ignacio-rocco/detectorch

Modulelist instead of Sequential

ruotianluo opened this issue · 3 comments

ruotianluo commented

https://github.com/ignacio-rocco/detectorch/blob/master/lib/model/detector.py#L17

If you want to group layers, what about using ModuleList

ignacio-rocco commented

Hey, thanks. Sounds like a good idea.

I was not aware of nn.ModuleList.

Should we also do it for https://github.com/ignacio-rocco/detectorch/blob/master/lib/model/detector.py#L202?

In this case the modules don't have any parameters so maybe it's less necessary.

ruotianluo commented

I would suggest having that. When you are switching training mode, it will also affect all submodules. Although currently the proposal layer seems not care about the training mode at all, just it's good for later.