Transformer

Installation

bashCopy

$ pip install -U %%SPACY_PKG_NAME[transformers] %%SPACY_PKG_FLAGS

This component is available via the extension package spacy-transformers. It exposes the component via entry points, so if you have the package installed, using factory = "transformer" in your training config or nlp.add_pipe("transformer") will work out-of-the-box.

This pipeline component lets you use transformer models in your pipeline. It supports all models that are available via the HuggingFace transformers library. Usually you will connect subsequent components to the shared transformer using the TransformerListener layer. This works similarly to spaCy's Tok2Vec component and Tok2VecListener sublayer.

The component assigns the output of the transformer to the Doc's extension attributes. We also calculate an alignment between the word-piece tokens and the spaCy tokenization, so that we can use the last hidden states to set the Doc.tensor attribute. When multiple word-piece tokens align to the same spaCy token, the spaCy token receives the sum of their values. To access the values, you can use the custom Doc._.trf_data attribute. The package also adds the function registries @span_getters and @annotation_setters with several built-in registered functions. For more details, see the usage documentation.

Assigned Attributes {id="assigned-attributes"}

The component sets the following custom extension attribute:

Config and implementation {id="config"}

The default config is defined by the pipeline component factory and describes how the component should be configured. You can override its settings via the config argument on nlp.add_pipe or in your config.cfg for training. See the model architectures documentation for details on the transformer architectures and their arguments and hyperparameters.

Example

pythonCopy

from spacy_transformers import Transformer
from spacy_transformers.pipeline_component import DEFAULT_CONFIG

nlp.add_pipe("transformer", config=DEFAULT_CONFIG["transformer"])

https://github.com/explosion/spacy-transformers/blob/master/spacy_transformers/pipeline_component.py

Transformer.__init__ {id="init",tag="method"}

Example

pythonCopy

# Construction via add_pipe with default model
trf = nlp.add_pipe("transformer")

# Construction via add_pipe with custom config
config = {
    "model": {
        "@architectures": "spacy-transformers.TransformerModel.v3",
        "name": "bert-base-uncased",
        "tokenizer_config": {"use_fast": True},
        "transformer_config": {"output_attentions": True},
        "mixed_precision": True,
        "grad_scaler_config": {"init_scale": 32768}
    }
}
trf = nlp.add_pipe("transformer", config=config)

# Construction from class
from spacy_transformers import Transformer
trf = Transformer(nlp.vocab, model)

Construct a Transformer component. One or more subsequent spaCy components can use the transformer outputs as features in its model, with gradients backpropagated to the single shared weights. The activations from the transformer are saved in the Doc._.trf_data extension attribute. You can also provide a callback to set additional annotations. In your application, you would normally use a shortcut for this and instantiate the component using its string name and nlp.add_pipe.

Transformer.__call__ {id="call",tag="method"}

Apply the pipe to one document. The document is modified in place, and returned. This usually happens under the hood when the nlp object is called on a text and all pipeline components are applied to the Doc in order. Both __call__ and pipe delegate to the predict and set_annotations methods.

Example

pythonCopy

doc = nlp("This is a sentence.")
trf = nlp.add_pipe("transformer")
# This usually happens under the hood
processed = transformer(doc)

Transformer.pipe {id="pipe",tag="method"}

Apply the pipe to a stream of documents. This usually happens under the hood when the nlp object is called on a text and all pipeline components are applied to the Doc in order. Both __call__ and pipe delegate to the predict and set_annotations methods.

Example

pythonCopy

trf = nlp.add_pipe("transformer")
for doc in trf.pipe(docs, batch_size=50):
    pass

Transformer.initialize {id="initialize",tag="method"}

Initialize the component for training and return an Optimizer. get_examples should be a function that returns an iterable of Example objects. At least one example should be supplied. The data examples are used to initialize the model of the component and can either be the full training data or a representative sample. Initialization includes validating the network, inferring missing shapes and setting up the label scheme based on the data. This method is typically called by Language.initialize.

