Token vectors are empty using en_core_web_trf model

[polm]

This is a copy of #8037 / #8047 because for some reason the migration on that won't finish. This was originally posted by @erip.

---

How to reproduce the behaviour

>>> import spacy
>>> spacy.__version__
'3.0.6'
>>> nlp = spacy.load('en_core_web_trf')
>>> s = "Hello, my name is Elijah."
>>> doc = nlp(s)
>>> doc[0].vector
array([], dtype=float32)

Your Environment

Info about spaCy

• spaCy version: 3.0.6
• Platform: macOS-10.16-x86_64-i386-64bit
• Python version: 3.8.8
• Pipelines: en_core_web_lg (3.0.0), en_core_web_trf (3.0.0)

---

I imagine the model is trained on subwords, so maybe the alignment between those and tokens in the spaCy since is causing issues?

[polm]

Sorry for the delayed reply on this, not sure what's up with the issue migration.

This is a design decision and not a bug. Basically the .vector api is only for static word vectors, not for contextual vectors like those generated by the Transformer. The Transformer models in spaCy don't include static word vectors because if you have Transformers you usually don't need them. If you need per-token representations, what you can do instead is use the data in doc._.trf_data, which contains tensors, wordpieces, and an alignment between spaCy tokens and the wordpieces. (I'm not sure there's a guide to this anywhere yet.)

[polm]

I will add this is a bit confusing because the small models don't have static vectors, but fall back to the tensor attribute when you use the vector API. The Transformers models don't have the tensor attribute either, due to the Wordpiece/spaCy token mapping issue, so you just get an empty array.

The consistency of this could probably be improved but I don't think we have any concrete plans now.