set_optimizable wrong param?

[Ska-p]

in reconstruction function, inside train_across_scene.py while iterating through the steps_inner value, it is written when satysfying the condition j%steps_inner==steps_inner-3

model.set_optimizable['mlp', 'basis', 'renderer'], True)

but mlp is not a parameter in the set_optmizable function

Should it be model.set_optimizable(['proj','basis','renderer'], True) ?

[Kilichbek]

• yes, I think aslo this might be wrong. Also, when I run the code for image set regression and after training I try to fit to new test image by freezing projection layer and basis layers. But it turns out these layers are being updated as well. When I change the method in model to the following it is working as expected but performance is not good:

  def set_optimizable(self, items, statue):
      for item in items:
          if item == 'basis' and self.cfg.model.basis_type != 'none':
              for item in self.basises:
                  for param in item.parameters():
                      param.requires_grad = statue
                  #item.requires_grad = statue
          elif item == 'coeff' and self.cfg.model.coeff_type != 'none':
              for item in self.basises:
                  item.requires_grad = statue
          elif item == 'proj':
              for param in self.linear_mat.parameters():
                      param.requires_grad = statue

[Ska-p]

I'm working with 3D reconstruction. I've changed the set_optimizable function to the following:

    def set_optimizable(self, items, status):
        for item in items:
            if item == 'basis' and self.cfg.model.basis_type != 'none':
                self.basises.requires_grad_(requires_grad=status)
            elif item == 'coeff' and self.cfg.model.coeff_type != 'none':
                self.coeffs.requires_grad_(requires_grad=status)
            elif item == 'proj':
                self.linear_mat.requires_grad_(requires_grad=status)
            elif item == 'renderer':
                self.renderModule.requires_grad_(requires_grad=status)

In the older version, when the method entered the 'coeff' elif, it was actually working on the basises module so basically the coefficient where never stopped from updating (in your code the problem is still there). After changing the function I tracked with pdb.set_trace() and apparently the function requiresgrad(value) correctly sets the value as the argument. I'm having as well worse results than the one stated in the paper but they are close.

Update: I checked as well the values of the parameters in the scenario where basises, proj and render do NOT have to be optimized. With my settings the values remains frozen so i believe it should be ok like this.