[transforms] small random resize improvement

doctr/transforms/modules/pytorch.py

@@ -220,18 +220,30 @@ class RandomResize(torch.nn.Module):
 
     >>> import torch
     >>> from doctr.transforms import RandomResize
-    >>> transfo = RandomResize((0.3, 0.9), p=0.5)
+    >>> transfo = RandomResize((0.3, 0.9), preserve_aspect_ratio=True, symmetric_pad=True, p=0.5)
     >>> out = transfo(torch.rand((3, 64, 64)))
 
     Args:
     ----
         scale_range: range of the resizing factor for width and height (independently)
+        preserve_aspect_ratio: whether to preserve the aspect ratio of the image,
+            given a float value, the aspect ratio will be preserved with this probability
+        symmetric_pad: whether to symmetrically pad the image,
+            given a float value, the symmetric padding will be applied with this probability
         p: probability to apply the transformation
     """
 
-    def __init__(self, scale_range: Tuple[float, float] = (0.3, 0.9), p: float = 0.5) -> None:
+    def __init__(
+        self,
+        scale_range: Tuple[float, float] = (0.3, 0.9),
+        preserve_aspect_ratio: Union[bool, float] = False,
+        symmetric_pad: Union[bool, float] = False,
+        p: float = 0.5,
+    ) -> None:
         super().__init__()
         self.scale_range = scale_range
+        self.preserve_aspect_ratio = preserve_aspect_ratio
+        self.symmetric_pad = symmetric_pad
         self.p = p
         self._resize = Resize
 
@@ -241,10 +253,18 @@ def forward(self, img: torch.Tensor, target: np.ndarray) -> Tuple[torch.Tensor,
             scale_w = np.random.uniform(*self.scale_range)
             new_size = (int(img.shape[-2] * scale_h), int(img.shape[-1] * scale_w))
 
-            _img, _target = self._resize(new_size, preserve_aspect_ratio=True, symmetric_pad=True)(img, target)
+            _img, _target = self._resize(
+                new_size,
+                preserve_aspect_ratio=self.preserve_aspect_ratio
+                if isinstance(self.preserve_aspect_ratio, bool)
+                else bool(torch.rand(1) <= self.symmetric_pad),
+                symmetric_pad=self.symmetric_pad
+                if isinstance(self.symmetric_pad, bool)
+                else bool(torch.rand(1) <= self.symmetric_pad),
+            )(img, target)
 
             return _img, _target
         return img, target
 
     def extra_repr(self) -> str:
-        return f"scale_range={self.scale_range}, p={self.p}"
+        return f"scale_range={self.scale_range}, preserve_aspect_ratio={self.preserve_aspect_ratio}, symmetric_pad={self.symmetric_pad}, p={self.p}"  # noqa: E501

doctr/transforms/modules/tensorflow.py

@@ -523,18 +523,30 @@ class RandomResize(NestedObject):
 
     >>> import tensorflow as tf
     >>> from doctr.transforms import RandomResize
-    >>> transfo = RandomResize((0.3, 0.9), p=0.5)
+    >>> transfo = RandomResize((0.3, 0.9), preserve_aspect_ratio=True, symmetric_pad=True, p=0.5)
     >>> out = transfo(tf.random.uniform(shape=[64, 64, 3], minval=0, maxval=1))
 
     Args:
     ----
         scale_range: range of the resizing factor for width and height (independently)
+        preserve_aspect_ratio: whether to preserve the aspect ratio of the image,
+            given a float value, the aspect ratio will be preserved with this probability
+        symmetric_pad: whether to symmetrically pad the image,
+            given a float value, the symmetric padding will be applied with this probability
         p: probability to apply the transformation
     """
 
-    def __init__(self, scale_range: Tuple[float, float] = (0.3, 0.9), p: float = 0.5) -> None:
+    def __init__(
+        self,
+        scale_range: Tuple[float, float] = (0.3, 0.9),
+        preserve_aspect_ratio: Union[bool, float] = False,
+        symmetric_pad: Union[bool, float] = False,
+        p: float = 0.5,
+    ):
         super().__init__()
         self.scale_range = scale_range
+        self.preserve_aspect_ratio = preserve_aspect_ratio
+        self.symmetric_pad = symmetric_pad
         self.p = p
         self._resize = Resize
 
@@ -544,10 +556,18 @@ def __call__(self, img: tf.Tensor, target: np.ndarray) -> Tuple[tf.Tensor, np.nd
             scale_w = random.uniform(*self.scale_range)
             new_size = (int(img.shape[-3] * scale_h), int(img.shape[-2] * scale_w))
 
