spaCy/website/docs/usage/saving-loading.md

title	menu
Saving and Loading	Basics basics Serialization Methods serialization-methods Entry Points entry-points Models models

Basics

import Serialization101 from 'usage/101/_serialization.md'

In spaCy v2.0, the API for saving and loading has changed to only use the four methods listed above consistently across objects and classes. For an overview of the changes, see this table and the notes on migrating.

Serializing the pipeline

When serializing the pipeline, keep in mind that this will only save out the binary data for the individual components to allow spaCy to restore them – not the entire objects. This is a good thing, because it makes serialization safe. But it also means that you have to take care of storing the language name and pipeline component names as well, and restoring them separately before you can load in the data.

Saving the model meta

The nlp.meta attribute is a JSON-serializable dictionary and contains all model meta information, like the language and pipeline, but also author and license information.

### Serialize
bytes_data = nlp.to_bytes()
lang = nlp.meta["lang"]  # "en"
pipeline = nlp.meta["pipeline"]  # ["tagger", "parser", "ner"]

### Deserialize
nlp = spacy.blank(lang)
for pipe_name in pipeline:
    pipe = nlp.create_pipe(pipe_name)
    nlp.add_pipe(pipe)
nlp.from_bytes(bytes_data)

This is also how spaCy does it under the hood when loading a model: it loads the model's meta.json containing the language and pipeline information, initializes the language class, creates and adds the pipeline components and then loads in the binary data. You can read more about this process here.

Serializing Doc objects efficiently

If you're working with lots of data, you'll probably need to pass analyses between machines, either to use something like Dask or Spark, or even just to save out work to disk. Often it's sufficient to use the Doc.to_array functionality for this, and just serialize the numpy arrays – but other times you want a more general way to save and restore Doc objects.

The DocBin class makes it easy to serialize and deserialize a collection of Doc objects together, and is much more efficient than calling Doc.to_bytes on each individual Doc object. You can also control what data gets saved, and you can merge pallets together for easy map/reduce-style processing.

### {highlight="4,8,9,13,14"}
import spacy
from spacy.tokens import DocBin

doc_bin = DocBin(attrs=["LEMMA", "ENT_IOB", "ENT_TYPE"], store_user_data=True)
texts = ["Some text", "Lots of texts...", "..."]
nlp = spacy.load("en_core_web_sm")
for doc in nlp.pipe(texts):
    doc_bin.add(doc)
bytes_data = docbin.to_bytes()

# Deserialize later, e.g. in a new process
nlp = spacy.blank("en")
doc_bin = DocBin().from_bytes(bytes_data)
docs = list(doc_bin.get_docs(nlp.vocab))

Using Pickle

Example

doc = nlp("This is a text.")
data = pickle.dumps(doc)

When pickling spaCy's objects like the Doc or the EntityRecognizer, keep in mind that they all require the shared Vocab (which includes the string to hash mappings, label schemes and optional vectors). This means that their pickled representations can become very large, especially if you have word vectors loaded, because it won't only include the object itself, but also the entire shared vocab it depends on.

If you need to pickle multiple objects, try to pickle them together instead of separately. For instance, instead of pickling all pipeline components, pickle the entire pipeline once. And instead of pickling several Doc objects separately, pickle a list of Doc objects. Since the all share a reference to the same Vocab object, it will only be included once.

### Pickling objects with shared data {highlight="8-9"}
doc1 = nlp("Hello world")
doc2 = nlp("This is a test")

doc1_data = pickle.dumps(doc1)
doc2_data = pickle.dumps(doc2)
print(len(doc1_data) + len(doc2_data))  # 6636116 😞

doc_data = pickle.dumps([doc1, doc2])
print(len(doc_data))  # 3319761 😃

Pickling Token and Span objects isn't supported. They're only views of the Doc and can't exist on their own. Pickling them would always mean pulling in the parent document and its vocabulary, which has practically no advantage over pickling the parent Doc.

