Update Rotate transform to use device of input rotation

theschnitz · facebook-github-bot · commit cd5af2521a56 · 2021-05-19T10:04:12.000-07:00
Summary: Currently the Rotate transform does not consider the R's device at all, resulting in errors if you're expecting it to be on cuda but it gets the default casting to cpu. This updates the transform to respect R's device.

Reviewed By: nikhilaravi

Differential Revision: D27828118

fbshipit-source-id: ddd99f73eadbd990688eb22f3d1ffbacbe168c81
diff --git a/pytorch3d/transforms/transform3d.py b/pytorch3d/transforms/transform3d.py
@@ -434,7 +434,7 @@ def cuda(self):
 
 
 class Translate(Transform3d):
-    def __init__(self, x, y=None, z=None, dtype=torch.float32, device="cpu"):
+    def __init__(self, x, y=None, z=None, dtype=torch.float32, device=None):
         """
         Create a new Transform3d representing 3D translations.
 
@@ -448,11 +448,11 @@ def __init__(self, x, y=None, z=None, dtype=torch.float32, device="cpu"):
                 - A torch scalar
                 - A 1D torch tensor
         """
-        super().__init__(device=device)
         xyz = _handle_input(x, y, z, dtype, device, "Translate")
+        super().__init__(device=xyz.device)
         N = xyz.shape[0]
 
-        mat = torch.eye(4, dtype=dtype, device=device)
+        mat = torch.eye(4, dtype=dtype, device=self.device)
         mat = mat.view(1, 4, 4).repeat(N, 1, 1)
         mat[:, 3, :3] = xyz
         self._matrix = mat
@@ -468,7 +468,7 @@ def _get_matrix_inverse(self):
 
 
 class Scale(Transform3d):
-    def __init__(self, x, y=None, z=None, dtype=torch.float32, device="cpu"):
+    def __init__(self, x, y=None, z=None, dtype=torch.float32, device=None):
         """
         A Transform3d representing a scaling operation, with different scale
         factors along each coordinate axis.
@@ -485,12 +485,12 @@ def __init__(self, x, y=None, z=None, dtype=torch.float32, device="cpu"):
                 - torch scalar
                 - 1D torch tensor
         """
-        super().__init__(device=device)
         xyz = _handle_input(x, y, z, dtype, device, "scale", allow_singleton=True)
+        super().__init__(device=xyz.device)
         N = xyz.shape[0]
 
         # TODO: Can we do this all in one go somehow?
-        mat = torch.eye(4, dtype=dtype, device=device)
+        mat = torch.eye(4, dtype=dtype, device=self.device)
         mat = mat.view(1, 4, 4).repeat(N, 1, 1)
         mat[:, 0, 0] = xyz[:, 0]
         mat[:, 1, 1] = xyz[:, 1]
@@ -509,7 +509,7 @@ def _get_matrix_inverse(self):
 
 class Rotate(Transform3d):
     def __init__(
-        self, R, dtype=torch.float32, device="cpu", orthogonal_tol: float = 1e-5
+        self, R, dtype=torch.float32, device=None, orthogonal_tol: float = 1e-5
     ):
         """
         Create a new Transform3d representing 3D rotation using a rotation
@@ -520,6 +520,7 @@ def __init__(
             orthogonal_tol: tolerance for the test of the orthogonality of R
 
         """
+        device = _get_device(R, device)
         super().__init__(device=device)
         if R.dim() == 2:
             R = R[None]
@@ -548,7 +549,7 @@ def __init__(
         axis: str = "X",
         degrees: bool = True,
         dtype=torch.float64,
-        device="cpu",
+        device=None,
     ):
         """
         Create a new Transform3d representing 3D rotation about an axis
@@ -578,7 +579,7 @@ 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=device, R=R)
+        super().__init__(device=angle.device, R=R)
 
 
 def _handle_coord(c, dtype, device):
@@ -595,9 +596,24 @@ def _handle_coord(c, dtype, device):
         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)
     return c
 
 
+def _get_device(x, device=None):
+    if device is not None:
+        # User overriding device, leave
+        device = device
+    elif torch.is_tensor(x):
+        # Set device based on input tensor
+        device = x.device
+    else:
+        # Default device is cpu
+        device = "cpu"
+    return device
+
+
 def _handle_input(x, y, z, dtype, device, name: str, allow_singleton: bool = False):
     """
     Helper function to handle parsing logic for building transforms. The output
@@ -626,6 +642,7 @@ def _handle_input(x, y, z, dtype, device, name: str, allow_singleton: bool = Fal
     Returns:
         xyz: Tensor of shape (N, 3)
     """
+    device = _get_device(x, device)
     # If x is actually a tensor of shape (N, 3) then just return it
     if torch.is_tensor(x) and x.dim() == 2:
         if x.shape[1] != 3:
@@ -634,7 +651,7 @@ def _handle_input(x, y, z, dtype, device, name: str, allow_singleton: bool = Fal
         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
+        return x.to(device=device)
 
     if allow_singleton and y is None and z is None:
         y = x
@@ -665,6 +682,7 @@ def _handle_angle_input(x, dtype, device, name: str):
         - Python scalar
         - Torch scalar
     """
+    device = _get_device(x, device)
     if torch.is_tensor(x) and x.dim() > 1:
         msg = "Expected tensor of shape (N,); got %r (in %s)"
         raise ValueError(msg % (x.shape, name))
