read and write methods for nn.Module.

[dominikgrewe]

Allows modules that want to implement custom read and write methods to
call parent.read/write.

[nicholas-leonard]

@dominikgrewe This won't break anything, but now all modules will be serialized through those methods you just created.

[dominikgrewe]

Yes, but they're just doing what torch serialization would have done anyway (apart from filtering out non-writable objects).

[soumith]

looks pretty useful! Thanks.

[soumith]

@nicholas-leonard this is useful when implementing custom read-write, you can change some fields and then just call parent.read /write

[eladhoffer]

@soumith can't we use this to avoid the need for #526 by adding an option to write just weight, bias and running_mean/var (empty tensors for all other fields)?

[dominikgrewe]

There are various other things you might want to save though, e.g. the convolution settings (kernel size, stride, padding etc) for a convolution module. So we could only do without clearState if many modules had custom read & write functions. It's worth thinking about though. What other uses for clearState are there aside from serialization?

[eladhoffer]

@dominikgrewe yes, what I meant is removing the output, gradInput and buffer tensors. Other fields should be kept.

[soumith]

@eladhoffer not a good idea. if the intention is to save a model, it's to save the whole model as is. we can introduce an nn.save that calls clearState before hand, and that's okay.