BackProp through Progressive nets

[arcosin]

BackProp through Progressive nets

[mbelalsh]

Thanks for your response. If I have any plans of implementing it, I will let you know.

[arcosin]

I'm glad you like it!

Regarding backpropagation, that is generally computed automatically by autograd in pytorch. As its a module, a column can be trained just like any other neural net in pytorch using autograd, but the whole Progressive net can as well if freezing is used correctly (freezing all previous columns before making a new one).

As for the adapters modification, we may add that in later. It hasn't been needed thus far. If you want to implement that as an option for your own uses on a fork and then create a pull request, that would be greatly appreciated :)