[proto] Fixed bug in ScaleJitter with params (#6541)

* [proto] Fixed bug in ScaleJitter with params

* Updated tests

Author: vfdev-5

test/test_prototype_transforms.py

@@ -1263,18 +1263,24 @@ def test__get_params(self, mocker):
         scale_range = (0.5, 1.5)
 
         transform = transforms.ScaleJitter(target_size=target_size, scale_range=scale_range)
-
         sample = mocker.MagicMock(spec=features.Image, num_channels=3, image_size=image_size)
-        params = transform._get_params(sample)
 
-        assert "size" in params
-        size = params["size"]
+        n_samples = 5
+        for _ in range(n_samples):
 
-        assert isinstance(size, tuple) and len(size) == 2
-        height, width = size
+            params = transform._get_params(sample)
+
+            assert "size" in params
+            size = params["size"]
+
+            assert isinstance(size, tuple) and len(size) == 2
+            height, width = size
+
+            r_min = min(target_size[1] / image_size[0], target_size[0] / image_size[1]) * scale_range[0]
+            r_max = min(target_size[1] / image_size[0], target_size[0] / image_size[1]) * scale_range[1]
 
-        assert int(target_size[0] * scale_range[0]) <= height <= int(target_size[0] * scale_range[1])
-        assert int(target_size[1] * scale_range[0]) <= width <= int(target_size[1] * scale_range[1])
+            assert int(image_size[0] * r_min) <= height <= int(image_size[0] * r_max)
+            assert int(image_size[1] * r_min) <= width <= int(image_size[1] * r_max)
 
     def test__transform(self, mocker):
         interpolation_sentinel = mocker.MagicMock()

torchvision/prototype/transforms/_geometry.py

@@ -727,9 +727,10 @@ def __init__(
     def _get_params(self, sample: Any) -> Dict[str, Any]:
         _, orig_height, orig_width = query_chw(sample)
 
-        r = self.scale_range[0] + torch.rand(1) * (self.scale_range[1] - self.scale_range[0])
-        new_width = int(self.target_size[1] * r)
-        new_height = int(self.target_size[0] * r)
+        scale = self.scale_range[0] + torch.rand(1) * (self.scale_range[1] - self.scale_range[0])
+        r = min(self.target_size[1] / orig_height, self.target_size[0] / orig_width) * scale
+        new_width = int(orig_width * r)
+        new_height = int(orig_height * r)
 
         return dict(size=(new_height, new_width))