Grad All Zero

[zichenwang01]

Hi,

I coded a custom integrator, defined a loss function between the reference image and the custom rendered image, and want to backpropagate the loss to the custom rendered image. However, I am having the grads all zero right now.

image1 = mi.render(scene, spp=64)

sdf_integrator = mi.load_dict({
   'type': 'sdf-direct-ps'
})
image2 = sdf_integrator.render(scene=scene, sdf=sdf, sensor=sensor, spp=64)

dr.enable_grad(image1)
dr.enable_grad(image2)
loss = l1_loss(image1, image2)
# dr.backward(loss)
dr.backward_to(image2)
grad = dr.grad(image2)

The weird place is that if I backpropagate to image1, then everything is fine, but things do not work when I backpropagate to image2.

Thanks for any help in advance!

[zichenwang01]

Nvm, I think I figured out the problem: I need to dr.set_grad(loss, 1) before backpropagating!