CuratedTransformer

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

This pipeline component lets you use a curated set of transformer models in your pipeline. spaCy Curated Transformers currently supports the following model types:

• ALBERT
• BERT
• CamemBERT
• RoBERTa
• XLM-RoBERT

If you want to use another type of model, use spacy-transformers, which allows you to use all Hugging Face transformer models with spaCy.

You will usually connect downstream components to a shared Curated Transformer pipe using one of the Curated Transformer listener layers. This works similarly to spaCy's Tok2Vec, and the Tok2VecListener sublayer. The component assigns the output of the transformer to the Doc's extension attributes. To access the values, you can use the custom Doc._.trf_data attribute.

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 curated transformer architectures and their arguments and hyperparameters.

Example

pythonCopy

from spacy_curated_transformers.pipeline.transformer import DEFAULT_CONFIG

nlp.add_pipe("curated_transformer", config=DEFAULT_CONFIG)

https://github.com/explosion/spacy-curated-transformers/blob/main/spacy_curated_transformers/pipeline/transformer.py

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

Example

pythonCopy

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

# Construction via add_pipe with custom config
config = {
    "model": {
        "@architectures": "spacy-curated-transformers.XlmrTransformer.v1",
        "vocab_size": 250002,
        "num_hidden_layers": 12,
        "hidden_width": 768,
        "piece_encoder": {
            "@architectures": "spacy-curated-transformers.XlmrSentencepieceEncoder.v1"
        }
    }
}
trf = nlp.add_pipe("curated_transformer", config=config)

# Construction from class
from spacy_curated_transformers import CuratedTransformer
trf = CuratedTransformer(nlp.vocab, model)

Construct a CuratedTransformer 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.

CuratedTransformer.__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("curated_transformer")
# This usually happens under the hood
processed = trf(doc)

CuratedTransformer.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("curated_transformer")
for doc in trf.pipe(docs, batch_size=50):
    pass

CuratedTransformer.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("curated_transformer")
trf.initialize(lambda: examples, nlp=nlp)

CuratedTransformer.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("curated_transformer")
scores = trf.predict([doc1, doc2])

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

Assign the extracted features to the Doc objects. By default, the DocTransformerOutput 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("curated_transformer")
scores = trf.predict(docs)
trf.set_annotations(docs, scores)

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

Prepare for an update to the transformer.

Like the Tok2Vec component, the CuratedTransformer 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 CuratedTransformer 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 transformer listener 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("curated_transformer")
optimizer = nlp.initialize()
losses = trf.update(examples, sgd=optimizer)

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

Create an optimizer for the pipeline component.

Example

pythonCopy

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

CuratedTransformer.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("curated_transformer")
with trf.use_params(optimizer.averages):
    trf.to_disk("/best_model")

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

Serialize the pipe to disk.

Example

pythonCopy

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

CuratedTransformer.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("curated_transformer")
trf.from_disk("/path/to/transformer")

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

Example

pythonCopy

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

Serialize the pipe to a bytestring.

CuratedTransformer.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("curated_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"])

DocTransformerOutput {id="doctransformeroutput",tag="dataclass"}

Curated Transformer outputs for one Doc object. Stores the dense representations generated by the transformer for each piece identifier. Piece identifiers are grouped by token. 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.")
tensors = doc._.trf_data.last_hidden_layer_state[1]

DocTransformerOutput.embedding_layer {id="doctransformeroutput-embeddinglayer",tag="property"}

Return the output of the transformer's embedding layer or None if last_layer_only is True.

DocTransformerOutput.last_hidden_layer_state {id="doctransformeroutput-lasthiddenlayerstate",tag="property"}

Return the output of the transformer's last hidden layer.

DocTransformerOutput.all_hidden_layer_states {id="doctransformeroutput-allhiddenlayerstates",tag="property"}

Return the outputs of all transformer layers (excluding the embedding layer).

DocTransformerOutput.num_outputs {id="doctransformeroutput-numoutputs",tag="property"}

Return the number of layer outputs stored in the DocTransformerOutput instance (including the embedding layer).

Span Getters {id="span_getters",source="github.com/explosion/spacy-transformers/blob/master/spacy_curated_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.with_spans] block of the config to customize the sequences processed by the transformer.

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

Example config

iniCopy

[transformer.model.with_spans]
@architectures = "spacy-curated-transformers.WithStridedSpans.v1"
stride = 96
window = 128

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.

Model Loaders

Curated Transformer models are constructed with default hyperparameters and randomized weights when the pipeline is created. To load the weights of an existing pre-trained model into the pipeline, one of the following loader callbacks can be used. The pre-trained model must have the same hyperparameters as the model used by the pipeline.

HFTransformerEncoderLoader.v1 {id="hf_trfencoder_loader",tag="registered_function"}

Construct a callback that initializes a supported transformer model with weights from a corresponding HuggingFace model.

PyTorchCheckpointLoader.v1 {id="pytorch_checkpoint_loader",tag="registered_function"}

Construct a callback that initializes a supported transformer model with weights from a PyTorch checkpoint.

Tokenizer Loaders

Curated Transformer models must be paired with a matching tokenizer (piece encoder) model in a spaCy pipeline. As with the transformer models, tokenizers are constructed with an empty vocabulary during pipeline creation - They need to be initialized with an appropriate loader before use in training/inference.

ByteBPELoader.v1 {id="bytebpe_loader",tag="registered_function"}

Construct a callback that initializes a Byte-BPE piece encoder model.

CharEncoderLoader.v1 {id="charencoder_loader",tag="registered_function"}

Construct a callback that initializes a character piece encoder model.

HFPieceEncoderLoader.v1 {id="hf_pieceencoder_loader",tag="registered_function"}

Construct a callback that initializes a HuggingFace piece encoder model. Used in conjunction with the HuggingFace model loader.

SentencepieceLoader.v1 {id="sentencepiece_loader",tag="registered_function"}

Construct a callback that initializes a SentencePiece piece encoder model.

WordpieceLoader.v1 {id="wordpiece_loader",tag="registered_function"}

Construct a callback that initializes a WordPiece piece encoder model.

Callbacks

gradual_transformer_unfreezing.v1 {id="gradual_transformer_unfreezing",tag="registered_function"}

Construct a callback that can be used to gradually unfreeze the weights of one or more Transformer components during training. This can be used to prevent catastrophic forgetting during fine-tuning.