Skip to content

Latest commit

 

History

History
196 lines (142 loc) · 7.84 KB

readme.md

File metadata and controls

196 lines (142 loc) · 7.84 KB

spaCy-wrap: For Wrapping fine-tuned transformers in spaCy pipelines

PyPI version python version Code style: black github actions pytest github actions docs github coverage

spaCy-wrap is a minimal library intended for wrapping fine-tuned transformers from the Huggingface model hub in your spaCy pipeline allowing the inclusion of existing models within SpaCy workflows.

As far as possible it follows a similar API as spacy-transformers.

NOTE: Since the release of spaCy-wrap, Explosion released the spacy-huggingface-pipelines it takes the approach of wrapping the Huggingface pipeline as opposed to the transformer. That means token aggregation and conversion into spans happens at the Huggingface pipeline, while in spaCy-wrap it happens at the logits of the model which can sometimes lead to unfortunate differences in results. I generally recommend using the spacy-huggingface-pipelines for most use cases, but if you need to use the transformer output more directly spaCy-wrap can have its uses.

Installation

Installing spacy-wrap is simple using pip:

pip install spacy_wrap

Examples

The following shows a simple example of how you can quickly add a fine-tuned transformer model from the Huggingface model hub for either text classification, named entity or token classification.

Sequence Classification

In this example, we will use a model fine-tuned for sentiment classification on SST2. This model classifies whether a text is positive or negative. We will add this model to a blank English pipeline:

import spacy
import spacy_wrap

nlp = spacy.blank("en")

config = {
    "doc_extension_trf_data": "clf_trf_data",  # document extention for the forward pass
    "doc_extension_prediction": "sentiment",  # document extention for the prediction
    "model": {
        # the model name or path of huggingface model
        "name": "distilbert-base-uncased-finetuned-sst-2-english",  
    },
}

transformer = nlp.add_pipe("sequence_classification_transformer", config=config)

doc = nlp("spaCy is a wonderful tool")

print(doc.cats)
# {'NEGATIVE': 0.001, 'POSITIVE': 0.999}
print(doc._.sentiment)
# 'POSITIVE'
print(doc._.clf_trf_data)
# TransformerData(wordpieces=...

These pipelines can also easily be applied to multiple documents using the nlp.pipe as one would expect from a spaCy component:

docs = nlp.pipe(
    [
        "I hate wrapping my own models",
        "Isn't there a tool for this?!",
        "spacy-wrap is great for wrapping models",
    ]
)

for doc in docs:
    print(doc._.sentiment)
# 'NEGATIVE'
# 'NEGATIVE'
# 'POSITIVE'

More Examples

It is always nice to have more than one example. Here is another one where we add the Hate speech model for Danish to a blank Danish pipeline:

import spacy
import spacy_wrap

nlp = spacy.blank("da")

config = {
    "doc_extension_trf_data": "clf_trf_data",  # document extention for the forward pass
    "doc_extension_prediction": "hate_speech",  # document extention for the prediction
    # choose custom labels
    "labels": ["Not hate Speech", "Hate speech"],
    "model": {
        "name": "DaNLP/da-bert-hatespeech-detection",  # the model name or path of huggingface model
    },
}

transformer = nlp.add_pipe("classification_transformer", config=config)

doc = nlp("Senile gamle idiot") # old senile idiot

doc._.clf_trf_data
# TransformerData(wordpieces=...
doc._.hate_speech
# "Hate speech"
doc._.hate_speech_prob
# {'prob': array([0.013, 0.987], dtype=float32), 'labels': ['Not hate Speech', 'Hate speech']}

Token Classification

We can also use the model for token classification:

import spacy
import spacy_wrap
nlp = spacy.blank("en")

config = {"model": {"name": "vblagoje/bert-english-uncased-finetuned-pos"}, 
          # "predictions_to": ["pos"]  # optional, can be "pos", "tag" or "ents"
}

snlp.add_pipe("token_classification_transformer", config=config)

text = "My name is Wolfgang and I live in Berlin"

doc = nlp(text)
print(doc._.tok_clf_predictions)
# ['PRON', 'NOUN', 'AUX', 'PROPN', 'CCONJ', 'PRON', 'VERB', 'ADP', 'PROPN']

By default, spacy-wrap will automatically detect it the labels follow the universal POS tags as well. If so it will also assign it to the token.pos, similar regular spacy pipelines:

print(doc[0].pos_)
# 'PRON'

Named Entity Recognition

In this example, we use a model fine-tuned for named entity recognition. spacy-wrap will in this case infer from the IOB tags that the model is intended for named entity recognition and assign it to doc.ents.

import spacy
import spacy_wrap
nlp = spacy.blank("en")

# specify model from the hub
config = {"model": {"name": "dslim/bert-base-NER"}, 
          "predictions_to": ["ents"]} # forced to be named entity recognition, if left out it will be estimated from the labels

# add it to the pipe
nlp.add_pipe("token_classification_transformer", config=config)

doc = nlp("My name is Wolfgang and I live in Berlin.")

print(doc.ents)
# (Wolfgang, Berlin)

📖 Documentation

Documentation
🔧 Installation Installation instructions for spacy-wrap.
📰 News and changelog New additions, changes and version history.
🎛 Documentation The reference for spacy-wrap's API.

💬 Where to ask questions

Type
🚨 FAQ FAQ
🚨 Bug Reports GitHub Issue Tracker
🎁 Feature Requests & Ideas GitHub Issue Tracker
👩‍💻 Usage Questions GitHub Discussions
🗯 General Discussion GitHub Discussions