- data = pickle.dumps(doc[10:20])
+ data = pickle.dumps(doc)

If you really only need a span – for example, a particular sentence – you can use Span.as_doc to make a copy of it and convert it to a Doc object. However, note that this will not let you recover contextual information from outside the span.

+ span_doc = doc[10:20].as_doc()
data = pickle.dumps(span_doc)

Implementing serialization methods

When you call nlp.to_disk, nlp.from_disk or load a model package, spaCy will iterate over the components in the pipeline, check if they expose a to_disk or from_disk method and if so, call it with the path to the model directory plus the string name of the component. For example, if you're calling nlp.to_disk("/path"), the data for the named entity recognizer will be saved in /path/ner.

If you're using custom pipeline components that depend on external data – for example, model weights or terminology lists – you can take advantage of spaCy's built-in component serialization by making your custom component expose its own to_disk and from_disk or to_bytes and from_bytes methods. When an nlp object with the component in its pipeline is saved or loaded, the component will then be able to serialize and deserialize itself. The following example shows a custom component that keeps arbitrary JSON-serializable data, allows the user to add to that data and saves and loads the data to and from a JSON file.

Real-world example

To see custom serialization methods in action, check out the new EntityRuler component and its source. Patterns added to the component will be saved to a .jsonl file if the pipeline is serialized to disk, and to a bytestring if the pipeline is serialized to bytes. This allows saving out a model with a rule-based entity recognizer and including all rules with the model data.

### {highlight="15-19,21-26"}
class CustomComponent(object):
    name = "my_component"

    def __init__(self):
        self.data = []

    def __call__(self, doc):
        # Do something to the doc here
        return doc

    def add(self, data):
        # Add something to the component's data
        self.data.append(data)

    def to_disk(self, path):
        # This will receive the directory path + /my_component
        data_path = path / "data.json"
        with data_path.open("w", encoding="utf8") as f:
            f.write(json.dumps(self.data))

    def from_disk(self, path, **cfg):
        # This will receive the directory path + /my_component
        data_path = path / "data.json"
        with data_path.open("r", encoding="utf8") as f:
            self.data = json.loads(f)
        return self

After adding the component to the pipeline and adding some data to it, we can serialize the nlp object to a directory, which will call the custom component's to_disk method.

### {highlight="2-4"}
nlp = spacy.load("en_core_web_sm")
my_component = CustomComponent()
my_component.add({"hello": "world"})
nlp.add_pipe(my_component)
nlp.to_disk("/path/to/model")

The contents of the directory would then look like this. CustomComponent.to_disk converted the data to a JSON string and saved it to a file data.json in its subdirectory:

### Directory structure {highlight="2-3"}
└── /path/to/model
    ├── my_component     # data serialized by "my_component"
    |   └── data.json
    ├── ner              # data for "ner" component
    ├── parser           # data for "parser" component
    ├── tagger           # data for "tagger" component
    ├── vocab            # model vocabulary
    ├── meta.json        # model meta.json with name, language and pipeline
    └── tokenizer        # tokenization rules

When you load the data back in, spaCy will call the custom component's from_disk method with the given file path, and the component can then load the contents of data.json, convert them to a Python object and restore the component state. The same works for other types of data, of course – for instance, you could add a wrapper for a model trained with a different library like TensorFlow or PyTorch and make spaCy load its weights automatically when you load the model package.

When you load a model from disk, spaCy will check the "pipeline" in the model's meta.json and look up the component name in the internal factories. To make sure spaCy knows how to initialize "my_component", you'll need to add it to the factories:

from spacy.language import Language
Language.factories["my_component"] = lambda nlp, **cfg: CustomComponent()

For more details, see the documentation on adding factories or use entry points to make your extension package expose your custom components to spaCy automatically.

Using entry points

Entry points let you expose parts of a Python package you write to other Python packages. This lets one application easily customize the behavior of another, by exposing an entry point in its setup.py. For a quick and fun intro to entry points in Python, check out this excellent blog post. spaCy can load custom function from several different entry points to add pipeline component factories, language classes and other settings. To make spaCy use your entry points, your package needs to expose them and it needs to be installed in the same environment – that's it.

