Normalize silently ignores in_min or in_max when only one is provided

[Sal-sal-sal]

Description

Normalize ignores an explicitly supplied in_min when in_max is None, and vice versa.
Both bounds are calculated from the input data instead.

The parameter documentation describes automatic calculation when the corresponding bound is None, so I would expect an explicitly supplied bound to be preserved.

Reproduction

import torch
import torchio as tio

image = tio.ScalarImage(
    torch.tensor([0., 50., 100.]).reshape(1, 1, 1, 3)
)

for bounds in ({"in_min": 50.}, {"in_max": 50.}):
    result = tio.Normalize(
        out_min=0.,
        out_max=1.,
        **bounds,
    )(image)
    print(bounds, result.data.flatten().tolist())

Actual behavior

{'in_min': 50.0} [0.0, 0.5, 1.0]
{'in_max': 50.0} [0.0, 0.5, 1.0]

Expected behavior

{'in_min': 50.0} [0.0, 0.0, 1.0]
{'in_max': 50.0} [0.0, 1.0, 1.0]

Possible cause

Normalize.make_params() uses the explicit bounds only when both are non-None:

if self.in_min is not None and self.in_max is not None:

Otherwise, it computes both bounds from percentiles.
Each bound could be resolved independently, calculating only the missing one.

Version

Reproduced on main, commit 2b019d2a9659c838408c73ae77327ac5eb87b190, TorchIO 2.0.0a2, using synthetic CPU tensors.