spaCy/website/docs/usage/models.md

title	next	menu
Models & Languages	usage/facts-figures	Quickstart quickstart Language Support languages Installation & Usage download Production Use production

spaCy's models can be installed as Python packages. This means that they're a component of your application, just like any other module. They're versioned and can be defined as a dependency in your requirements.txt. Models can be installed from a download URL or a local directory, manually or via pip. Their data can be located anywhere on your file system.

Important note

If you're upgrading to spaCy v3.x, you need to download the new models. If you've trained statistical models that use spaCy's annotations, you should retrain your models after updating spaCy. If you don't retrain, you may suffer train/test skew, which might decrease your accuracy.

Quickstart

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

Language support

spaCy currently provides support for the following languages. You can help by improving the existing language data and extending the tokenization patterns. See here for details on how to contribute to model development.

Usage note

If a model is available for a language, you can download it using the spacy download command. In order to use languages that don't yet come with a model, you have to import them directly, or use spacy.blank:

from spacy.lang.fi import Finnish
nlp = Finnish()  # use directly
nlp = spacy.blank("fi")  # blank instance

If lemmatization rules are available for your language, make sure to install spaCy with the lookups option, or install spacy-lookups-data separately in the same environment:

$ pip install spacy[lookups]

import Languages from 'widgets/languages.js'

Multi-language support

# Standard import
from spacy.lang.xx import MultiLanguage
nlp = MultiLanguage()

# With lazy-loading
from spacy.util import get_lang_class
nlp = get_lang_class('xx')

spaCy also supports models trained on more than one language. This is especially useful for named entity recognition. The language ID used for multi-language or language-neutral models is xx. The language class, a generic subclass containing only the base language data, can be found in lang/xx.

To load your model with the neutral, multi-language class, simply set "language": "xx" in your model package's meta.json. You can also import the class directly, or call util.get_lang_class() for lazy-loading.

Chinese language support

The Chinese language class supports three word segmentation options:

from spacy.lang.zh import Chinese

# Disable jieba to use character segmentation
Chinese.Defaults.use_jieba = False
nlp = Chinese()

# Disable jieba through tokenizer config options
cfg = {"use_jieba": False}
nlp = Chinese(meta={"tokenizer": {"config": cfg}})

# Load with "default" model provided by pkuseg
cfg = {"pkuseg_model": "default", "require_pkuseg": True}
nlp = Chinese(meta={"tokenizer": {"config": cfg}})

1. Jieba: Chinese uses Jieba for word segmentation by default. It's enabled when you create a new Chinese language class or call spacy.blank("zh").
2. Character segmentation: Character segmentation is supported by disabling jieba and setting Chinese.Defaults.use_jieba = False before initializing the language class. As of spaCy v2.3.0, the meta tokenizer config options can be used to configure use_jieba.
3. PKUSeg: In spaCy v2.3.0, support for PKUSeg has been added to support better segmentation for Chinese OntoNotes and the new Chinese models.

Note that pkuseg doesn't yet ship with pre-compiled wheels for Python 3.8. If you're running Python 3.8, you can install it from our fork and compile it locally:

$ pip install https://github.com/honnibal/pkuseg-python/archive/master.zip

The meta argument of the Chinese language class supports the following following tokenizer config settings:

Name	Type	Description
pkuseg_model	str	Required: Name of a model provided by pkuseg or the path to a local model directory.
pkuseg_user_dict	str	Optional path to a file with one word per line which overrides the default pkuseg user dictionary.
require_pkuseg	bool	Overrides all jieba settings (optional but strongly recommended).

