Fix dtype propagation (#1141)

Summary:
Previously, dtypes were not propagated correctly in composed transforms, resulting in errors when different dtypes were mixed. Even specifying a dtype in the constructor does not fix this. Neither does specifying the dtype for each composition function invocation (e.g. as a `kwarg` in `rotate_axis_angle`).

With the change, I also had to modify the default dtype of `RotateAxisAngle`, which was `torch.float64`; it is now `torch.float32` like for all other transforms. This was required because the fix in propagation broke some tests due to dtype mismatches.

This change in default dtype in turn broke two tests due to precision changes (calculations that were previously done in `torch.float64` were now done in `torch.float32`), so I changed the precision tolerances to be less strict. I chose the lowest power of ten that passed the tests here.

Pull Request resolved: #1141

Reviewed By: patricklabatut

Differential Revision: D35192970

Pulled By: bottler

fbshipit-source-id: ba0293e8b3595dfc94b3cf8048e50b7a5e5ed7cf

Author: janEbert

pytorch3d/transforms/transform3d.py

@@ -390,16 +390,24 @@ def transform_normals(self, normals) -> torch.Tensor:
         return normals_out
 
     def translate(self, *args, **kwargs) -> "Transform3d":
-        return self.compose(Translate(device=self.device, *args, **kwargs))
+        return self.compose(
+            Translate(device=self.device, dtype=self.dtype, *args, **kwargs)
+        )
 
     def scale(self, *args, **kwargs) -> "Transform3d":
-        return self.compose(Scale(device=self.device, *args, **kwargs))
+        return self.compose(
+            Scale(device=self.device, dtype=self.dtype, *args, **kwargs)
+        )
 
     def rotate(self, *args, **kwargs) -> "Transform3d":
-        return self.compose(Rotate(device=self.device, *args, **kwargs))
+        return self.compose(
+            Rotate(device=self.device, dtype=self.dtype, *args, **kwargs)
+        )
 
     def rotate_axis_angle(self, *args, **kwargs) -> "Transform3d":
-        return self.compose(RotateAxisAngle(device=self.device, *args, **kwargs))
+        return self.compose(
+            RotateAxisAngle(device=self.device, dtype=self.dtype, *args, **kwargs)
+        )
 
     def clone(self) -> "Transform3d":
         """
@@ -488,7 +496,7 @@ def __init__(
                 - A 1D torch tensor
         """
         xyz = _handle_input(x, y, z, dtype, device, "Translate")
-        super().__init__(device=xyz.device)
+        super().__init__(device=xyz.device, dtype=dtype)
         N = xyz.shape[0]
 
         mat = torch.eye(4, dtype=dtype, device=self.device)
@@ -532,7 +540,7 @@ def __init__(
                 - 1D torch tensor
         """
         xyz = _handle_input(x, y, z, dtype, device, "scale", allow_singleton=True)
-        super().__init__(device=xyz.device)
+        super().__init__(device=xyz.device, dtype=dtype)
         N = xyz.shape[0]
 
         # TODO: Can we do this all in one go somehow?
@@ -571,7 +579,7 @@ def __init__(
 
         """
         device_ = get_device(R, device)
-        super().__init__(device=device_)
+        super().__init__(device=device_, dtype=dtype)
         if R.dim() == 2:
             R = R[None]
         if R.shape[-2:] != (3, 3):
@@ -598,7 +606,7 @@ def __init__(
         angle,
         axis: str = "X",
         degrees: bool = True,
-        dtype: torch.dtype = torch.float64,
+        dtype: torch.dtype = torch.float32,
         device: Optional[Device] = None,
     ) -> None:
         """
@@ -629,7 +637,7 @@ def __init__(
         # is for transforming column vectors. Therefore we transpose this matrix.
         # R will always be of shape (N, 3, 3)
         R = _axis_angle_rotation(axis, angle).transpose(1, 2)
-        super().__init__(device=angle.device, R=R)
+        super().__init__(device=angle.device, R=R, dtype=dtype)
 
 
 def _handle_coord(c, dtype: torch.dtype, device: torch.device) -> torch.Tensor:
@@ -646,8 +654,8 @@ def _handle_coord(c, dtype: torch.dtype, device: torch.device) -> torch.Tensor:
         c = torch.tensor(c, dtype=dtype, device=device)
     if c.dim() == 0:
         c = c.view(1)
-    if c.device != device:
-        c = c.to(device=device)
+    if c.device != device or c.dtype != dtype:
+        c = c.to(device=device, dtype=dtype)
     return c
 
 
@@ -696,7 +704,7 @@ def _handle_input(
         if y is not None or z is not None:
             msg = "Expected y and z to be None (in %s)" % name
             raise ValueError(msg)
-        return x.to(device=device_)
+        return x.to(device=device_, dtype=dtype)
 
     if allow_singleton and y is None and z is None:
         y = x

tests/test_transforms.py

@@ -87,6 +87,36 @@ def test_to(self):
             t = t.cuda()
             t = t.cpu()
 
+    def test_dtype_propagation(self):
+        """
+        Check that a given dtype is correctly passed along to child
+        transformations.
+        """
+        # Use at least two dtypes so we avoid only testing on the
+        # default dtype.
+        for dtype in [torch.float32, torch.float64]:
+            R = torch.tensor(
+                [[0.0, 1.0, 0.0], [0.0, 0.0, 1.0], [1.0, 0.0, 0.0]],
+                dtype=dtype,
+            )
+            tf = (
+                Transform3d(dtype=dtype)
+                .rotate(R)
+                .rotate_axis_angle(
+                    R[0],
+                    "X",
+                )
+                .translate(3, 2, 1)
+                .scale(0.5)
+            )
+
+            self.assertEqual(tf.dtype, dtype)
+            for inner_tf in tf._transforms:
+                self.assertEqual(inner_tf.dtype, dtype)
+
+            transformed = tf.transform_points(R)
+            self.assertEqual(transformed.dtype, dtype)
+
     def test_clone(self):
         """
         Check that cloned transformations contain different _matrix objects.
@@ -219,8 +249,8 @@ def test_rotate_axis_angle(self):
         normals_out_expected = torch.tensor(
             [[0.0, 1.0, 0.0], [0.0, 1.0, 0.0], [0.0, 1.0, 0.0]]
         ).view(1, 3, 3)
-        self.assertTrue(torch.allclose(points_out, points_out_expected))
-        self.assertTrue(torch.allclose(normals_out, normals_out_expected))
+        self.assertTrue(torch.allclose(points_out, points_out_expected, atol=1e-7))
+        self.assertTrue(torch.allclose(normals_out, normals_out_expected, atol=1e-7))
 
     def test_transform_points_fail(self):
         t1 = Scale(0.1, 0.1, 0.1)
@@ -951,7 +981,7 @@ def test_rotate_x_python_scalar(self):
         self.assertTrue(
             torch.allclose(transformed_points.squeeze(), expected_points, atol=1e-7)
         )
-        self.assertTrue(torch.allclose(t._matrix, matrix))
+        self.assertTrue(torch.allclose(t._matrix, matrix, atol=1e-7))
 
     def test_rotate_x_torch_scalar(self):
         angle = torch.tensor(90.0)