Variable number of input channels for vision models

[goergen95]

Hello, thanks again for the great work on making torch available to R users! I was wondering if you are interested in allowing users to specify the number of input channels for the vision models that this package provides. Of course, that would not allow to use pre-trained weights, but there are a number of domains (e.g. Earth Observation) where training a model from scratch with a different number of input channels can be useful. I have branch implementing this over here. If you are interested I can open a pull-request. Best, Darius.