-            _img, _target = self._resize(new_size, preserve_aspect_ratio=True, symmetric_pad=True)(img, target)
+            _img, _target = self._resize(
+                new_size,
+                preserve_aspect_ratio=self.preserve_aspect_ratio
+                if isinstance(self.preserve_aspect_ratio, bool)
+                else bool(np.random.rand(1) <= self.symmetric_pad),
+                symmetric_pad=self.symmetric_pad
+                if isinstance(self.symmetric_pad, bool)
+                else bool(np.random.rand(1) <= self.symmetric_pad),
+            )(img, target)
 
             return _img, _target
         return img, target
 
     def extra_repr(self) -> str:
-        return f"scale_range={self.scale_range}, p={self.p}"
+        return f"scale_range={self.scale_range}, preserve_aspect_ratio={self.preserve_aspect_ratio}, symmetric_pad={self.symmetric_pad}, p={self.p}"  # noqa: E501

references/detection/train_pytorch.py

@@ -279,7 +279,7 @@ def main(args):
                 T.RandomHorizontalFlip(0.15),
                 T.OneOf([
                     T.RandomApply(T.RandomCrop(ratio=(0.6, 1.33)), 0.25),
-                    T.RandomResize(scale_range=(0.4, 0.9), p=0.25),
+                    T.RandomResize(scale_range=(0.4, 0.9), preserve_aspect_ratio=0.5, symmetric_pad=0.5, p=0.25),
                 ]),
                 T.Resize((args.input_size, args.input_size), preserve_aspect_ratio=True, symmetric_pad=True),
             ]
@@ -288,7 +288,7 @@ def main(args):
                 T.RandomHorizontalFlip(0.15),
                 T.OneOf([
                     T.RandomApply(T.RandomCrop(ratio=(0.6, 1.33)), 0.25),
-                    T.RandomResize(scale_range=(0.4, 0.9), p=0.25),
+                    T.RandomResize(scale_range=(0.4, 0.9), preserve_aspect_ratio=0.5, symmetric_pad=0.5, p=0.25),
                 ]),
                 # Rotation augmentation
                 T.Resize(args.input_size, preserve_aspect_ratio=True),

references/detection/train_tensorflow.py

@@ -237,7 +237,7 @@ def main(args):
                 T.RandomHorizontalFlip(0.15),
                 T.OneOf([
                     T.RandomApply(T.RandomCrop(ratio=(0.6, 1.33)), 0.25),
-                    T.RandomResize(scale_range=(0.4, 0.9), p=0.25),
+                    T.RandomResize(scale_range=(0.4, 0.9), preserve_aspect_ratio=0.5, symmetric_pad=0.5, p=0.25),
                 ]),
                 T.Resize((args.input_size, args.input_size), preserve_aspect_ratio=True, symmetric_pad=True),
             ]
@@ -246,7 +246,7 @@ def main(args):
                 T.RandomHorizontalFlip(0.15),
                 T.OneOf([
                     T.RandomApply(T.RandomCrop(ratio=(0.6, 1.33)), 0.25),
-                    T.RandomResize(scale_range=(0.4, 0.9), p=0.25),
+                    T.RandomResize(scale_range=(0.4, 0.9), preserve_aspect_ratio=0.5, symmetric_pad=0.5, p=0.25),
                 ]),
                 # Rotation augmentation
                 T.Resize(args.input_size, preserve_aspect_ratio=True),

tests/pytorch/test_transforms_pt.py

@@ -204,7 +204,7 @@ def test_random_crop(target):
     # Check the scale
     assert img.shape[-1] * img.shape[-2] >= 0.4 * input_t.shape[-1] * input_t.shape[-2]
     # Check aspect ratio
-    assert 0.65 <= img.shape[-2] / img.shape[-1] <= 1.5
+    assert 0.65 <= img.shape[-2] / img.shape[-1] <= 1.6
     # Check the target
     assert np.all(target >= 0)
     if target.ndim == 2:
@@ -330,17 +330,24 @@ def test_random_shadow(input_dtype, input_shape):
 
