Add pipelines 101 and rewrite pipelines workflow

This commit is contained in:
ines 2017-05-24 19:25:13 +02:00
parent 54885b5e88
commit 8aaed8bea7
4 changed files with 349 additions and 151 deletions

View File

@ -105,7 +105,7 @@
},
"language-processing-pipeline": {
"title": "Natural language processing pipelines",
"title": "Language processing pipelines",
"next": "deep-learning"
},

View File

@ -0,0 +1,44 @@
//- 💫 DOCS > USAGE > SPACY 101 > PIPELINES
p
| When you call #[code nlp] on a text, spaCy first tokenizes the text to
| produce a #[code Doc] object. The #[code Doc] is the processed in several
| different steps this is also referred to as the
| #[strong processing pipeline]. The pipeline used by our
| #[+a("/docs/usage/models") default models] consists of a
| vectorizer, a tagger, a parser and an entity recognizer. Each pipeline
| component returns the processed #[code Doc], which is then passed on to
| the next component.
+image
include ../../../assets/img/docs/pipeline.svg
.u-text-right
+button("/assets/img/docs/pipeline.svg", false, "secondary").u-text-tag View large graphic
+table(["Name", "Component", "Creates"])
+row
+cell tokenizer
+cell #[+api("tokenizer") #[code Tokenizer]]
+cell #[code Doc]
+row("divider")
+cell vectorizer
+cell #[code Vectorizer]
+cell #[code Doc.tensor]
+row
+cell tagger
+cell #[+api("tagger") #[code Tagger]]
+cell #[code Doc[i].tag]
+row
+cell parser
+cell #[+api("dependencyparser") #[code DependencyParser]]
+cell
| #[code Doc[i].head], #[code Doc[i].dep], #[code Doc.sents],
| #[code Doc.noun_chunks]
+row
+cell ner
+cell #[+api("entityrecognizer") #[code EntityRecognizer]]
+cell #[code Doc.ents], #[code Doc[i].ent_iob], #[code Doc[i].ent_type]

