Last active
February 6, 2022 08:26
-
-
Save etienne87/24c2a42d5816cfe61c1e1444f3a42a5d to your computer and use it in GitHub Desktop.
super slow torch.jit.script for loop
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
@torch.jit.script | |
def drloop(V: torch.Tensor,T: torch.Tensor) -> torch.Tensor: | |
n = len(V) | |
Ri = torch.zeros((3,), dtype=torch.float32, device=V.device) | |
Ri[0] = -T[0][1] | |
Ri[1] = T[0][0] | |
Ri[2] = 0 | |
Ris = [Ri[None]] | |
for i in range(n - 1): | |
V1 = V[i] | |
C: float = V1.dot(V1) | |
Ri_L: torch.Tensor = Ri - 2/C * Ri.dot(V1) * V1 | |
Ti_L: torch.Tensor = T[i] - 2/C * T[i].dot(V1) * V1 | |
V2: torch.Tensor = T[i+1] - Ti_L | |
C2: float = V2.dot(V2) | |
Ri = Ri_L - (2/C2) * V2.dot(Ri_L) * V2 | |
Ris.append(Ri[None]) | |
# type: torch.Tensor | |
return torch.cat(Ris) | |
def dr(points): | |
# https://www.microsoft.com/en-us/research/wp-content/uploads/2016/12/Computation-of-rotation-minimizing-frames.pdf | |
n = len(points) | |
V = torch.gradient(points, dim=0)[0] | |
norms = torch.linalg.norm(V, dim=1)[:,None] | |
T = V[...] / norms | |
R = drloop(V,T) | |
R = R / torch.linalg.norm(R, dim=1)[:,None] | |
S = torch.cross(R,T) | |
return T,S,R | |
X = th.randn(1000,3, requires_grad=True) | |
t1 = time.time() | |
T,U,V = dr(X) | |
t2 = time.time() | |
print(t2-t1) | |
loss = (T+U+V).sum() | |
loss.backward() |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment