Overview

Model Architectures

Pre-defined model architectures included with the core library

A model architecture is a function that wires up a Model instance, which you can then use in a pipeline component or as a layer of a larger network. This page documents spaCy’s built-in architectures that are used for different NLP tasks. All trainable built-in components expect a model argument defined in the config and document their the default architecture. Custom architectures can be registered using the @spacy.registry.architectures decorator and used as part of the training config. Also see the usage documentation on layers and model architectures.

Tok2Vec architectures

spacy.Tok2Vec.v2

Construct a tok2vec model out of two subnetworks: one for embedding and one for encoding. See the “Embed, Encode, Attend, Predict” blog post for background.

NameDescription
embedEmbed tokens into context-independent word vector representations. For example, CharacterEmbed or MultiHashEmbed. Model[List[Doc], List[Floats2d]]
encodeEncode context into the embeddings, using an architecture such as a CNN, BiLSTM or transformer. For example, MaxoutWindowEncoder. Model[List[Floats2d], List[Floats2d]]

spacy.HashEmbedCNN.v1

Build spaCy’s “standard” tok2vec layer. This layer is defined by a MultiHashEmbed embedding layer that uses subword features, and a MaxoutWindowEncoder encoding layer consisting of a CNN and a layer-normalized maxout activation function.

NameDescription
widthThe width of the input and output. These are required to be the same, so that residual connections can be used. Recommended values are 96, 128 or 300. int
depthThe number of convolutional layers to use. Recommended values are between 2 and 8. int
embed_sizeThe number of rows in the hash embedding tables. This can be surprisingly small, due to the use of the hash embeddings. Recommended values are between 2000 and 10000. int
window_sizeThe number of tokens on either side to concatenate during the convolutions. The receptive field of the CNN will be depth * (window_size * 2 + 1), so a 4-layer network with a window size of 2 will be sensitive to 17 words at a time. Recommended value is 1. int
maxout_piecesThe number of pieces to use in the maxout non-linearity. If 1, the Mish non-linearity is used instead. Recommended values are 1-3. int
subword_featuresWhether to also embed subword features, specifically the prefix, suffix and word shape. This is recommended for alphabetic languages like English, but not if single-character tokens are used for a language such as Chinese. bool
pretrained_vectorsWhether to also use static vectors. bool

spacy.Tok2VecListener.v1

A listener is used as a sublayer within a component such as a DependencyParser, EntityRecognizeror TextCategorizer. Usually you’ll have multiple listeners connecting to a single upstream Tok2Vec component that’s earlier in the pipeline. The listener layers act as proxies, passing the predictions from the Tok2Vec component into downstream components, and communicating gradients back upstream.

Instead of defining its own Tok2Vec instance, a model architecture like Tagger can define a listener as its tok2vec argument that connects to the shared tok2vec component in the pipeline.

NameDescription
widthThe width of the vectors produced by the “upstream” Tok2Vec component. int
upstreamA string to identify the “upstream” Tok2Vec component to communicate with. By default, the upstream name is the wildcard string "*", but you could also specify the name of the Tok2Vec component. You’ll almost never have multiple upstream Tok2Vec components, so the wildcard string will almost always be fine. str

spacy.MultiHashEmbed.v1

Construct an embedding layer that separately embeds a number of lexical attributes using hash embedding, concatenates the results, and passes it through a feed-forward subnetwork to build a mixed representation. The features used can be configured with the attrs argument. The suggested attributes are NORM, PREFIX, SUFFIX and SHAPE. This lets the model take into account some subword information, without construction a fully character-based representation. If pretrained vectors are available, they can be included in the representation as well, with the vectors table will be kept static (i.e. it’s not updated).

NameDescription
widthThe output width. Also used as the width of the embedding tables. Recommended values are between 64 and 300. If static vectors are included, a learned linear layer is used to map the vectors to the specified width before concatenating it with the other embedding outputs. A single maxout layer is then used to reduce the concatenated vectors to the final width. int
attrsThe token attributes to embed. A separate embedding table will be constructed for each attribute. List[Union[int, str]]
rowsThe number of rows for each embedding tables. Can be low, due to the hashing trick. Recommended values are between 1000 and 10000. The layer needs surprisingly few rows, due to its use of the hashing trick. Generally between 2000 and 10000 rows is sufficient, even for very large vocabularies. A number of rows must be specified for each table, so the rows list must be of the same length as the attrs parameter. List[int]
include_static_vectorsWhether to also use static word vectors. Requires a vectors table to be loaded in the Doc objects’ vocab. bool

spacy.CharacterEmbed.v1

Construct an embedded representation based on character embeddings, using a feed-forward network. A fixed number of UTF-8 byte characters are used for each word, taken from the beginning and end of the word equally. Padding is used in the center for words that are too short.