Example

pythonCopy

trf = nlp.add_pipe("transformer")
trf.initialize(lambda: examples, nlp=nlp)

Transformer.predict {id="predict",tag="method"}

Apply the component's model to a batch of Doc objects without modifying them.

Example

pythonCopy

trf = nlp.add_pipe("transformer")
scores = trf.predict([doc1, doc2])

Transformer.set_annotations {id="set_annotations",tag="method"}

Assign the extracted features to the Doc objects. By default, the TransformerData object is written to the Doc._.trf_data attribute. Your set_extra_annotations callback is then called, if provided.

Example

pythonCopy

trf = nlp.add_pipe("transformer")
scores = trf.predict(docs)
trf.set_annotations(docs, scores)

Transformer.update {id="update",tag="method"}

Prepare for an update to the transformer. Like the Tok2Vec component, the Transformer component is unusual in that it does not receive "gold standard" annotations to calculate a weight update. The optimal output of the transformer data is unknown – it's a hidden layer inside the network that is updated by backpropagating from output layers.

The Transformer component therefore does not perform a weight update during its own update method. Instead, it runs its transformer model and communicates the output and the backpropagation callback to any downstream components that have been connected to it via the TransformerListener sublayer. If there are multiple listeners, the last layer will actually backprop to the transformer and call the optimizer, while the others simply increment the gradients.

Example

pythonCopy

trf = nlp.add_pipe("transformer")
optimizer = nlp.initialize()
losses = trf.update(examples, sgd=optimizer)

Transformer.create_optimizer {id="create_optimizer",tag="method"}

Create an optimizer for the pipeline component.

Example

pythonCopy

trf = nlp.add_pipe("transformer")
optimizer = trf.create_optimizer()

Transformer.use_params {id="use_params",tag="method, contextmanager"}

Modify the pipe's model to use the given parameter values. At the end of the context, the original parameters are restored.

Example

pythonCopy

trf = nlp.add_pipe("transformer")
with trf.use_params(optimizer.averages):
    trf.to_disk("/best_model")

Transformer.to_disk {id="to_disk",tag="method"}

Serialize the pipe to disk.

Example

pythonCopy

trf = nlp.add_pipe("transformer")
trf.to_disk("/path/to/transformer")

Transformer.from_disk {id="from_disk",tag="method"}

Load the pipe from disk. Modifies the object in place and returns it.

Example

pythonCopy

trf = nlp.add_pipe("transformer")
trf.from_disk("/path/to/transformer")

Transformer.to_bytes {id="to_bytes",tag="method"}

Example

pythonCopy

trf = nlp.add_pipe("transformer")
trf_bytes = trf.to_bytes()

Serialize the pipe to a bytestring.

Transformer.from_bytes {id="from_bytes",tag="method"}

Load the pipe from a bytestring. Modifies the object in place and returns it.

Example

pythonCopy

trf_bytes = trf.to_bytes()
trf = nlp.add_pipe("transformer")
trf.from_bytes(trf_bytes)

Serialization fields {id="serialization-fields"}

During serialization, spaCy will export several data fields used to restore different aspects of the object. If needed, you can exclude them from serialization by passing in the string names via the exclude argument.

Example

pythonCopy

data = trf.to_disk("/path", exclude=["vocab"])

TransformerData {id="transformerdata",tag="dataclass"}

Transformer tokens and outputs for one Doc object. The transformer models return tensors that refer to a whole padded batch of documents. These tensors are wrapped into the FullTransformerBatch object. The FullTransformerBatch then splits out the per-document data, which is handled by this class. Instances of this class are typically assigned to the Doc._.trf_data extension attribute.

Example

pythonCopy

# Get the last hidden layer output for "is" (token index 1)
doc = nlp("This is a text.")
indices = doc._.trf_data.align[1].data.flatten()
last_hidden_state = doc._.trf_data.model_output.last_hidden_state
dim = last_hidden_state.shape[-1]
tensors = last_hidden_state.reshape(-1, dim)[indices]

TransformerData.empty {id="transformerdata-empty",tag="classmethod"}

Create an empty TransformerData container.

