Guillaume Klein

There are currently no plans to add this but I would love if someone could look into this.

Correct. Only the PyTorch version has this feature.

We could exclude the `pyonmttok` package on this platform using environment markers. Something like this: ```diff diff --git a/setup.py b/setup.py index c086b849..ae285bcd 100644 --- a/setup.py +++ b/setup.py @@ -27,7 +27,7...

`pyonmttok` now includes Python wheels for ARM64 macOS, so the installation should now succeed.

This sounds interesting. We will probably not integrate this specific library, but allowing users to provide a file with callbacks is something we could do.

Currently it does not. Marking this as "contributions welcome", it would be nice to support this.

Thank you for the request. 1. Do you confirm it is the feature described in [this forum thread](http://forum.opennmt.net/t/teacher-forcing/104)? 2. The sentence probability is already [returned by the model](https://github.com/OpenNMT/OpenNMT-tf/blob/master/opennmt/models/sequence_to_sequence.py#L210) (not printed...

*(Answering here.)* If you want to work on this, you should mostly focus on the decoding phase, e.g. for a self-attention decoder: * [decoders/self_attention_decoder.py](https://github.com/OpenNMT/OpenNMT-tf/blob/master/opennmt/decoders/self_attention_decoder.py) * [utils/beam_search.py](https://github.com/OpenNMT/OpenNMT-tf/blob/master/opennmt/utils/beam_search.py) You might also need...

It's an old issue and you probably moved on but for reference, 2. and 3. of the inital request are now supported with the `score` run type and the following...

You are correct, additional word features are only supported on the source side using parallel inputs.