spaCy/website/docs/api/transformer.md
2020-11-19 14:11:38 +01:00

33 KiB
Raw Blame History

title teaser tag source new api_base_class api_string_name
Transformer Pipeline component for multi-task learning with transformer models class github.com/explosion/spacy-transformers/blob/master/spacy_transformers/pipeline_component.py 3 /api/pipe transformer

Installation

$ 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.

Config and implementation

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

from spacy_transformers import Transformer, DEFAULT_CONFIG

nlp.add_pipe("transformer", config=DEFAULT_CONFIG)
Setting Description
max_batch_items Maximum size of a padded batch. Defaults to 4096. int
set_extra_annotations Function that takes a batch of Doc objects and transformer outputs to set additional annotations on the Doc. The Doc._.trf_data attribute is set prior to calling the callback. Defaults to null_annotation_setter (no additional annotations). Callable[[List[Doc], FullTransformerBatch], None]
model The Thinc Model wrapping the transformer. Defaults to TransformerModel. Model[List[Doc], FullTransformerBatch]
https://github.com/explosion/spacy-transformers/blob/master/spacy_transformers/pipeline_component.py

Transformer.__init__

Example

# 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.v1",
        "name": "bert-base-uncased",
        "tokenizer_config": {"use_fast": True}
    }
}
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.

Name Description
vocab The shared vocabulary. Vocab
model The Thinc Model wrapping the transformer. Usually you will want to use the TransformerModel layer for this. Model[List[Doc], FullTransformerBatch]
set_extra_annotations Function that takes a batch of Doc objects and transformer outputs and stores the annotations on the Doc. The Doc._.trf_data attribute is set prior to calling the callback. By default, no additional annotations are set. Callable[[List[Doc], FullTransformerBatch], None]
keyword-only
name String name of the component instance. Used to add entries to the losses during training. str
max_batch_items Maximum size of a padded batch. Defaults to 128*32. int

Transformer.__call__

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

doc = nlp("This is a sentence.")
trf = nlp.add_pipe("transformer")
# This usually happens under the hood
processed = transformer(doc)
Name Description
doc The document to process. Doc
RETURNS The processed document. Doc

Transformer.pipe

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

trf = nlp.add_pipe("transformer")
for doc in trf.pipe(docs, batch_size=50):
    pass
Name Description
stream A stream of documents. Iterable[Doc]
keyword-only
batch_size The number of documents to buffer. Defaults to 128. int
YIELDS The processed documents in order. Doc

Transformer.initialize

Initialize the component for training and return an Optimizer. get_examples should be a function that returns an iterable of Example objects. 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

trf = nlp.add_pipe("transformer")
trf.initialize(lambda: [], nlp=nlp)
Name Description
get_examples Function that returns gold-standard annotations in the form of Example objects. Callable], Iterable[Example
keyword-only
nlp The current nlp object. Defaults to None. Optional[Language]

Transformer.predict

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

Example

trf = nlp.add_pipe("transformer")
scores = trf.predict([doc1, doc2])
Name Description
docs The documents to predict. Iterable[Doc]
RETURNS The model's prediction for each document.

Transformer.set_annotations

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

trf = nlp.add_pipe("transformer")
scores = trf.predict(docs)
trf.set_annotations(docs, scores)
Name Description
docs The documents to modify. Iterable[Doc]
scores The scores to set, produced by Transformer.predict.

Transformer.update

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

trf = nlp.add_pipe("transformer")
optimizer = nlp.initialize()
losses = trf.update(examples, sgd=optimizer)
Name Description
examples A batch of Example objects. Only the Example.predicted Doc object is used, the reference Doc is ignored. Iterable[Example]
keyword-only
drop The dropout rate. float
set_annotations Whether or not to update the Example objects with the predictions, delegating to set_annotations. bool
sgd An optimizer. Will be created via create_optimizer if not set. Optional[Optimizer]
losses Optional record of the loss during training. Updated using the component name as the key. Optional[Dict[str, float]]
RETURNS The updated losses dictionary. Dict[str, float]

Transformer.create_optimizer

Create an optimizer for the pipeline component.

