[Question] Setting a constraint for the ConstantMean()

[vr308]

Hi - I am trying to set a GreaterThan() constraint for the constant of the mean_module -- which I thought is trival following the examples from the docs, but unfortunately, the training doesn't seem to obey it.

class ExactGPModel(gpytorch.models.ExactGP):
    
    def __init__(self, train_x, train_y, likelihood, kernel):
        super(ExactGPModel, self).__init__(train_x, train_y, likelihood)
        self.mean_module = gpytorch.means.ConstantMean(constant_constraint=gpytorch.constraints.GreaterThan(0.5))
        self.covar_module = kernel

    def forward(self, x):
        mean_x = self.mean_module(x)
        covar_x = self.covar_module(x)
        return gpytorch.distributions.MultivariateNormal(mean_x, covar_x)

   model = ExactGPModel(train_x, train_y, likelihood, kernel).double().cuda()
   model.mean_module.register_constraint('constant', gpytorch.constraints.GreaterThan(0.5))
  

Anything missing here?

Upon some digging, I noticed that the constant mean is slightly different to the other hyperparameters which are tensors while the constant mean is a Parameter -- could that be it?

model.likelihood.noise
tensor([0.2515], device='cuda:0', dtype=torch.float64, grad_fn=<AddBackward0>)

model.mean_module.constant
Parameter containing:
tensor([0.0770], device='cuda:0', dtype=torch.float64, requires_grad=True)

[gpleiss]

Hmm this is unexpected... I'll dig into that.

[gpleiss]

Turns out ConstantMean never supported constraints, but I'll put up a PR to fix that!