In spacy-transformers v1.0, the model output is stored in TransformerData.tensors as List[Union[FloatsXd]] and only includes the activations for the Doc from the transformer. Usually the last tensor that is 3-dimensional will be the most important, as that will provide the final hidden state. Generally activations that are 2-dimensional will be attention weights. Details of this variable will differ depending on the underlying transformer model.

FullTransformerBatch {id="fulltransformerbatch",tag="dataclass"}

Holds a batch of input and output objects for a transformer model. The data can then be split to a list of TransformerData objects to associate the outputs to each Doc in the batch.

FullTransformerBatch.unsplit_by_doc {id="fulltransformerbatch-unsplit_by_doc",tag="method"}

Return a new FullTransformerBatch from a split batch of activations, using the current object's spans, tokens and alignment. This is used during the backward pass, in order to construct the gradients to pass back into the transformer model.

FullTransformerBatch.split_by_doc {id="fulltransformerbatch-split_by_doc",tag="method"}

Split a TransformerData object that represents a batch into a list with one TransformerData per Doc.

In spacy-transformers v1.0, the model output is stored in FullTransformerBatch.tensors as List[torch.Tensor].

Span getters {id="span_getters",source="github.com/explosion/spacy-transformers/blob/master/spacy_transformers/span_getters.py"}

Span getters are functions that take a batch of Doc objects and return a lists of Span objects for each doc to be processed by the transformer. This is used to manage long documents by cutting them into smaller sequences before running the transformer. The spans are allowed to overlap, and you can also omit sections of the Doc if they are not relevant.

Span getters can be referenced in the [components.transformer.model.get_spans] block of the config to customize the sequences processed by the transformer. You can also register custom span getters using the @spacy.registry.span_getters decorator.

Example

pythonCopy

@spacy.registry.span_getters("custom_sent_spans")
def configure_get_sent_spans() -> Callable:
    def get_sent_spans(docs: Iterable[Doc]) -> List[List[Span]]:
        return [list(doc.sents) for doc in docs]

    return get_sent_spans

doc_spans.v1 {id="doc_spans",tag="registered function"}

Example config

iniCopy

[transformer.model.get_spans]
@span_getters = "spacy-transformers.doc_spans.v1"

Create a span getter that uses the whole document as its spans. This is the best approach if your Doc objects already refer to relatively short texts.

sent_spans.v1 {id="sent_spans",tag="registered function"}

Example config

iniCopy

[transformer.model.get_spans]
@span_getters = "spacy-transformers.sent_spans.v1"

Create a span getter that uses sentence boundary markers to extract the spans. This requires sentence boundaries to be set (e.g. by the Sentencizer), and may result in somewhat uneven batches, depending on the sentence lengths. However, it does provide the transformer with more meaningful windows to attend over.

To set sentence boundaries with the sentencizer during training, add a sentencizer to the beginning of the pipeline and include it in [training.annotating_components] to have it set the sentence boundaries before the transformer component runs.

strided_spans.v1 {id="strided_spans",tag="registered function"}

Example config

iniCopy

[transformer.model.get_spans]
@span_getters = "spacy-transformers.strided_spans.v1"
window = 128
stride = 96

Create a span getter for strided spans. If you set the window and stride to the same value, the spans will cover each token once. Setting stride lower than window will allow for an overlap, so that some tokens are counted twice. This can be desirable, because it allows all tokens to have both a left and right context.

Annotation setters {id="annotation_setters",tag="registered functions",source="github.com/explosion/spacy-transformers/blob/master/spacy_transformers/annotation_setters.py"}

Annotation setters are functions that take a batch of Doc objects and a FullTransformerBatch and can set additional annotations on the Doc, e.g. to set custom or built-in attributes. You can register custom annotation setters using the @registry.annotation_setters decorator.

Example

pythonCopy

@registry.annotation_setters("spacy-transformers.null_annotation_setter.v1")
def configure_null_annotation_setter() -> Callable:
    def setter(docs: List[Doc], trf_data: FullTransformerBatch) -> None:
        pass

    return setter

The following built-in functions are available: