spaCy pipelines for pre-trained BERT, XLNet and GPT-2

spacy-transformers

This package (previously spacy-pytorch-transformers) provides spaCy model pipelines that wrap Hugging Face’s transformers package, so you can use them in spaCy. The result is convenient access to state-of-the-art transformer architectures, such as BERT, GPT-2, XLNet, etc. For more details and background, check out our blog post.

Features

• Use BERT, RoBERTa, XLNet and GPT-2 directly in your spaCy pipeline.
• Fine-tune pretrained transformer models on your task using spaCy’s API.
• Custom component for text classification using transformer features.
• Automatic alignment of wordpieces and outputs to linguistic tokens.
• Process multi-sentence documents with intelligent per-sentence prediction.
• Built-in hooks for context-sensitive vectors and similarity.
• Out-of-the-box serialization and model packaging.

🚀 Quickstart

Installing the package from pip will automatically install all dependencies, including PyTorch and spaCy. Make sure you install this package before you install the models. Also note that this package requires Python 3.6+ and the latest version of spaCy, v2.2.1 or above.

pip install spacy-transformers

For GPU installation, find your CUDA version using nvcc --version and add the version in brackets, e.g. spacy-transformers[cuda92] for CUDA9.2 or spacy-transformers[cuda100] for CUDA10.0.

If you are having trouble installing PyTorch, follow the instructions on the official website for your specific operation system and requirements.

We’ve also pre-packaged some of the pretrained models as spaCy model packages. You can either use the spacy download command or download the packages from the model releases.

python -m spacy download en_trf_bertbaseuncased_lg
python -m spacy download de_trf_bertbasecased_lg
python -m spacy download en_trf_xlnetbasecased_lg
python -m spacy download en_trf_robertabase_lg
python -m spacy download en_trf_distilbertbaseuncased_lg

Once the model is installed, you can load it in spaCy like any other model package.

import spacy

nlp = spacy.load("en_trf_bertbaseuncased_lg")
doc = nlp("Apple shares rose on the news. Apple pie is delicious.")
print(doc[0].similarity(doc[7]))
print(doc._.trf_last_hidden_state.shape)

💡 If you’re seeing an error like No module named 'spacy.lang.trf', double-check that spacy-transformers is installed. It needs to be available so it can register its language entry points. Also make sure that you’re running spaCy v2.2.1 or higher.

📖 Usage

⚠️ Important note: This package was previously called spacy-pytorch-transformers and used attributes and pipeline components prefixed with pytt. It’s now called spacy-transformers and uses the prefix trf.

Transfer learning

The main use case for pretrained transformer models is transfer learning. You load in a large generic model pretrained on lots of text, and start training on your smaller dataset with labels specific to your problem. This package has custom pipeline components that make this especially easy. We provide an example component for text categorization. Development of analogous components for other tasks should be quite straight-forward.

The trf_textcat component is based on spaCy’s built-in TextCategorizer and supports using the features assigned by the transformers models, via the trf_tok2vec component. This lets you use a model like BERT to predict contextual token representations, and then learn a text categorizer on top as a task-specific “head”. The API is the same as any other spaCy pipeline:

TRAIN_DATA = [
    ("text1", {"cats": {"POSITIVE": 1.0, "NEGATIVE": 0.0}})
]

import spacy
from spacy.util import minibatch
import random
import torch

is_using_gpu = spacy.prefer_gpu()
if is_using_gpu:
    torch.set_default_tensor_type("torch.cuda.FloatTensor")

nlp = spacy.load("en_trf_bertbaseuncased_lg")
print(nlp.pipe_names) # ["sentencizer", "trf_wordpiecer", "trf_tok2vec"]
textcat = nlp.create_pipe("trf_textcat", config={"exclusive_classes": True})
for label in ("POSITIVE", "NEGATIVE"):
    textcat.add_label(label)
nlp.add_pipe(textcat)

optimizer = nlp.resume_training()
for i in range(10):
    random.shuffle(TRAIN_DATA)
    losses = {}
    for batch in minibatch(TRAIN_DATA, size=8):
        texts, cats = zip(*batch)
        nlp.update(texts, cats, sgd=optimizer, losses=losses)
    print(i, losses)
nlp.to_disk("/bert-textcat")

For a full example, see the examples/train_textcat.py script.

Vectors and similarity

The TransformersTok2Vec component of the model sets custom hooks that override the default behaviour of the .vector attribute and .similarity method of the Token, Span and Doc objects. By default, these usually refer to the word vectors table at nlp.vocab.vectors. Naturally, in the transformer models we’d rather use the doc.tensor attribute, since it holds a much more informative context-sensitive representation.

apple1 = nlp("Apple shares rose on the news.")
apple2 = nlp("Apple sold fewer iPhones this quarter.")
apple3 = nlp("Apple pie is delicious.")
print(apple1[0].similarity(apple2[0]))
print(apple1[0].similarity(apple3[0]))

Serialization

Saving and loading pretrained transformer models and packaging them as spaCy models ✨just works ✨ (at least, it should). The wrapper and components follow spaCy’s API, so when you save and load the nlp object, it…

• Writes the pretrained weights to disk / bytes and loads them back in.
• Adds "lang_factory": "trf" in the meta.json so spaCy knows how to initialize the Language class when you load the model.
• Adds this package and its version to the "requirements" in the meta.json, so when you run spacy package to create an installable Python package it’s automatically added to the setup’s install_requires.

