AD Update Parameter

[zichenwang01]

Hi,

I want to forward a single parameter to the rendered image and optimize the parameter. However, right now once I update the parameter, the gradients all become zero. Am I updating the parameter incorrectly?

image = my_integrator.render(p)
dr.set_grad(p, 1)
dr.forward_to(p, image)
print(dr.grad(image)) # works normally

p += 0.001
image = my_integrator.render(p)
dr.set_grad(p, 1)
dr.forward_to(p, image)
print(dr.grad(image)) # grads all zero

[zichenwang01]

Btw, if I do not update the parameter with p += 0.001, then the second grad is correct

[njroussel]

Hi @zichenwang01

Could you post a full reproducer? There are many possible causes to this issue. I'd say what most commonly happens here is that you're not explicitly calling params.update().

[zichenwang01]

Indeed, I did not call params.update() because p in my setting is not inside params but instead a global variable. Also, I did not use an optimizer but instead wrote a simple gradient descent.

    for epoch in range(100):
    total_loss = 0
    for i in range(1):
        img = sdf_integrator.render(scene=scene, params=params, sdf=sdf, 
                                sensor=sensors[i], spp=16)
        
        loss = l2_loss(ref_imgs[i], img)
        print("loss: ", loss)
        
        dr.set_grad(loss, 1.0)
        # dr.backward(loss)
        dr.backward_to(p)
        total_loss += loss[0]
        
        p = p - 1e-2 * dr.grad(loss)
        params.update()
        print("p: ", p)

Here is a snippet of my custom integrator. No other places use p

    def sample(self, 
          scene:mi.Scene, sdf:SDF, sampler:mi.Sampler, 
          ray:mi.Ray3f, depth = 0, 
          active = True, **kwargs
    ):       
    # ray intersection with the scene 
    itx : SurfaceInteraction3f = scene.ray_intersect(ray)
    itx.p = dr.replace_grad(itx.p, Vector3f(0, p, 0)) # sphere move up
    # itx.p = dr.replace_grad(itx.p, Vector3f(p, 0, 0)) # sphere move right

The full reproducer contains a base integrator, a custom direct integrator, a scene XML, and an optimization file. I can provide these if needed!

[njroussel]

Hi @zichenwang01

I think this just basic Python at play After the p = p - 1e-2 * dr.grad(loss) the p variable in the body no longer matches the p in the integrator sdf_integrator. You're then trying to backward_to(p) which is the p defined in the loop body and not the integrator and is therefore completely detached from the loss computation (this will obviously give a 0 gradient) .
This is hard to write out clearly, I hope it was clear. Your issue should be solved by doing something like sdf_integrator.p = p.

Without a full minimal reproducer I can't tell for sure.

[zichenwang01]

I see. Thanks!