spaCy/website/docs/models/index.md
2022-05-23 09:15:51 +02:00

10 KiB
Raw Blame History

title teaser menu
Trained Models & Pipelines Downloadable trained pipelines and weights for spaCy
Quickstart
quickstart
Conventions
conventions
Pipeline Design
design

Quickstart

📖 Installation and usage

For more details on how to use trained pipelines with spaCy, see the usage guide.

import QuickstartModels from 'widgets/quickstart-models.js'

Package naming conventions

In general, spaCy expects all pipeline packages to follow the naming convention of [lang]\_[name]. For spaCy's pipelines, we also chose to divide the name into three components:

  1. Type: Capabilities (e.g. core for general-purpose pipeline with tagging, parsing, lemmatization and named entity recognition, or dep for only tagging, parsing and lemmatization).

  2. Genre: Type of text the pipeline is trained on, e.g. web or news.

  3. Size: Package size indicator, sm, md, lg or trf.

    sm and trf pipelines have no static word vectors.

    For pipelines with default vectors, md has a reduced word vector table with 20k unique vectors for ~500k words and lg has a large word vector table with ~500k entries.

    For pipelines with floret vectors, md vector tables have 50k entries and lg vector tables have 200k entries.

For example, en_core_web_sm is a small English pipeline trained on written web text (blogs, news, comments), that includes vocabulary, syntax and entities.

Package versioning

Additionally, the pipeline package versioning reflects both the compatibility with spaCy, as well as the model version. A package version a.b.c translates to:

  • a: spaCy major version. For example, 2 for spaCy v2.x.
  • b: spaCy minor version. For example, 3 for spaCy v2.3.x.
  • c: Model version. Different model config: e.g. from being trained on different data, with different parameters, for different numbers of iterations, with different vectors, etc.

For a detailed compatibility overview, see the compatibility.json. This is also the source of spaCy's internal compatibility check, performed when you run the download command.

Trained pipeline design

The spaCy v3 trained pipelines are designed to be efficient and configurable. For example, multiple components can share a common "token-to-vector" model and it's easy to swap out or disable the lemmatizer. The pipelines are designed to be efficient in terms of speed and size and work well when the pipeline is run in full.

When modifying a trained pipeline, it's important to understand how the components depend on each other. Unlike spaCy v2, where the tagger, parser and ner components were all independent, some v3 components depend on earlier components in the pipeline. As a result, disabling or reordering components can affect the annotation quality or lead to warnings and errors.

Main changes from spaCy v2 models:

  • The Tok2Vec component may be a separate, shared component. A component like a tagger or parser can listen to an earlier tok2vec or transformer rather than having its own separate tok2vec layer.
  • Rule-based exceptions move from individual components to the attribute_ruler. Lemma and POS exceptions move from the tokenizer exceptions to the attribute ruler and the tag map and morph rules move from the tagger to the attribute ruler.
  • The lemmatizer tables and processing move from the vocab and tagger to a separate lemmatizer component.

CNN/CPU pipeline design

Components and their dependencies in the CNN pipelines

In the sm/md/lg models:

  • The tagger, morphologizer and parser components listen to the tok2vec component. If the lemmatizer is trainable (v3.3+), lemmatizer also listens to tok2vec.
  • The attribute_ruler maps token.tag to token.pos if there is no morphologizer. The attribute_ruler additionally makes sure whitespace is tagged consistently and copies token.pos to token.tag if there is no tagger. For English, the attribute ruler can improve its mapping from token.tag to token.pos if dependency parses from a parser are present, but the parser is not required.
  • The lemmatizer component for many languages requires token.pos annotation from either tagger+attribute_ruler or morphologizer.
  • The ner component is independent with its own internal tok2vec layer.

CNN/CPU pipelines with floret vectors

The Finnish, Korean and Swedish md and lg pipelines use floret vectors instead of default vectors. If you're running a trained pipeline on texts and working with Doc objects, you shouldn't notice any difference with floret vectors. With floret vectors no tokens are out-of-vocabulary, so Token.is_oov will return False for all tokens.

If you access vectors directly for similarity comparisons, there are a few differences because floret vectors don't include a fixed word list like the vector keys for default vectors.

