Cannot add custom attributes to a flambe module

[asappresearch]

Cannot add custom attributes to a flambe module

[nmatthews-asapp]

Hey @peterAsapp nice catch, you are right that this error shouldn't be thrown, but the problem is actually misuse of the load method in the trainer. I'll try to explain why:

• The code you point out is actually correct; _load_from_state_dict in Pytorch will still call our hook. Notice in Component.__init__ how we register the hook that appears in the else case. The hook mechanism is new in PyTorch 1.1+ I believe, and you can learn more by diving into the source code for torch.nn.Module if you're curious
• On the other hand, because our saving mechanism uses these hooks to add extra things to the state_dict that pytorch doesn't expect, one should basically always use strict=False when using either our state or PyTorch's native state dict methods. (see https://flambe.ai/en/v0.4.6/autoapi/flambe/compile/component/index.html#flambe.compile.component.Component.load_state)
• The error you see here is actually caused by Trainer not using strict=False when loading the weights of the best model. I'll submit a PR to fix this shortly.