How to optimize SVBRDF(Texture)

[tim37021]

it generates mipmap automatically in pyredner module. I am not quite sure how I can optimize texture texels

text = torch.tensor(np.ones()*0.5, requires_grad=True, device=pyredner.device()))
material.diffuse_reflectance = pyredner.Texture(text)
...
optimizer = torch.optim.Adam([text])

for ...:
  // I have to add retain_graph=True here to get it work.
  render()
  loss.backward(retain_graph=True)
  // text update correctly, but rendered image does not change..
  // then text will be gradually broken...
  optimizer.step()

[tim37021]

should I re-assign text to diffuse_reflectance??
it seems the texels is not updated even if text is change.

scene.materials[-2].diffuse_reflectance = pyredner.Texture(text)

[tim37021]

If I try to rebuild the graph by reassign texture to diffuse reflectance.
Cuda memory access error will occur

[BachiLi]

Hi,
The key is to do the re-assignment also in the gradient descent loop, so that PyTorch will propagate the derivatives for you. I added an example (and fixed a bug related to roughness texture derivatives) in https://github.com/BachiLi/redner/blob/master/tests/test_svbrdf.py
Let me know if there's other issues.

[BachiLi]

Closing this issue. Feel free to reopen if there are other questions.