Fix sample from prior for ConstantMean

[dme65]

The use of fill_ doesn't play well with the shape of the priors.

[Balandat]

Thanks. What exactly goes wrong when using fill_ here? Is the issue that priors.sample() produces a tensor of some shape rather than a scalar and so fill_ does weird things? Would be good to add some type annotation to understand what _constant_closure is supposed to operate on.

Also, does that mean all the other modules should be updated in the same way?

[Balandat]

is this gating necessary or could we just always apply torch.as_tensor (should be a nullop)?

[dme65]

I'm just following what is done for other closures, for example: https://github.com/cornellius-gp/gpytorch/blob/master/gpytorch/kernels/kernel.py#L262.

[dme65]

Is the issue that priors.sample() produces a tensor of some shape rather than a scalar and so fill_ does weird things

Yes, fill_ will fail if you call it with a tensor with more than 0-dimensions. For example:

x = torch.randn(5)
y = torch.randn(1)
x.fill_(y)

results in RuntimeError: fill_ only supports 0-dimension value tensor but got tensor with 1 dimensions.