For instance, let’s say nC=4, and the word is “jumping”. The characters used will be "jung" (two from the start, two from the end). If we had nC=8, the characters would be "jumpping": 4 from the start, 4 from the end. This ensures that the final character is always in the last position, instead of being in an arbitrary position depending on the word length.

The characters are embedded in a embedding table with a given number of rows, and the vectors concatenated. A hash-embedded vector of the NORM of the word is also concatenated on, and the result is then passed through a feed-forward network to construct a single vector to represent the information.

NameDescription
widthThe width of the output vector and the NORM hash embedding. int
rowsThe number of rows in the NORM hash embedding table. int
nMThe dimensionality of the character embeddings. Recommended values are between 16 and 64. int
nCThe number of UTF-8 bytes to embed per word. Recommended values are between 3 and 8, although it may depend on the length of words in the language. int

spacy.MaxoutWindowEncoder.v2

Encode context using convolutions with maxout activation, layer normalization and residual connections.

NameDescription
widthThe input and output width. These are required to be the same, to allow residual connections. This value will be determined by the width of the inputs. Recommended values are between 64 and 300. int
window_sizeThe number of words to concatenate around each token to construct the convolution. Recommended value is 1. int
maxout_piecesThe number of maxout pieces to use. Recommended values are 2 or 3. int
depthThe number of convolutional layers. Recommended value is 4. int

spacy.MishWindowEncoder.v2

Encode context using convolutions with Mish activation, layer normalization and residual connections.

NameDescription
widthThe input and output width. These are required to be the same, to allow residual connections. This value will be determined by the width of the inputs. Recommended values are between 64 and 300. int
window_sizeThe number of words to concatenate around each token to construct the convolution. Recommended value is 1. int
depthThe number of convolutional layers. Recommended value is 4. int

spacy.TorchBiLSTMEncoder.v1

Encode context using bidirectional LSTM layers. Requires PyTorch.

NameDescription
widthThe input and output width. These are required to be the same, to allow residual connections. This value will be determined by the width of the inputs. Recommended values are between 64 and 300. int
depthThe number of recurrent layers, for instance depth=2 results in stacking two LSTMs together. int
dropoutCreates a Dropout layer on the outputs of each LSTM layer except the last layer. Set to 0.0 to disable this functionality. float

spacy.StaticVectors.v1

Embed Doc objects with their vocab’s vectors table, applying a learned linear projection to control the dimensionality. See the documentation on static vectors for details.

Name Description
nOThe output width of the layer, after the linear projection. Optional[int]
nMThe width of the static vectors. Optional[int]
dropoutOptional dropout rate. If set, it’s applied per dimension over the whole batch. Defaults to None. Optional[float]
init_WThe initialization function. Defaults to glorot_uniform_init. Callable[[Ops, Tuple[int,]]], FloatsXd]
key_attrDefaults to "ORTH". str

spacy.FeatureExtractor.v1

Extract arrays of input features from Doc objects. Expects a list of feature names to extract, which should refer to token attributes.

Name Description
columnsThe token attributes to extract. List[Union[int, str]]

Transformer architectures

The following architectures are provided by the package spacy-transformers. See the usage documentation for how to integrate the architectures into your training config.

spacy-transformers.TransformerModel.v1

Load and wrap a transformer model from the HuggingFace transformers library. You can use any transformer that has pretrained weights and a PyTorch implementation. The name variable is passed through to the underlying library, so it can be either a string or a path. If it’s a string, the pretrained weights will be downloaded via the transformers library if they are not already available locally.

In order to support longer documents, the TransformerModel layer allows you to pass in a get_spans function that will divide up the Doc objects before passing them through the transformer. Your spans are allowed to overlap or exclude tokens. This layer is usually used directly by the Transformer component, which allows you to share the transformer weights across your pipeline. For a layer that’s configured for use in other components, see Tok2VecTransformer.

NameDescription
nameAny model name that can be loaded by transformers.AutoModel. str
get_spansFunction that takes a batch of Doc object and returns lists of Span objects to process by the transformer. See here for built-in options and examples. Callable[[List[Doc]], List[Span]]
tokenizer_configTokenizer settings passed to transformers.AutoTokenizer. Dict[str, Any]

spacy-transformers.TransformerListener.v1

Create a TransformerListener layer, which will connect to a Transformer component earlier in the pipeline. The layer takes a list of Doc objects as input, and produces a list of 2-dimensional arrays as output, with each array having one row per token. Most spaCy models expect a sublayer with this signature, making it easy to connect them to a transformer model via this sublayer. Transformer models usually operate over wordpieces, which usually don’t align one-to-one against spaCy tokens. The layer therefore requires a reduction operation in order to calculate a single token vector given zero or more wordpiece vectors.