 
 @pytest.mark.parametrize(
-    "p,target",
+    "p,preserve_aspect_ratio,symmetric_pad,target",
     [
-        [1, np.array([[0.1, 0.1, 0.3, 0.4]], dtype=np.float32)],
-        [0, np.array([[0.1, 0.1, 0.3, 0.4]], dtype=np.float32)],
-        [1, np.array([[[0.1, 0.8], [0.3, 0.1], [0.3, 0.4], [0.8, 0.4]]], dtype=np.float32)],
-        [0, np.array([[[0.1, 0.8], [0.3, 0.1], [0.3, 0.4], [0.8, 0.4]]], dtype=np.float32)],
+        [1, True, False, np.array([[0.1, 0.1, 0.3, 0.4]], dtype=np.float32)],
+        [0, True, False, np.array([[0.1, 0.1, 0.3, 0.4]], dtype=np.float32)],
+        [1, True, False, np.array([[[0.1, 0.8], [0.3, 0.1], [0.3, 0.4], [0.8, 0.4]]], dtype=np.float32)],
+        [0, True, False, np.array([[[0.1, 0.8], [0.3, 0.1], [0.3, 0.4], [0.8, 0.4]]], dtype=np.float32)],
+        [1, 0.5, 0.5, np.array([[0.1, 0.1, 0.3, 0.4]], dtype=np.float32)],
+        [0, 0.5, 0.5, np.array([[0.1, 0.1, 0.3, 0.4]], dtype=np.float32)],
     ],
 )
-def test_random_resize(p, target):
-    transfo = RandomResize(scale_range=(0.3, 1.3), p=p)
-    assert repr(transfo) == f"RandomResize(scale_range=(0.3, 1.3), p={p})"
+def test_random_resize(p, preserve_aspect_ratio, symmetric_pad, target):
+    transfo = RandomResize(
+        scale_range=(0.3, 1.3), preserve_aspect_ratio=preserve_aspect_ratio, symmetric_pad=symmetric_pad, p=p
+    )
+    assert (
+        repr(transfo)
+        == f"RandomResize(scale_range=(0.3, 1.3), preserve_aspect_ratio={preserve_aspect_ratio}, symmetric_pad={symmetric_pad}, p={p})"  # noqa: E501
+    )
 
     img = torch.rand((3, 64, 64))
     # Apply the transformation

tests/tensorflow/test_transforms_tf.py

@@ -332,7 +332,7 @@ def test_random_crop(target):
     # Check the scale (take a margin)
     assert img.shape[0] * img.shape[1] >= 0.4 * input_t.shape[0] * input_t.shape[1]
     # Check aspect ratio (take a margin)
-    assert 0.65 <= img.shape[0] / img.shape[1] <= 1.5
+    assert 0.65 <= img.shape[0] / img.shape[1] <= 1.6
     # Check the target
     assert np.all(target >= 0)
     if target.ndim == 2:
@@ -471,17 +471,24 @@ def test_random_shadow(input_dtype, input_shape):
 
 
 @pytest.mark.parametrize(
-    "p,target",
+    "p,preserve_aspect_ratio,symmetric_pad,target",
     [
-        [1, np.array([[0.1, 0.1, 0.3, 0.4]], dtype=np.float32)],
-        [0, np.array([[0.1, 0.1, 0.3, 0.4]], dtype=np.float32)],
-        [1, np.array([[[0.1, 0.8], [0.3, 0.1], [0.3, 0.4], [0.8, 0.4]]], dtype=np.float32)],
-        [0, np.array([[[0.1, 0.8], [0.3, 0.1], [0.3, 0.4], [0.8, 0.4]]], dtype=np.float32)],
+        [1, True, False, np.array([[0.1, 0.1, 0.3, 0.4]], dtype=np.float32)],
+        [0, True, False, np.array([[0.1, 0.1, 0.3, 0.4]], dtype=np.float32)],
+        [1, True, False, np.array([[[0.1, 0.8], [0.3, 0.1], [0.3, 0.4], [0.8, 0.4]]], dtype=np.float32)],
+        [0, True, False, np.array([[[0.1, 0.8], [0.3, 0.1], [0.3, 0.4], [0.8, 0.4]]], dtype=np.float32)],
+        [1, 0.5, 0.5, np.array([[0.1, 0.1, 0.3, 0.4]], dtype=np.float32)],
+        [0, 0.5, 0.5, np.array([[0.1, 0.1, 0.3, 0.4]], dtype=np.float32)],
     ],
 )
-def test_random_resize(p, target):
-    transfo = T.RandomResize(scale_range=(0.3, 1.3), p=p)
-    assert repr(transfo) == f"RandomResize(scale_range=(0.3, 1.3), p={p})"
+def test_random_resize(p, preserve_aspect_ratio, symmetric_pad, target):
+    transfo = T.RandomResize(
+        scale_range=(0.3, 1.3), preserve_aspect_ratio=preserve_aspect_ratio, symmetric_pad=symmetric_pad, p=p
+    )
+    assert (
+        repr(transfo)
+        == f"RandomResize(scale_range=(0.3, 1.3), preserve_aspect_ratio={preserve_aspect_ratio}, symmetric_pad={symmetric_pad}, p={p})"  # noqa: E501
+    )
 
     img = tf.random.uniform((64, 64, 3))
     # Apply the transformation