Feature/sg 849 add replace in channels

[Louis-Dupont]

Adding replace_input_channels and get_input_channels to SgModel (and other required stems/backbones)

[shaydeci]

Adding replace_input_channels and get_input_channels to SgModel (and other required stems/backbones)

Question

How to handle models that were implemented with self.in_channels attribute, since we now have get_input_channels() method that dynamically checks the value without relying on an attribute?

a. Should we remove any self.in_channels attribute?

pro: cleaner IMO con: risk of not being backward compatible if models are built on top of these models.

b. Or should we update the value of self.in_channels when calling replace_input_channels() ? pro: 100% backward compatible con: risk of bug if we miss one place where self.in_channels should be updated con: redundant and maybe a bit more confusing

Most important thing this PR is missin, is a mechanism to replace the stem out of the box when using pretrained weights.
One should be able to use pretrained checkpoints, replacing the stem behind the scenes.

[Louis-Dupont]

Most important thing this PR is missin, is a mechanism to replace the stem out of the box when using pretrained weights.
One should be able to use pretrained checkpoints, replacing the stem behind the scenes.

Added here 287b28f

[Louis-Dupont]

Added update of self.in_channels when calling replace_input_channels: ab760e8

[shaydeci]

LGTM