Matthias Sperber

I see, yeah I had suspected something like that. In that case I think this can be merged (once the merge conflicts are resolved)! -- Matthias _(this comment is "not...

Hi Sruti, it seems that this recipe was not adjusted to the latest change in how decoders are specified. I can't send code right now unfortunately, but if you look...

@SrutiBh you are correct that the rnn_layer parameter needs to be renamed. The config specifies a (bidirectional) pyramidal LSTM as encoder, and a unidirectional LSTM as decoder. This is correct,...

The example config folder has gotten a bit big, some parts are also redundant with test configs, others even outdated. Shall we begin by only keeping some xnmt core features...

Yes, we should definitely make things modular and move them to the modelparts package. However, the transformer needs some refactoring because it doesn't use the standard inference mechanism that's used...

I'm wondering what the cleanest way to implement this would be, requiring little or no changes to the existing model components. Preferably, we could do some stuff behind the scenes...

Any hints, maybe @philip30 ? I'll be happy to fix this myself but don't want to undo any previous design decisions and would like to know what the best way...

No, I think nothing has been done along these lines yet.

Making config files and saved experiments compatible has been implemented by #491. Some thoughts on what would need to be done to support resuming crashed experiments: - make .mod and...

I think we're currently only testing that loading a model doesn't cause a crash, so adding a correctness test seems like a nice idea.