For example, if you’ve trained your own text classifier, you can package it like this:

python -m spacy package /bert-textcat /output
cd /output/en_trf_bertbaseuncased_lg-1.0.0
python setup.py sdist
pip install dist/en_trf_bertbaseuncased_lg-1.0.0.tar.gz

Extension attributes

This wrapper sets the following custom extension attributes on the Doc, Span and Token objects:

The values can be accessed via the ._ attribute. For example:

doc = nlp("This is a text.")
print(doc._.trf_word_pieces_)

Setting up the pipeline

In order to run, the nlp object created using TransformersLanguage requires a few components to run in order: a component that assigns sentence boundaries (e.g. spaCy’s built-in Sentencizer), the TransformersWordPiecer, which assigns the wordpiece tokens and the TransformersTok2Vec, which assigns the token vectors. The trf_name argument defines the name of the pretrained model to use. The from_pretrained methods load the pretrained model via transformers.

from spacy_transformers import TransformersLanguage, TransformersWordPiecer, TransformersTok2Vec

name = "bert-base-uncased"
nlp = TransformersLanguage(trf_name=name, meta={"lang": "en"})
nlp.add_pipe(nlp.create_pipe("sentencizer"))
nlp.add_pipe(TransformersWordPiecer.from_pretrained(nlp.vocab, name))
nlp.add_pipe(TransformersTok2Vec.from_pretrained(nlp.vocab, name))
print(nlp.pipe_names)  # ['sentencizer', 'trf_wordpiecer', 'trf_tok2vec']

You can also use the init_model.py script in the examples.

Loading models from a path

transformers models can also be loaded from a file path instead of just a name. For instance, let’s say you want to use Allen AI’s scibert. First, download the PyTorch model files, unpack them, unpack the weights.tar, rename the bert_config.json to config.json and put everything into one directory. Your directory should now have a pytorch_model.bin, vocab.txt and config.json. Also make sure that your path includes the name of the model. You can then initialize the nlp object like this:

from spacy_transformers import TransformersLanguage, TransformersWordPiecer, TransformersTok2Vec

name = "scibert-scivocab-uncased"
path = "/path/to/scibert-scivocab-uncased"

nlp = TransformersLanguage(trf_name=name, meta={"lang": "en"})
nlp.add_pipe(nlp.create_pipe("sentencizer"))
nlp.add_pipe(TransformersWordPiecer.from_pretrained(nlp.vocab, path))
nlp.add_pipe(TransformersTok2Vec.from_pretrained(nlp.vocab, path))

Tokenization alignment

Transformer models are usually trained on text preprocessed with the “wordpiece” algorithm, which limits the number of distinct token-types the model needs to consider. Wordpiece is convenient for training neural networks, but it doesn’t produce segmentations that match up to any linguistic notion of a “word”. Most rare words will map to multiple wordpiece tokens, and occasionally the alignment will be many-to-many. spacy-transformers calculates this alignment, which you can access at doc._.trf_alignment. It’s a list of length equal to the number of spaCy tokens. Each value in the list is a list of consecutive integers, which are indexes into the wordpieces list.

If you can work on representations that aren’t aligned to actual words, it’s best to use the raw outputs of the transformer, which can be accessed at doc._.trf_last_hidden_state. This variable gives you a tensor with one row per wordpiece token.

If you’re working on token-level tasks such as part-of-speech tagging or spelling correction, you’ll want to work on the token-aligned features, which are stored in the doc.tensor variable.

We’ve taken care to calculate the aligned doc.tensor representation as faithfully as possible, with priority given to avoid information loss. The alignment has been calculated such that doc.tensor.sum(axis=1) == doc._.trf_last_hidden_state.sum(axis=1). To make this work, each row of the doc.tensor (which corresponds to a spaCy token) is set to a weighted sum of the rows of the last_hidden_state tensor that the token is aligned to, where the weighting is proportional to the number of other spaCy tokens aligned to that row. To include the information from the (often important — see Clark et al., 2019) boundary tokens, we imagine that these are also “aligned” to all of the tokens in the sentence.

Batching, padding and per-sentence processing

Transformer models have cubic runtime and memory complexity with respect to sequence length. This means that longer texts need to be divided into sentences in order to achieve reasonable efficiency.

spacy-transformers handles this internally, and requires that a sentence-boundary detection component has been added to the pipeline. We recommend:

sentencizer = nlp.create_pipe("sentencizer")
nlp.add_pipe(sentencizer, first=True)

Internally, the transformer model will predict over sentences, and the resulting tensor features will be reconstructed to produce document-level annotations.