Entry point	Description
spacy_factories	Group of entry points for pipeline component factories to add to Language.factories, keyed by component name.
spacy_languages	Group of entry points for custom Language subclasses, keyed by language shortcut.
spacy_lookups 2.2	Group of entry points for custom Lookups, including lemmatizer data. Used by spaCy's spacy-lookups-data package.
spacy_displacy_colors 2.2	Group of entry points of custom label colors for the displaCy visualizer. The key name doesn't matter, but it should point to a dict of labels and color values. Useful for custom models that predict different entity types.

Custom components via entry points

When you load a model, spaCy will generally use the model's meta.json to set up the language class and construct the pipeline. The pipeline is specified as a list of strings, e.g. "pipeline": ["tagger", "paser", "ner"]. For each of those strings, spaCy will call nlp.create_pipe and look up the name in the built-in factories. If your model wanted to specify its own custom components, you usually have to write to Language.factories before loading the model.

pipe = nlp.create_pipe("custom_component")  # fails 👎

Language.factories["custom_component"] = CustomComponentFactory
pipe = nlp.create_pipe("custom_component")  # works 👍

This is inconvenient and usually required shipping a bunch of component initialization code with the model. Using entry points, model packages and extension packages can now define their own "spacy_factories", which will be added to the built-in factories when the Language class is initialized. If a package in the same environment exposes spaCy entry points, all of this happens automatically and no further user action is required.

To stick with the theme of this entry points blog post, consider the following custom spaCy extension which is initialized with the shared nlp object and will print a snake when it's called as a pipeline component.

Package directory structure

├── snek.py   # the extension code
└── setup.py  # setup file for pip installation

### snek.py
snek = """
    --..,_                     _,.--.
       `'.'.                .'`__ o  `;__.
          '.'.            .'.'`  '---'`  `
            '.`'--....--'`.'
              `'--....--'`
"""

class SnekFactory(object):
    def __init__(self, nlp, **cfg):
        self.nlp = nlp

    def __call__(self, doc):
        print(snek)
        return doc

Since it's a very complex and sophisticated module, you want to split it off into its own package so you can version it and upload it to PyPi. You also want your custom model to be able to define "pipeline": ["snek"] in its meta.json. For that, you need to be able to tell spaCy where to find the factory for "snek". If you don't do this, spaCy will raise an error when you try to load the model because there's no built-in "snek" factory. To add an entry to the factories, you can now expose it in your setup.py via the entry_points dictionary:

### setup.py {highlight="5-7"}
from setuptools import setup

setup(
    name="snek",
    entry_points={
        "spacy_factories": ["snek = snek:SnekFactory"]
    }
)

The entry point definition tells spaCy that the name snek can be found in the module snek (i.e. snek.py) as SnekFactory. The same package can expose multiple entry points. To make them available to spaCy, all you need to do is install the package:

$ python setup.py develop

spaCy is now able to create the pipeline component 'snek':

>>> from spacy.lang.en import English
>>> nlp = English()
>>> snek = nlp.create_pipe("snek")  # this now works! 🐍🎉
>>> nlp.add_pipe(snek)
>>> doc = nlp("I am snek")
    --..,_                     _,.--.
       `'.'.                .'`__ o  `;__.
          '.'.            .'.'`  '---'`  `
            '.`'--....--'`.'
              `'--....--'`

Arguably, this gets even more exciting when you train your en_core_snek_sm model. To make sure snek is installed with the model, you can add it to the model's setup.py. You can then tell spaCy to construct the model pipeline with the snek component by setting "pipeline": ["snek"] in the meta.json.

meta.json

{
    "lang": "en",
    "name": "core_snek_sm",
    "version": "1.0.0",
+   "pipeline": ["snek"]
}

In theory, the entry point mechanism also lets you overwrite built-in factories – including the tokenizer. By default, spaCy will output a warning in these cases, to prevent accidental overwrites and unintended results.