  • If your workflow iterates over the vector keys, you need to use an external word list instead:

    - lexemes = [nlp.vocab[orth] for orth in nlp.vocab.vectors]
    + lexemes = [nlp.vocab[word] for word in external_word_list]
    
  • Vectors.most_similar is not supported because there's no fixed list of vectors to compare your vectors to.

Transformer pipeline design

In the transformer (trf) models, the tagger, parser and ner (if present) all listen to the transformer component. The attribute_ruler and lemmatizer have the same configuration as in the CNN models.

Modifying the default pipeline

For faster processing, you may only want to run a subset of the components in a trained pipeline. The disable and exclude arguments to spacy.load let you control which components are loaded and run. Disabled components are loaded in the background so it's possible to reenable them in the same pipeline in the future with nlp.enable_pipe. To skip loading a component completely, use exclude instead of disable.

Disable part-of-speech tagging and lemmatization

To disable part-of-speech tagging and lemmatization, disable the tagger, morphologizer, attribute_ruler and lemmatizer components.

# Note: English doesn't include a morphologizer
nlp = spacy.load("en_core_web_sm", disable=["tagger", "attribute_ruler", "lemmatizer"])
nlp = spacy.load("en_core_web_trf", disable=["tagger", "attribute_ruler", "lemmatizer"])

The lemmatizer depends on tagger+attribute_ruler or morphologizer for a number of languages. If you disable any of these components, you'll see lemmatizer warnings unless the lemmatizer is also disabled.

v3.3: Catalan, English, French, Russian and Spanish

v3.0-v3.2: Catalan, Dutch, English, French, Greek, Italian, Macedonian, Norwegian, Polish, Russian and Spanish

Use senter rather than parser for fast sentence segmentation

If you need fast sentence segmentation without dependency parses, disable the parser use the senter component instead:

nlp = spacy.load("en_core_web_sm")
nlp.disable_pipe("parser")
nlp.enable_pipe("senter")

The senter component is ~10× faster than the parser and more accurate than the rule-based sentencizer.

Switch from trainable lemmatizer to default lemmatizer

Since v3.3, a number of pipelines use a trainable lemmatizer. You can check whether the lemmatizer is trainable:

nlp = spacy.load("de_core_web_sm")
assert nlp.get_pipe("lemmatizer").is_trainable

If you'd like to switch to a non-trainable lemmatizer that's similar to v3.2 or earlier, you can replace the trainable lemmatizer with the default non-trainable lemmatizer:

# Requirements: pip install spacy-lookups-data
nlp = spacy.load("de_core_web_sm")
# Remove existing lemmatizer
nlp.remove_pipe("lemmatizer")
# Add non-trainable lemmatizer from language defaults
# and load lemmatizer tables from spacy-lookups-data
nlp.add_pipe("lemmatizer").initialize()

Switch from rule-based to lookup lemmatization

For the Dutch, English, French, Greek, Macedonian, Norwegian and Spanish pipelines, you can swap out a trainable or rule-based lemmatizer for a lookup lemmatizer:

# Requirements: pip install spacy-lookups-data
nlp = spacy.load("en_core_web_sm")
nlp.remove_pipe("lemmatizer")
nlp.add_pipe("lemmatizer", config={"mode": "lookup"}).initialize()

Disable everything except NER

For the non-transformer models, the ner component is independent, so you can disable everything else:

nlp = spacy.load("en_core_web_sm", disable=["tok2vec", "tagger", "parser", "attribute_ruler", "lemmatizer"])

In the transformer models, ner listens to the transformer component, so you can disable all components related tagging, parsing, and lemmatization.

nlp = spacy.load("en_core_web_trf", disable=["tagger", "parser", "attribute_ruler", "lemmatizer"])

Move NER to the end of the pipeline

As of v3.1, the NER component is at the end of the pipeline by default.

For access to POS and LEMMA features in an entity_ruler, move ner to the end of the pipeline after attribute_ruler and lemmatizer:

# load without NER
nlp = spacy.load("en_core_web_sm", exclude=["ner"])

# source NER from the same pipeline package as the last component
nlp.add_pipe("ner", source=spacy.load("en_core_web_sm"))

# insert the entity ruler
nlp.add_pipe("entity_ruler", before="ner")