2022-11-10 nightly release (ffd5a56)

test/test_prototype_transforms_consistency.py

@@ -251,6 +251,8 @@ def __init__(
             ArgsKwargs(p=0),
             ArgsKwargs(p=1),
         ],
+        # Use default tolerances of `torch.testing.assert_close`
+        closeness_kwargs=dict(rtol=None, atol=None),
     ),
     ConsistencyConfig(
         prototype_transforms.RandomAdjustSharpness,

torchvision/prototype/transforms/functional/_color.py

@@ -79,9 +79,14 @@ def adjust_contrast_image_tensor(image: torch.Tensor, contrast_factor: float) ->
     c = image.shape[-3]
     if c not in [1, 3]:
         raise TypeError(f"Input image tensor permitted channel values are {[1, 3]}, but found {c}")
-    dtype = image.dtype if torch.is_floating_point(image) else torch.float32
-    grayscale_image = _rgb_to_gray(image) if c == 3 else image
-    mean = torch.mean(grayscale_image.to(dtype), dim=(-3, -2, -1), keepdim=True)
+    fp = image.is_floating_point()
+    if c == 3:
+        grayscale_image = _rgb_to_gray(image, cast=False)
+        if not fp:
+            grayscale_image = grayscale_image.floor_()
+    else:
+        grayscale_image = image if fp else image.to(torch.float32)
+    mean = torch.mean(grayscale_image, dim=(-3, -2, -1), keepdim=True)
     return _blend(image, mean, contrast_factor)
 
 
@@ -372,17 +377,23 @@ def autocontrast_image_tensor(image: torch.Tensor) -> torch.Tensor:
         return image
 
     bound = _FT._max_value(image.dtype)
-    dtype = image.dtype if torch.is_floating_point(image) else torch.float32
+    fp = image.is_floating_point()
+    float_image = image if fp else image.to(torch.float32)
 
-    minimum = image.amin(dim=(-2, -1), keepdim=True).to(dtype)
-    maximum = image.amax(dim=(-2, -1), keepdim=True).to(dtype)
+    minimum = float_image.amin(dim=(-2, -1), keepdim=True)
+    maximum = float_image.amax(dim=(-2, -1), keepdim=True)
 
-    scale = bound / (maximum - minimum)
     eq_idxs = maximum == minimum
+    inv_scale = maximum.sub_(minimum).div_(bound)
     minimum[eq_idxs] = 0.0
-    scale[eq_idxs] = 1.0
+    inv_scale[eq_idxs] = 1.0
+
+    if fp:
+        diff = float_image.sub(minimum)
+    else:
+        diff = float_image.sub_(minimum)
 
-    return (image - minimum).mul_(scale).clamp_(0, bound).to(image.dtype)
+    return diff.div_(inv_scale).clamp_(0, bound).to(image.dtype)
 
 
 autocontrast_image_pil = _FP.autocontrast

torchvision/prototype/transforms/functional/_meta.py

@@ -213,10 +213,12 @@ def _gray_to_rgb(grayscale: torch.Tensor) -> torch.Tensor:
     return grayscale.repeat(repeats)
 
 
-def _rgb_to_gray(image: torch.Tensor) -> torch.Tensor:
+def _rgb_to_gray(image: torch.Tensor, cast: bool = True) -> torch.Tensor:
     r, g, b = image.unbind(dim=-3)
-    l_img = (0.2989 * r).add_(g, alpha=0.587).add_(b, alpha=0.114)
-    l_img = l_img.to(image.dtype).unsqueeze(dim=-3)
+    l_img = r.mul(0.2989).add_(g, alpha=0.587).add_(b, alpha=0.114)
+    if cast:
+        l_img = l_img.to(image.dtype)
+    l_img = l_img.unsqueeze(dim=-3)
     return l_img