### Examples
# Load "default" model
cfg = {"pkuseg_model": "default", "require_pkuseg": True}
nlp = Chinese(meta={"tokenizer": {"config": cfg}})

# Load local model
cfg = {"pkuseg_model": "/path/to/pkuseg_model", "require_pkuseg": True}
nlp = Chinese(meta={"tokenizer": {"config": cfg}})

# Override the user directory
cfg = {"pkuseg_model": "default", "require_pkuseg": True, "pkuseg_user_dict": "/path"}
nlp = Chinese(meta={"tokenizer": {"config": cfg}})

You can also modify the user dictionary on-the-fly:

# Append words to user dict
nlp.tokenizer.pkuseg_update_user_dict(["中国", "ABC"])

# Remove all words from user dict and replace with new words
nlp.tokenizer.pkuseg_update_user_dict(["中国"], reset=True)

# Remove all words from user dict
nlp.tokenizer.pkuseg_update_user_dict([], reset=True)

The Chinese models provided by spaCy include a custom pkuseg model trained only on Chinese OntoNotes 5.0, since the models provided by pkuseg include data restricted to research use. For research use, pkuseg provides models for several different domains ("default", "news" "web", "medicine", "tourism") and for other uses, pkuseg provides a simple training API:

import pkuseg
from spacy.lang.zh import Chinese

# Train pkuseg model
pkuseg.train("train.utf8", "test.utf8", "/path/to/pkuseg_model")
# Load pkuseg model in spaCy Chinese tokenizer
nlp = Chinese(meta={"tokenizer": {"config": {"pkuseg_model": "/path/to/pkuseg_model", "require_pkuseg": True}}})

Japanese language support

from spacy.lang.ja import Japanese

# Load SudachiPy with split mode A (default)
nlp = Japanese()

# Load SudachiPy with split mode B
cfg = {"split_mode": "B"}
nlp = Japanese(meta={"tokenizer": {"config": cfg}})

The Japanese language class uses SudachiPy for word segmentation and part-of-speech tagging. The default Japanese language class and the provided Japanese models use SudachiPy split mode A. The meta argument of the Japanese language class can be used to configure the split mode to A, B or C.

If you run into errors related to sudachipy, which is currently under active development, we suggest downgrading to sudachipy==0.4.5, which is the version used for training the current Japanese models.

Installing and using models

The easiest way to download a model is via spaCy's download command. It takes care of finding the best-matching model compatible with your spaCy installation.

Important note for v3.0

Note that as of spaCy v3.0, model shortcut links that create (potentially brittle) symlinks in your spaCy installation are deprecated. To download and load an installed model, use its full name:

- python -m spacy download en
+ python -m spacy dowmload en_core_web_sm

- nlp = spacy.load("en")
+ nlp = spacy.load("en_core_web_sm")

# Download best-matching version of specific model for your spaCy installation
python -m spacy download en_core_web_sm

# Download exact model version
python -m spacy download en_core_web_sm-2.2.0 --direct

The download command will install the model via pip and place the package in your site-packages directory.

pip install spacy
python -m spacy download en_core_web_sm

import spacy
nlp = spacy.load("en_core_web_sm")
doc = nlp("This is a sentence.")

Installation via pip

To download a model directly using pip, point pip install to the URL or local path of the archive file. To find the direct link to a model, head over to the model releases, right click on the archive link and copy it to your clipboard.

# With external URL
pip install https://github.com/explosion/spacy-models/releases/download/en_core_web_sm-3.0.0/en_core_web_sm-3.0.0.tar.gz

# With local file
pip install /Users/you/en_core_web_sm-3.0.0.tar.gz

By default, this will install the model into your site-packages directory. You can then use spacy.load() to load it via its package name or import it explicitly as a module. If you need to download models as part of an automated process, we recommend using pip with a direct link, instead of relying on spaCy's download command.

You can also add the direct download link to your application's requirements.txt. For more details, see the section on working with models in production.

Manual download and installation

In some cases, you might prefer downloading the data manually, for example to place it into a custom directory. You can download the model via your browser from the latest releases, or configure your own download script using the URL of the archive file. The archive consists of a model directory that contains another directory with the model data.

### Directory structure {highlight="7"}
└── en_core_web_md-3.0.0.tar.gz       # downloaded archive
    ├── meta.json                     # model meta data
    ├── setup.py                      # setup file for pip installation
    └── en_core_web_md                # 📦 model package
        ├── __init__.py               # init for pip installation
        ├── meta.json                 # model meta data
        └── en_core_web_md-3.0.0      # model data

You can place the model package directory anywhere on your local file system.