Example

trf = nlp.add_pipe("transformer")
optimizer = trf.create_optimizer()
Name Description
RETURNS The optimizer. Optimizer

Transformer.use_params

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

Example

trf = nlp.add_pipe("transformer")
with trf.use_params(optimizer.averages):
    trf.to_disk("/best_model")
Name Description
params The parameter values to use in the model. dict

Transformer.to_disk

Serialize the pipe to disk.

Example

trf = nlp.add_pipe("transformer")
trf.to_disk("/path/to/transformer")
Name Description
path A path to a directory, which will be created if it doesn't exist. Paths may be either strings or Path-like objects. Union[str, Path]
keyword-only
exclude String names of serialization fields to exclude. Iterable[str]

Transformer.from_disk

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

Example

trf = nlp.add_pipe("transformer")
trf.from_disk("/path/to/transformer")
Name Description
path A path to a directory. Paths may be either strings or Path-like objects. Union[str, Path]
keyword-only
exclude String names of serialization fields to exclude. Iterable[str]
RETURNS The modified Transformer object. Transformer

Transformer.to_bytes

Example

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

Serialize the pipe to a bytestring.

Name Description
keyword-only
exclude String names of serialization fields to exclude. Iterable[str]
RETURNS The serialized form of the Transformer object. bytes

Transformer.from_bytes

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

Example

trf_bytes = trf.to_bytes()
trf = nlp.add_pipe("transformer")
trf.from_bytes(trf_bytes)
Name Description
bytes_data The data to load from. bytes
keyword-only
exclude String names of serialization fields to exclude. Iterable[str]
RETURNS The Transformer object. Transformer

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

data = trf.to_disk("/path", exclude=["vocab"])
Name Description
vocab The shared Vocab.
cfg The config file. You usually don't want to exclude this.
model The binary model data. You usually don't want to exclude this.

TransformerData

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.

Name Description
tokens A slice of the tokens data produced by the tokenizer. This may have several fields, including the token IDs, the texts and the attention mask. See the transformers.BatchEncoding object for details. dict
tensors 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. List[FloatsXd]
align Alignment from the Doc's tokenization to the wordpieces. This is a ragged array, where align.lengths[i] indicates the number of wordpiece tokens that token i aligns against. The actual indices are provided at align[i].dataXd. Ragged
width The width of the last hidden layer. int

TransformerData.empty

Create an empty TransformerData container.

Name Description
RETURNS The container. TransformerData

FullTransformerBatch

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.

Name Description
spans The batch of input spans. The outer list refers to the Doc objects in the batch, and the inner list are the spans for that Doc. Note that spans are allowed to overlap or exclude tokens, but each Span can only refer to one Doc (by definition). This means that within a Doc, the regions of the output tensors that correspond to each Span may overlap or have gaps, but for each Doc, there is a non-overlapping contiguous slice of the outputs. List[List[Span]]
tokens The output of the tokenizer. transformers.BatchEncoding
tensors The output of the transformer model. List[torch.Tensor]
align Alignment from the spaCy tokenization to the wordpieces. This is a ragged array, where align.lengths[i] indicates the number of wordpiece tokens that token i aligns against. The actual indices are provided at align[i].dataXd. Ragged
doc_data The outputs, split per Doc object. List[TransformerData]

FullTransformerBatch.unsplit_by_doc

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.

Name Description
arrays The split batch of activations. List[List[Floats3d]]
RETURNS The transformer batch. FullTransformerBatch

FullTransformerBatch.split_by_doc

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

Name Description
RETURNS The split batch. List[TransformerData]

Span getters

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

@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
Name Description
docs A batch of Doc objects. Iterable[Doc]
RETURNS The spans to process by the transformer. List[List[Span]]

doc_spans.v1

Example config

[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

Example config

[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.

strided_spans.v1

Example config

[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.

Name Description
window The window size. int
stride The stride size. int

Annotation setters

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

@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
Name Description
docs A batch of Doc objects. List[Doc]
trf_data The transformers data for the batch. FullTransformerBatch

The following built-in functions are available:

Name Description
spacy-transformers.null_annotation_setter.v1 Don't set any additional annotations.

Custom attributes

The component sets the following custom extension attributes:

Name Description
Doc._.trf_data Transformer tokens and outputs for the Doc object. TransformerData