π₯ Use Hugging Face text and token classification pipelines directly in spaCy
MIT License
This package provides spaCy components to use pretrained Hugging Face Transformers pipelines for inference only.
dslim/bert-base-NER
anddistilbert-base-uncased-finetuned-sst-2-english
.Installing the package from pip will automatically install all dependencies, including PyTorch and spaCy.
pip install -U pip setuptools wheel
pip install spacy-huggingface-pipelines
For GPU installation, follow the spaCy installation quickstart with GPU, e.g.
pip install -U spacy[cuda12x]
If you are having trouble installing PyTorch, follow the instructions on the official website for your specific operating system and requirements.
This module provides spaCy wrappers for the inference-only transformers
TokenClassificationPipeline
and
TextClassificationPipeline
pipelines.
The models are downloaded on initialization from the Hugging Face Hub if they're not already in your local cache, or alternatively they can be loaded from a local path.
Note that the transformer model data is not saved with the pipeline when you
call nlp.to_disk
, so if you are loading pipelines in an environment with
limited internet access, make sure the model is available in your
transformers cache directory
and enable offline mode if needed.
Config settings for hf_token_pipe
:
[components.hf_token_pipe]
factory = "hf_token_pipe"
model = "dslim/bert-base-NER" # Model name or path
revision = "main" # Model revision
aggregation_strategy = "average" # "simple", "first", "average", "max"
stride = 16 # If stride >= 0, process long texts in
# overlapping windows of the model max
# length. The value is the length of the
# window overlap in transformer tokenizer
# tokens, NOT the length of the stride.
kwargs = {} # Any additional arguments for
# TokenClassificationPipeline
alignment_mode = "strict" # "strict", "contract", "expand"
annotate = "ents" # "ents", "pos", "spans", "tag"
annotate_spans_key = null # Doc.spans key for annotate = "spans"
scorer = null # Optional scorer
TokenClassificationPipeline
settingsmodel
: The model name or path.revision
: The model revision. For production use, a specific git commit ismain
.stride
: For stride >= 0
, the text is processed in overlapping windowsstride
setting specifies the number of overlapping tokens betweenstride
is None
, then the text may bestride
is only supported for fast tokenizers.aggregation_strategy
: The aggregation strategy determines the word-levelkwargs
: Any additional arguments toTokenClassificationPipeline
.alignment_mode
determines how transformer predictions are aligned to spaCyDoc.char_span
.annotate
and annotate_spans_key
configure how the annotation is saved totoken.tag_
, token.pos_
(only fordoc.ents
or doc.spans
.Doc.ents
:import spacy
nlp = spacy.blank("en")
nlp.add_pipe("hf_token_pipe", config={"model": "dslim/bert-base-NER"})
doc = nlp("My name is Sarah and I live in London")
print(doc.ents)
# (Sarah, London)
Doc.spans[spans_key]
and scores asDoc.spans[spans_key].attrs["scores"]
:import spacy
nlp = spacy.blank("en")
nlp.add_pipe(
"hf_token_pipe",
config={
"model": "dslim/bert-base-NER",
"annotate": "spans",
"annotate_spans_key": "bert-base-ner",
},
)
doc = nlp("My name is Sarah and I live in London")
print(doc.spans["bert-base-ner"])
# [Sarah, London]
print(doc.spans["bert-base-ner"].attrs["scores"])
# [0.99854773, 0.9996215]
Token.tag
:import spacy
nlp = spacy.blank("en")
nlp.add_pipe(
"hf_token_pipe",
config={
"model": "QCRI/bert-base-multilingual-cased-pos-english",
"annotate": "tag",
},
)
doc = nlp("My name is Sarah and I live in London")
print([t.tag_ for t in doc])
# ['PRP$', 'NN', 'VBZ', 'NNP', 'CC', 'PRP', 'VBP', 'IN', 'NNP']
Token.pos
:import spacy
nlp = spacy.blank("en")
nlp.add_pipe(
"hf_token_pipe",
config={"model": "vblagoje/bert-english-uncased-finetuned-pos", "annotate": "pos"},
)
doc = nlp("My name is Sarah and I live in London")
print([t.pos_ for t in doc])
# ['PRON', 'NOUN', 'AUX', 'PROPN', 'CCONJ', 'PRON', 'VERB', 'ADP', 'PROPN']
Config settings for hf_text_pipe
:
[components.hf_text_pipe]
factory = "hf_text_pipe"
model = "distilbert-base-uncased-finetuned-sst-2-english" # Model name or path
revision = "main" # Model revision
kwargs = {} # Any additional arguments for
# TextClassificationPipeline
scorer = null # Optional scorer
The input texts are truncated according to the transformers model max length.
TextClassificationPipeline
settingsmodel
: The model name or path.revision
: The model revision. For production use, a specific git commit ismain
.kwargs
: Any additional arguments toTextClassificationPipeline
.import spacy
nlp = spacy.blank("en")
nlp.add_pipe(
"hf_text_pipe",
config={"model": "distilbert-base-uncased-finetuned-sst-2-english"},
)
doc = nlp("This is great!")
print(doc.cats)
# {'POSITIVE': 0.9998694658279419, 'NEGATIVE': 0.00013048505934420973}
Both token and text classification support batching with nlp.pipe
:
for doc in nlp.pipe(texts, batch_size=256):
do_something(doc)
If the component runs into an error processing a batch (e.g. on an empty text),
nlp.pipe
will back off to processing each text individually. If it runs into
an error on an individual text, a warning is shown and the doc is returned
without additional annotation.
Switch to GPU:
import spacy
spacy.require_gpu()
for doc in nlp.pipe(texts):
do_something(doc)
Please report bugs in the spaCy issue tracker or open a new thread on the discussion board for other issues.