Taichi: Invalid input for autodiff, Kernel Simplicity Rule

[theorska]

Hi! I am trying to create a compute_loss function, where along the path of the object I calculate the closest distance, making that my loss. I am using this code:

@ti.kernel
def compute_loss(t: ti.i32):
    for i in range(1024):
        d=ti.sqrt(ti.sqr(help[1, 0][0] - path[i, 0][0]) + ti.sqr(help[1, 0][1] - path[i, 0][1]))
        if dist[0] > d:
            dist[0] = d
    dist[1] = 0.01
    loss[None] = ti.sqr(dist[1] - dist[0])

The error:
Traceback (most recent call last):
File "C:/Users/User/PycharmProjects/Try/Throwing.py", line 166, in
main()
File "C:/Users/User/PycharmProjects/Try/Throwing.py", line 140, in main
forward()
File "C:\Program Files (x86)\Microsoft Visual Studio\Shared\Python37_64\lib\site-packages\taichi\lang\tape.py", line 19, in __exit__
self.grad()
File "C:\Program Files (x86)\Microsoft Visual Studio\Shared\Python37_64\lib\site-packages\taichi\lang\tape.py", line 28, in grad
func.grad(*args)
File "C:\Program Files (x86)\Microsoft Visual Studio\Shared\Python37_64\lib\site-packages\taichi\lang\kernel.py", line 400, in __call__
return self.compiled_functionskey
File "C:\Program Files (x86)\Microsoft Visual Studio\Shared\Python37_64\lib\site-packages\taichi\lang\kernel.py", line 374, in func__
t_kernel()
RuntimeError: [reverse_segments.cpp:taichi::lang::irpass::reverse_segments@64] Invalid program input for autodiff. Please check the documentation for the "Kernel Simplicity Rule".
[E 03/09/20 17:10:42.325] [reverse_segments.cpp:taichi::lang::irpass::reverse_segments@64] Invalid program input for autodiff. Please check the documentation for the "Kernel Simplicity Rule".

Process finished with exit code 1

I don't see why to loop makes it invalid, without it, it runs fine. Can somebody point out the mistake for me?

[xumingkuan]

You can use ti.static(range(1024)) instead. Non-static for loops are not allowed with other statements.

Kernel Simplicity Rule: Kernel body consists of multiple simply nested for-loops. I.e., each for-loop can either contain exactly one (nested) for-loop (and no other statements), or a group of statements without loops.

[theorska]

Thank you, it solved that problem, but now it keeps crashing, I put it in a separate function to make it more readable.

@ti.kernel
def min_dis():
    for i in ti.static(range(1024)):
        d=ti.sqrt(ti.sqr(help[1, 0][0] - path[i, 0][0]) + ti.sqr(help[1, 0][1] - path[i, 0][1]))
        if dist[0] > d:
            dist[0] = ti.cast(d, real)

No error on the console, it just says on the cv2 image that loads up that it is not responding. In debug it crashes somewhere between

  min_dis()
  compute_loss(steps-1)

[xumingkuan]

Maybe you can change the for loop in min_dis() here back to the non-static one. One loop without other statements is OK. ti.static(range(1024)) in fact unrolls the loop and I'm not sure if the number 1024 is too large...

[archibate]

Maybe you want to use dist[0] = min(dist[0], d) instead of if, which is not supported by adjoint grad as far as I can see.

[theorska]

Thank you, changing it to non-static, and applying ti.min solved it.