DeprecationWarning from `spacy_legacy`

Hi there, I recently upgraded to spacy 3 and scispacy 0.4, but I am now getting a warning whenever I use the small scispacy model (I have not tried any other model).

I am getting a DeprecationWarning on a fresh install in python 3.8 with the latest version of scispacy and en_core_sci_sm.

Steps to reproduce:

pip install scispacy pip install https://s3-us-west-2.amazonaws.com/ai2-s2-scispacy/releases/v0.4.0/en_core_sci_sm-0.4.0.tar.gz

import spacy
nlp = spacy.load("en_core_sci_sm")

import warnings
warnings.filterwarnings("error")
nlp("Hello World")

Any input to the nlp model triggers the same warning:

/opt/miniconda3/envs/clean/lib/python3.8/site-packages/spacy_legacy/layers/staticvectors_v1.py in forward(model, docs, is_train)
     43     )
     44     try:
---> 45         vectors_data = model.ops.gemm(model.ops.as_contig(V[rows]), W, trans2=True)
     46     except ValueError:
     47         raise RuntimeError(Errors.E896)

DeprecationWarning: Out of bound index found. 
This was previously ignored when the indexing result contained no elements. 
In the future the index error will be raised. 
This error occurs either due to an empty slice, or if an array has zero elements even before indexing.
(Use `warnings.simplefilter('error')` to turn this DeprecationWarning into an error and get more details on the invalid index.)

Any ideas as to how to resolve this without manually ignoring the warning?

[adrianeboyd]

Hmm, this is one of those cases where I hesitate to describe a workaround, even though I think it’s technically possible (you’d have to replace MultiHashEmbed and StaticVectors with custom code).

But I really wouldn’t recommend it, especially for a widely distributed model, so my official answer is: no, I would retrain the model with the corrected config.

[adrianeboyd]

Hi, it looks like en_core_sci_sm has include_static_vectors = true even though there are no vectors, which leads to this deprecation warning in numpy>=1.20.0.