Daniël de Kok

[sticker](https://github.com/danieldk/sticker) is a neural sequence labeler (bidi RNNs or dilated convolution networks) that can do part-of-speech tagging. We use it for tagging German, with an accuracy of 99.10% for in-domain...

I think changing people's configuration files is a bit invasive and could lead to race conditions (e.g. if the user has left open an editor). What we could do is...

At this point graph compatibility is everything. I guess the ramification here would be that we need three additional `Option`al ops in `TaggerGraph`, right?

No more things in the configuration file, there is too much stuff in there already that is only relevant to training. Maybe `clap` offers some functionality to only unveil options...

There is `requires`, not sure if it hides an option if it is not present: https://kbknapp.github.io/clap-rs/clap/struct.Arg.html#method.requires

Yep. It's worth trying if they get hidden if the requirement is not given. But I guess at the very least it would also group the arguments together in usage...

The output is already a bit wide. I wonder if we really need the progress bar. It looks nice, but it takes up a lot of space and does not...

At least Tensorboard makes it possible to monitor a lot more things without things getting too crowded. In the terminal output I am typically only interested in the train/validation accuracy...

Currently, we primarily use sticker through the `sticker-tag` and `sticker-server` utilities and not as a library. It could be used as a library in the future, but some functionality from...

`transformers` updates have broken things in the past. So rather than removing the upper bound, it should be set to some version against which the package is tested.