NameDescription
poolingA reduction layer used to calculate the token vectors based on zero or more wordpiece vectors. If in doubt, mean pooling (see reduce_mean) is usually a good choice. Model[Ragged, Floats2d]
grad_factorReweight gradients from the component before passing them upstream. You can set this to 0 to “freeze” the transformer weights with respect to the component, or use it to make some components more significant than others. Leaving it at 1.0 is usually fine. float
upstreamA string to identify the “upstream” Transformer component to communicate with. By default, the upstream name is the wildcard string "*", but you could also specify the name of the Transformer component. You’ll almost never have multiple upstream Transformer components, so the wildcard string will almost always be fine. str

spacy-transformers.Tok2VecTransformer.v1

Use a transformer as a Tok2Vec layer directly. This does not allow multiple components to share the transformer weights and does not allow the transformer to set annotations into the Doc object, but it’s a simpler solution if you only need the transformer within one component.

NameDescription
get_spansFunction that takes a batch of Doc object and returns lists of Span objects to process by the transformer. See here for built-in options and examples. Callable[[List[Doc]], List[Span]]
tokenizer_configTokenizer settings passed to transformers.AutoTokenizer. Dict[str, Any]
poolingA reduction layer used to calculate the token vectors based on zero or more wordpiece vectors. If in doubt, mean pooling (see reduce_mean) is usually a good choice. Model[Ragged, Floats2d]
grad_factorReweight gradients from the component before passing them upstream. You can set this to 0 to “freeze” the transformer weights with respect to the component, or use it to make some components more significant than others. Leaving it at 1.0 is usually fine. float

Pretraining architectures

The spacy pretrain command lets you initialize a Tok2Vec layer in your pipeline with information from raw text. To this end, additional layers are added to build a network for a temporary task that forces the Tok2Vec layer to learn something about sentence structure and word cooccurrence statistics. Two pretraining objectives are available, both of which are variants of the cloze task Devlin et al. (2018) introduced for BERT.

For more information, see the section on pretraining.

spacy.PretrainVectors.v1

Predict the word’s vector from a static embeddings table as pretraining objective for a Tok2Vec layer. To use this objective, make sure that the initialize.vectors section in the config refers to a model with static vectors.

NameDescription
maxout_piecesThe number of maxout pieces to use. Recommended values are 2 or 3. int
hidden_sizeSize of the hidden layer of the model. int
lossThe loss function can be either “cosine” or “L2”. We typically recommend to use “cosine”. ~str

spacy.PretrainCharacters.v1

Predict some number of leading and trailing UTF-8 bytes as pretraining objective for a Tok2Vec layer.

NameDescription
maxout_piecesThe number of maxout pieces to use. Recommended values are 2 or 3. int
hidden_sizeSize of the hidden layer of the model. int
n_charactersThe window of characters - e.g. if n_characters = 2, the model will try to predict the first two and last two characters of the word. int

Parser & NER architectures

spacy.TransitionBasedParser.v2

Build a transition-based parser model. Can apply to NER or dependency parsing. Transition-based parsing is an approach to structured prediction where the task of predicting the structure is mapped to a series of state transitions. You might find this tutorial helpful for background information. The neural network state prediction model consists of either two or three subnetworks:

  • tok2vec: Map each token into a vector representation. This subnetwork is run once for each batch.
  • lower: Construct a feature-specific vector for each (token, feature) pair. This is also run once for each batch. Constructing the state representation is then simply a matter of summing the component features and applying the non-linearity.
  • upper (optional): A feed-forward network that predicts scores from the state representation. If not present, the output from the lower model is used as action scores directly.
NameDescription
tok2vecSubnetwork to map tokens into vector representations. Model[List[Doc], List[Floats2d]]
state_typeWhich task to extract features for. Possible values are “ner” and “parser”. str
extra_state_tokensWhether to use an expanded feature set when extracting the state tokens. Slightly slower, but sometimes improves accuracy slightly. Defaults to False. bool
hidden_widthThe width of the hidden layer. int
maxout_piecesHow many pieces to use in the state prediction layer. Recommended values are 1, 2 or 3. If 1, the maxout non-linearity is replaced with a Relu non-linearity if use_upper is True, and no non-linearity if False. int
use_upperWhether to use an additional hidden layer after the state vector in order to predict the action scores. It is recommended to set this to False for large pretrained models such as transformers, and True for smaller networks. The upper layer is computed on CPU, which becomes a bottleneck on larger GPU-based models, where it’s also less necessary. bool
nOThe number of actions the model will predict between. Usually inferred from data at the beginning of training, or loaded from disk. int

Tagging architectures

spacy.Tagger.v1

Build a tagger model, using a provided token-to-vector component. The tagger model simply adds a linear layer with softmax activation to predict scores given the token vectors.