Advanced components with settings

The **cfg keyword arguments that the factory receives are passed down all the way from spacy.load. This means that the factory can respond to custom settings defined when loading the model – for example, the style of the snake to load:

nlp = spacy.load("en_core_snek_sm", snek_style="cute")

SNEKS = {"basic": snek, "cute": cute_snek}  # collection of sneks

class SnekFactory(object):
    def __init__(self, nlp, **cfg):
        self.nlp = nlp
        self.snek_style = cfg.get("snek_style", "basic")
        self.snek = SNEKS[self.snek_style]

    def __call__(self, doc):
        print(self.snek)
        return doc

The factory can also implement other pipeline component like to_disk and from_disk for serialization, or even update to make the component trainable. If a component exposes a from_disk method and is included in a model's pipeline, spaCy will call it on load. This lets you ship custom data with your model. When you save out a model using nlp.to_disk and the component exposes a to_disk method, it will be called with the disk path.

def to_disk(self, path):
    snek_path = path / "snek.txt"
    with snek_path.open("w", encoding="utf8") as snek_file:
        snek_file.write(self.snek)

def from_disk(self, path, **cfg):
    snek_path = path / "snek.txt"
    with snek_path.open("r", encoding="utf8") as snek_file:
        self.snek = snek_file.read()
    return self

The above example will serialize the current snake in a snek.txt in the model data directory. When a model using the snek component is loaded, it will open the snek.txt and make it available to the component.

Custom language classes via entry points

To stay with the theme of the previous example and this blog post on entry points, let's imagine you wanted to implement your own SnekLanguage class for your custom model – but you don't necessarily want to modify spaCy's code to add a language. In your package, you could then implement the following:

### snek.py
from spacy.language import Language
from spacy.attrs import LANG

class SnekDefaults(Language.Defaults):
    lex_attr_getters = dict(Language.Defaults.lex_attr_getters)
    lex_attr_getters[LANG] = lambda text: "snk"


class SnekLanguage(Language):
    lang = "snk"
    Defaults = SnekDefaults
    # Some custom snek language stuff here

Alongside the spacy_factories, there's also an entry point option for spacy_languages, which maps language codes to language-specific Language subclasses:

### setup.py
from setuptools import setup

setup(
    name="snek",
    entry_points={
        "spacy_factories": ["snek = snek:SnekFactory"],
+       "spacy_languages": ["snk = snek:SnekLanguage"]
    }
)

In spaCy, you can then load the custom sk language and it will be resolved to SnekLanguage via the custom entry point. This is especially relevant for model packages, which could then specify "lang": "snk" in their meta.json without spaCy raising an error because the language is not available in the core library.

meta.json

{
-   "lang": "en",
+   "lang": "snk",
    "name": "core_snek_sm",
    "version": "1.0.0",
    "pipeline": ["snek"]
}

from spacy.util import get_lang_class

SnekLanguage = get_lang_class("snk")
nlp = SnekLanguage()

Custom displaCy colors via entry points

If you're training a named entity recognition model for a custom domain, you may end up training different labels that don't have pre-defined colors in the displacy visualizer. The spacy_displacy_colors entry point lets you define a dictionary of entity labels mapped to their color values. It's added to the pre-defined colors and can also overwrite existing values.

Domain-specific NER labels

Good examples of models with domain-specific label schemes are scispaCy and Blackstone.

### snek.py
displacy_colors = {"SNEK": "#3dff74", "HUMAN": "#cfc5ff"}

Given the above colors, the entry point can be defined as follows. Entry points need to have a name, so we use the key colors. However, the name doesn't matter and whatever is defined in the entry point group will be used.

### setup.py
from setuptools import setup

setup(
    name="snek",
    entry_points={
+       "spacy_displacy_colors": ["colors = snek:displacy_colors"]
    }
)

After installing the package, the the custom colors will be used when visualizing text with displacy. Whenever the label SNEK is assigned, it will be displayed in #3dff74.

import DisplaCyEntSnekHtml from 'images/displacy-ent-snek.html'