👽️ API changes in PyTorch 1.11 (#25)

Author: francois-rozet

piqa/ssim.py

@@ -244,7 +244,7 @@ def forward(self, input: Tensor, target: Tensor) -> Tensor:
         assert_type(
             input, target,
             device=self.kernel.device,
-            dim_range=(3, -1),
+            dim_range=(3, 5),
             n_channels=self.kernel.size(0),
             value_range=(0., self.value_range),
         )

piqa/utils/color.py

@@ -29,7 +29,7 @@ def color_conv(
         weight: A weight kernel, :math:`(C', C)`.
     """
 
-    return F.conv1d(x, weight.view(weight.shape + (1,) * spatial(x)))
+    return F.linear(x.transpose(1, -1), weight).transpose(1, -1)
 
 
 RGB_TO_YIQ = torch.tensor([

piqa/utils/functional.py

@@ -36,7 +36,18 @@ def channel_conv(
                   [144., 153., 162.]]]])
     """
 
-    return F.conv1d(x, kernel, padding=padding, groups=x.size(1))
+    D = len(kernel.shape) - 2
+
+    assert D <= 3, "PyTorch only supports 1D, 2D or 3D convolutions."
+
+    if D == 3:
+        return F.conv3d(x, kernel, padding=padding, groups=x.size(-4))
+    elif D == 2:
+        return F.conv2d(x, kernel, padding=padding, groups=x.size(-3))
+    elif D == 1:
+        return F.conv1d(x, kernel, padding=padding, groups=x.size(-2))
+    else:
+        return F.linear(x, kernel.expand(x.size(-1)))
 
 
 def channel_convs(

setup.py

@@ -10,7 +10,7 @@
 
 setuptools.setup(
     name='piqa',
-    version='1.2.1',
+    version='1.2.2',
     packages=setuptools.find_packages(),
     description='PyTorch Image Quality Assessment',
     keywords='image quality processing metrics torch vision',