from ._geometry import HorizontalFlip, Resize, CenterCrop, RandomResizedCrop
-from ._meta_conversion import ConvertBoundingBoxFormat, ConvertImageDtype, ConvertColorSpace
+from ._meta_conversion import ConvertBoundingBoxFormat, ConvertImageDtype, ConvertImageColorSpace
 from ._misc import Identity, Normalize, ToDtype, Lambda
 from ._presets import CocoEval, ImageNetEval, VocEval, Kinect400Eval, RaftEval
 from ._type_conversion import DecodeImage, LabelToOneHot

torchvision/prototype/transforms/_augment.py

@@ -3,7 +3,6 @@
 import warnings
 from typing import Any, Dict, Tuple
 
-import PIL.Image
 import torch
 from torchvision.prototype import features
 from torchvision.prototype.transforms import Transform, functional as F
@@ -12,9 +11,6 @@
 
 
 class RandomErasing(Transform):
-    _DISPATCHER = F.erase
-    _FAIL_TYPES = {PIL.Image.Image, features.BoundingBox, features.SegmentationMask}
-
     def __init__(
         self,
         p: float = 0.5,
@@ -45,8 +41,8 @@ def __init__(
 
     def _get_params(self, sample: Any) -> Dict[str, Any]:
         image = query_image(sample)
-        img_h, img_w = F.get_image_size(image)
         img_c = F.get_image_num_channels(image)
+        img_w, img_h = F.get_image_size(image)
 
         if isinstance(self.value, (int, float)):
             value = [self.value]
@@ -93,16 +89,24 @@ def _get_params(self, sample: Any) -> Dict[str, Any]:
         return dict(zip("ijhwv", (i, j, h, w, v)))
 
     def _transform(self, input: Any, params: Dict[str, Any]) -> Any:
-        if torch.rand(1) >= self.p:
+        if isinstance(input, (features.BoundingBox, features.SegmentationMask)):
+            raise TypeError(f"{type(input).__name__}'s are not supported by {type(self).__name__}()")
+        elif isinstance(input, features.Image):
+            output = F.erase_image_tensor(input, **params)
+            return features.Image.new_like(input, output)
+        elif isinstance(input, torch.Tensor):
+            return F.erase_image_tensor(input, **params)
+        else:
             return input
 
-        return super()._transform(input, params)
+    def forward(self, *inputs: Any) -> Any:
+        if torch.rand(1) >= self.p:
+            return inputs if len(inputs) > 1 else inputs[0]
+
+        return super().forward(*inputs)
 
 
 class RandomMixup(Transform):
-    _DISPATCHER = F.mixup
-    _FAIL_TYPES = {features.BoundingBox, features.SegmentationMask}
-
     def __init__(self, *, alpha: float) -> None:
         super().__init__()
         self.alpha = alpha
@@ -111,11 +115,20 @@ def __init__(self, *, alpha: float) -> None:
     def _get_params(self, sample: Any) -> Dict[str, Any]:
         return dict(lam=float(self._dist.sample(())))
 
+    def _transform(self, input: Any, params: Dict[str, Any]) -> Any:
+        if isinstance(input, (features.BoundingBox, features.SegmentationMask)):
+            raise TypeError(f"{type(input).__name__}'s are not supported by {type(self).__name__}()")
+        elif isinstance(input, features.Image):
+            output = F.mixup_image_tensor(input, **params)
+            return features.Image.new_like(input, output)
+        elif isinstance(input, features.OneHotLabel):
+            output = F.mixup_one_hot_label(input, **params)
+            return features.OneHotLabel.new_like(input, output)
+        else:
+            return input
 
-class RandomCutmix(Transform):
-    _DISPATCHER = F.cutmix
-    _FAIL_TYPES = {features.BoundingBox, features.SegmentationMask}
 
+class RandomCutmix(Transform):
     def __init__(self, *, alpha: float) -> None:
         super().__init__()
         self.alpha = alpha
@@ -125,7 +138,7 @@ def _get_params(self, sample: Any) -> Dict[str, Any]:
         lam = float(self._dist.sample(()))
 
         image = query_image(sample)
-        H, W = F.get_image_size(image)
+        W, H = F.get_image_size(image)
 
         r_x = torch.randint(W, ())
         r_y = torch.randint(H, ())
@@ -143,3 +156,15 @@ def _get_params(self, sample: Any) -> Dict[str, Any]:
         lam_adjusted = float(1.0 - (x2 - x1) * (y2 - y1) / (W * H))
 
         return dict(box=box, lam_adjusted=lam_adjusted)
+
+    def _transform(self, input: Any, params: Dict[str, Any]) -> Any:
+        if isinstance(input, (features.BoundingBox, features.SegmentationMask)):
+            raise TypeError(f"{type(input).__name__}'s are not supported by {type(self).__name__}()")
+        elif isinstance(input, features.Image):
+            output = F.cutmix_image_tensor(input, box=params["box"])
+            return features.Image.new_like(input, output)
+        elif isinstance(input, features.OneHotLabel):
+            output = F.cutmix_one_hot_label(input, lam_adjusted=params["lam_adjusted"])
+            return features.OneHotLabel.new_like(input, output)
+        else:
+            return input