In order to further improve efficiency and reduce memory requirements, spacy-transformers also performs length-based subbatching internally. The subbatching regroups the batched sentences by sequence length, to minimise the amount of padding required. The configuration option words_per_batch controls this behaviour. You can set it to 0 to disable the subbatching, or set it to an integer to require a maximum limit on the number of words (including padding) per subbatch. The default value of 3000 words works reasonably well on a Tesla V100.

Many of the pretrained transformer models have a maximum sequence length. If a sentence is longer than the maximum, it is truncated and the affected ending tokens will receive zeroed vectors.

🎛 API

class TransformersLanguage

A subclass of Language that holds a Transformers pipeline. Transformers pipelines work only slightly differently from spaCy’s default pipelines. Specifically, we introduce a new pipeline component at the start of the pipeline, TransformersTok2Vec. We then modify the nlp.update function to run the TransformersTok2Vec before the other pipeline components, and backprop it after the other components are done.

staticmethod TransformersLanguage.install_extensions

Register the custom extension attributes on the Doc, Span and Token objects. If the extensions have already been registered, spaCy will raise an error. See here for the extension attributes that will be set. You shouldn’t have to call this method yourself – it already runs when you import the package.

method TransformersLanguage.__init__

See Language.__init__. Expects either a trf_name setting in the meta or as a keyword argument, specifying the pretrained model name. This is used to set up the model-specific tokenizer.

method TransformersLanguage.update

Update the models in the pipeline.

class TransformersWordPiecer

spaCy pipeline component to assign Transformers wordpiece tokenization to the Doc, which can then be used by the token vector encoder. Note that this component doesn’t modify spaCy’s tokenization. It only sets extension attributes trf_word_pieces_, trf_word_pieces and trf_alignment (alignment between wordpiece tokens and spaCy tokens).

The component is available as trf_wordpiecer and registered via an entry point, so it can also be created using nlp.create_pipe:

wordpiecer = nlp.create_pipe("trf_wordpiecer")

Config

The component can be configured with the following settings, usually passed in as the **cfg.

classmethod TransformersWordPiecer.from_nlp

Factory to add to Language.factories via entry point.

method TransformersWordPiecer.__init__

Initialize the component.

method TransformersWordPiecer.predict

Run the wordpiece tokenizer on a batch of docs and return the extracted strings.

method TransformersWordPiecer.set_annotations

Assign the extracted tokens and IDs to the Doc objects.

class TransformersTok2Vec

spaCy pipeline component to use transformers models. The component assigns the output of the transformer to extension attributes. We also calculate an alignment between the wordpiece tokens and the spaCy tokenization, so that we can use the last hidden states to set the doc.tensor attribute. When multiple wordpiece tokens align to the same spaCy token, the spaCy token receives the sum of their values.

The component is available as trf_tok2vec and registered via an entry point, so it can also be created using nlp.create_pipe:

tok2vec = nlp.create_pipe("trf_tok2vec")

Config

The component can be configured with the following settings, usually passed in as the **cfg.

classmethod TransformersTok2Vec.from_nlp

Factory to add to Language.factories via entry point.

classmethod TransformersTok2Vec.from_pretrained

Create a TransformersTok2Vec instance using pretrained weights from a transformers model, even if it’s not installed as a spaCy package.

from spacy_transformers import TransformersTok2Vec
from spacy.tokens import Vocab
tok2vec = TransformersTok2Vec.from_pretrained(Vocab(), "bert-base-uncased")

classmethod TransformersTok2Vec.Model

Create an instance of TransformersWrapper, which holds the transformers model.

method TransformersTok2Vec.__init__

Initialize the component.

method TransformersTok2Vec.__call__

Process a Doc and assign the extracted features.

method TransformersTok2Vec.pipe

Process Doc objects as a stream and assign the extracted features.

method TransformersTok2Vec.predict

Run the transformer model on a batch of docs and return the extracted features.

method TransformersTok2Vec.set_annotations

Assign the extracted features to the Doc objects and overwrite the vector and similarity hooks.

class TransformersTextCategorizer

Subclass of spaCy’s built-in TextCategorizer component that supports using the features assigned by the transformers models via the token vector encoder. It requires the TransformersTok2Vec to run before it in the pipeline.

The component is available as trf_textcat and registered via an entry point, so it can also be created using nlp.create_pipe:

textcat = nlp.create_pipe("trf_textcat")

classmethod TransformersTextCategorizer.from_nlp

Factory to add to Language.factories via entry point.

classmethod TransformersTextCategorizer.Model

Create a text classification model using a transformers model for token vector encoding.

dataclass Activations

Dataclass to hold the features produced by transformers.

Entry points

This package exposes several entry points that tell spaCy how to initialize its components. If spacy-transformers and spaCy are installed in the same environment, you’ll be able to run the following and it’ll work as expected:

tok2vec = nlp.create_pipe("trf_tok2vec")

This also means that your custom models can ship a trf_tok2vec component and define "trf_tok2vec" in their pipelines, and spaCy will know how to create those components when you deserialize the model. The following entry points are set: