[feature request] Accepting python scalar inputs for torch.minimum/torch.maximum and friends

[vadimkantorov]

It will simplify the user code and will allow to use some placeholder default values like float('inf') (btw torch.minimum could be nop in this case).

>>> torch.rand(3, 4).mul(5.0)
# works

>>> torch.rand(3, 4).minimum(5.0)
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
TypeError: minimum(): argument 'other' (position 1) must be Tensor, not float

>>> torch.rand(3, 4).minimum(torch.tensor(5.0))
# works

[zou3519]

Seems reasonable to have

[vadimkantorov]

It seems that torch.clamp now supports both scalar and vector min/max arguments and supports broadcasting. I'm not sure if it supports differentiating through min/max arguments.

Could torch.maximum / torch.minimum just redirect to torch.clamp? (and maybe all of them should support differentiation through min/max although that could be forward compat breaking)