Using models with spaCy

To load a model, use spacy.load with the model's package name or a path to the data directory:

Important note for v3.0

Note that as of spaCy v3.0, model shortcut links that create (potentially brittle) symlinks in your spaCy installation are deprecated. To load an installed model, use its full name:

- nlp = spacy.load("en")
+ nlp = spacy.load("en_core_web_sm")

import spacy
nlp = spacy.load("en_core_web_sm")           # load model package "en_core_web_sm"
nlp = spacy.load("/path/to/en_core_web_sm")  # load package from a directory

doc = nlp("This is a sentence.")

You can use the info command or spacy.info() method to print a model's meta data before loading it. Each Language object with a loaded model also exposes the model's meta data as the attribute meta. For example, nlp.meta['version'] will return the model's version.

Importing models as modules

If you've installed a model via spaCy's downloader, or directly via pip, you can also import it and then call its load() method with no arguments:

### {executable="true"}
import en_core_web_sm

nlp = en_core_web_sm.load()
doc = nlp("This is a sentence.")

How you choose to load your models ultimately depends on personal preference. However, for larger code bases, we usually recommend native imports, as this will make it easier to integrate models with your existing build process, continuous integration workflow and testing framework. It'll also prevent you from ever trying to load a model that is not installed, as your code will raise an ImportError immediately, instead of failing somewhere down the line when calling spacy.load().

For more details, see the section on working with models in production.

Using your own models

If you've trained your own model, for example for additional languages or custom named entities, you can save its state using the Language.to_disk() method. To make the model more convenient to deploy, we recommend wrapping it as a Python package.

For more information and a detailed guide on how to package your model, see the documentation on saving and loading models.

Using models in production

If your application depends on one or more models, you'll usually want to integrate them into your continuous integration workflow and build process. While spaCy provides a range of useful helpers for downloading, linking and loading models, the underlying functionality is entirely based on native Python packages. This allows your application to handle a model like any other package dependency.

Downloading and requiring model dependencies

spaCy's built-in download command is mostly intended as a convenient, interactive wrapper. It performs compatibility checks and prints detailed error messages and warnings. However, if you're downloading models as part of an automated build process, this only adds an unnecessary layer of complexity. If you know which models your application needs, you should be specifying them directly.

Because all models are valid Python packages, you can add them to your application's requirements.txt. If you're running your own internal PyPi installation, you can upload the models there. pip's requirements file format supports both package names to download via a PyPi server, as well as direct URLs.

### requirements.txt
spacy>=2.2.0,<3.0.0
https://github.com/explosion/spacy-models/releases/download/en_core_web_sm-2.2.0/en_core_web_sm-2.2.0.tar.gz#egg=en_core_web_sm

Specifying #egg= with the package name tells pip which package to expect from the download URL. This way, the package won't be re-downloaded and overwritten if it's already installed - just like when you're downloading a package from PyPi.

All models are versioned and specify their spaCy dependency. This ensures cross-compatibility and lets you specify exact version requirements for each model. If you've trained your own model, you can use the package command to generate the required meta data and turn it into a loadable package.

Loading and testing models

Models are regular Python packages, so you can also import them as a package using Python's native import syntax, and then call the load method to load the model data and return an nlp object:

import en_core_web_sm
nlp = en_core_web_sm.load()

In general, this approach is recommended for larger code bases, as it's more "native", and doesn't depend on symlinks or rely on spaCy's loader to resolve string names to model packages. If a model can't be imported, Python will raise an ImportError immediately. And if a model is imported but not used, any linter will catch that.

Similarly, it'll give you more flexibility when writing tests that require loading models. For example, instead of writing your own try and except logic around spaCy's loader, you can use pytest's importorskip() method to only run a test if a specific model or model version is installed. Each model package exposes a __version__ attribute which you can also use to perform your own version compatibility checks before loading a model.