NameDescription
tok2vecSubnetwork to map tokens into vector representations. Model[List[Doc], List[Floats2d]]
nOThe number of tags to output. Inferred from the data if None. Optional[int]

Text classification architectures

A text classification architecture needs to take a Doc as input, and produce a score for each potential label class. Textcat challenges can be binary (e.g. sentiment analysis) or involve multiple possible labels. Multi-label challenges can either have mutually exclusive labels (each example has exactly one label), or multiple labels may be applicable at the same time.

As the properties of text classification problems can vary widely, we provide several different built-in architectures. It is recommended to experiment with different architectures and settings to determine what works best on your specific data and challenge.

spacy.TextCatEnsemble.v2

Stacked ensemble of a linear bag-of-words model and a neural network model. The neural network is built upon a Tok2Vec layer and uses attention. The setting for whether or not this model should cater for multi-label classification, is taken from the linear model, where it is stored in model.attrs["multi_label"].

NameDescription
linear_modelThe linear bag-of-words model. Model[List[Doc], Floats2d]
tok2vecThe tok2vec layer to build the neural network upon. Model[List[Doc], List[Floats2d]]
nOOutput dimension, determined by the number of different labels. If not set, the TextCategorizer component will set it when initialize is called. Optional[int]

TextCatEnsemble.v1 was functionally similar, but used an internal tok2vec instead of taking it as argument:

NameDescription
exclusive_classesWhether or not categories are mutually exclusive. bool
pretrained_vectorsWhether or not pretrained vectors will be used in addition to the feature vectors. bool
widthOutput dimension of the feature encoding step. int
embed_sizeInput dimension of the feature encoding step. int
conv_depthDepth of the tok2vec layer. int
window_sizeThe number of contextual vectors to concatenate from the left and from the right. int
ngram_sizeDetermines the maximum length of the n-grams in the BOW model. For instance, ngram_size=3would give unigram, trigram and bigram features. int
dropoutThe dropout rate. float
nOOutput dimension, determined by the number of different labels. If not set, the TextCategorizer component will set it when initialize is called. Optional[int]

spacy.TextCatCNN.v1

A neural network model where token vectors are calculated using a CNN. The vectors are mean pooled and used as features in a feed-forward network. This architecture is usually less accurate than the ensemble, but runs faster.

NameDescription
exclusive_classesWhether or not categories are mutually exclusive. bool
tok2vecThe tok2vec layer of the model. Model
nOOutput dimension, determined by the number of different labels. If not set, the TextCategorizer component will set it when initialize is called. Optional[int]

spacy.TextCatBOW.v1

An n-gram “bag-of-words” model. This architecture should run much faster than the others, but may not be as accurate, especially if texts are short.

NameDescription
exclusive_classesWhether or not categories are mutually exclusive. bool
ngram_sizeDetermines the maximum length of the n-grams in the BOW model. For instance, ngram_size=3 would give unigram, trigram and bigram features. int
no_output_layerWhether or not to add an output layer to the model (Softmax activation if exclusive_classes is True, else Logistic). bool
nOOutput dimension, determined by the number of different labels. If not set, the TextCategorizer component will set it when initialize is called. Optional[int]

Entity linking architectures

An EntityLinker component disambiguates textual mentions (tagged as named entities) to unique identifiers, grounding the named entities into the “real world”. This requires 3 main components:

  • A KnowledgeBase (KB) holding the unique identifiers, potential synonyms and prior probabilities.
  • A candidate generation step to produce a set of likely identifiers, given a certain textual mention.
  • A machine learning Model that picks the most plausible ID from the set of candidates.

spacy.EntityLinker.v1

The EntityLinker model architecture is a Thinc Model with a Linear output layer.

NameDescription
tok2vecThe tok2vec layer of the model. Model
nOOutput dimension, determined by the length of the vectors encoding each entity in the KB. If the nO dimension is not set, the entity linking component will set it when initialize is called. Optional[int]

spacy.EmptyKB.v1

A function that creates an empty KnowledgeBase from a Vocab instance. This is the default when a new entity linker component is created.

NameDescription
entity_vector_lengthThe length of the vectors encoding each entity in the KB. Defaults to 64. int

spacy.KBFromFile.v1

A function that reads an existing KnowledgeBase from file.

NameDescription
kb_pathThe location of the KB that was stored to file. Path

spacy.CandidateGenerator.v1

A function that takes as input a KnowledgeBase and a Span object denoting a named entity, and returns a list of plausible Candidate objects. The default CandidateGenerator simply uses the text of a mention to find its potential aliases in the KnowledgeBase. Note that this function is case-dependent.