View File

@ -2,164 +2,316 @@
include ../../_includes/_mixins
p
| The standard entry point into spaCy is the #[code spacy.load()]
| function, which constructs a language processing pipeline. The standard
| variable name for the language processing pipeline is #[code nlp], for
| Natural Language Processing. The #[code nlp] variable is usually an
| instance of class #[code spacy.language.Language]. For English, the
| #[code spacy.en.English] class is the default.
+h(2, "101") Pipelines 101
include _spacy-101/_pipelines
+h(2, "pipelines") How pipelines work
p
| You'll use the nlp instance to produce #[+api("doc") #[code Doc]]
| objects. You'll then use the #[code Doc] object to access linguistic
| annotations to help you with whatever text processing task you're
| trying to do.
+code.
import spacy # See "Installing spaCy"
nlp = spacy.load('en') # You are here.
doc = nlp(u'Hello, spacy!') # See "Using the pipeline"
print((w.text, w.pos_) for w in doc) # See "Doc, Span and Token"
+aside("Why do we have to preload?")
| Loading the models takes ~200x longer than
| processing a document. We therefore want to amortize the start-up cost
| across multiple invocations. It's often best to wrap the pipeline as a
| singleton. The library avoids doing that for you, because it's a
| difficult design to back out of.
p The #[code load] function takes the following positional arguments:
+table([ "Name", "Description" ])
+row
+cell #[code lang_id]
+cell
| An ID that is resolved to a class or factory function by
| #[code spacy.util.get_lang_class()]. Common values are
| #[code 'en'] for the English pipeline, or #[code 'de'] for the
| German pipeline. You can register your own factory function or
| class with #[code spacy.util.set_lang_class()].
| spaCy makes it very easy to create your own pipelines consisting of
| reusable components this includes spaCy's default vectorizer, tagger,
| parser and entity regcognizer, but also your own custom processing
| functions. A pipeline component can be added to an already existing
| #[code nlp] object, specified when initialising a #[code Language] class,
| or defined within a
| #[+a("/docs/usage/saving-loading#models-generating") model package].
p
| All keyword arguments are passed forward to the pipeline factory. No
| keyword arguments are required. The built-in factories (e.g.
| #[code spacy.en.English], #[code spacy.de.German]), which are subclasses
| of #[+api("language") #[code Language]], respond to the following
| keyword arguments:
| When you load a model, spaCy first consults the model's
| #[+a("/docs/usage/saving-loading#models-generating") meta.json] for its
| #[code setup] details. This typically includes the ID of a language class,
| and an optional list of pipeline components. spaCy then does the
| following:
+table([ "Name", "Description"])
+row
+cell #[code path]
+cell
| Where to load the data from. If None, the default data path is
| fetched via #[code spacy.util.get_data_path()]. You can
| configure this default using #[code spacy.util.set_data_path()].
| The data path is expected to be either a string, or an object
| responding to the #[code pathlib.Path] interface. If the path is
| a string, it will be immediately transformed into a
| #[code pathlib.Path] object. spaCy promises to never manipulate
| or open file-system paths as strings. All access to the
| file-system is done via the #[code pathlib.Path] interface.
| spaCy also promises to never check the type of path objects.
| This allows you to customize the loading behaviours in arbitrary
| ways, by creating your own object that implements the
| #[code pathlib.Path] interface.
+aside-code("meta.json (excerpt)", "json").
{
"name": "example_model",
"description": "Example model for spaCy",
"setup": {
"lang": "en",
"pipeline": ["token_vectors", "tagger"]
}
}
+row
+cell #[code pipeline]
+cell
| A sequence of functions that take the Doc object and modify it
| in-place. See
| #[+a("customizing-pipeline") Customizing the pipeline].
+row
+cell #[code create_pipeline]
+cell
| Callback to construct the pipeline sequence. It should accept
| the #[code nlp] instance as its only argument, and return a
| sequence of functions that take the #[code Doc] object and
| modify it in-place.
| See #[+a("customizing-pipeline") Customizing the pipeline]. If
| a value is supplied to the pipeline keyword argument, the
| #[code create_pipeline] keyword argument is ignored.
+row
+cell #[code make_doc]
+cell A function that takes the input and returns a document object.
+row
+cell #[code create_make_doc]
+cell
| Callback to construct the #[code make_doc] function. It should
| accept the #[code nlp] instance as its only argument. To use the
| built-in annotation processes, it should return an object of
| type #[code Doc]. If a value is supplied to the #[code make_doc]
| keyword argument, the #[code create_make_doc] keyword argument
| is ignored.
+row
+cell #[code vocab]
+cell Supply a pre-built Vocab instance, instead of constructing one.
+row
+cell #[code add_vectors]
+cell
| Callback that installs word vectors into the Vocab instance. The
| #[code add_vectors] callback should take a
| #[+api("vocab") #[code Vocab]] instance as its only argument,
| and set the word vectors and #[code vectors_length] in-place. See
| #[+a("word-vectors-similarities") Word Vectors and Similarities].
+row
+cell #[code tagger]
+cell Supply a pre-built tagger, instead of creating one.
+row
+cell #[code parser]
+cell Supply a pre-built parser, instead of creating one.
+row
+cell #[code entity]
+cell Supply a pre-built entity recognizer, instead of creating one.
+row
+cell #[code matcher]
+cell Supply a pre-built matcher, instead of creating one.
+h(2, "customizing") Customizing the pipeline
+list("numbers")
+item
| Look up #[strong pipeline IDs] in the available
| #[strong pipeline factories].
+item
| Initialise the #[strong pipeline components] by calling their
| factories with the #[code Vocab] as an argument. This gives each
| factory and component access to the pipeline's shared data, like
| strings, morphology and annotation scheme.
+item
| Load the #[strong language class and data] for the given ID via
| #[+api("util.get_lang_class") #[code get_lang_class]].
+item
| Pass the path to the #[strong model data] to the #[code Language]
| class and return it.
p
| spaCy provides several linguistic annotation functions by default. Each
| function takes a Doc object, and modifies it in-place. The default
| pipeline is #[code [nlp.tagger, nlp.entity, nlp.parser]]. spaCy 1.0
| introduced the ability to customise this pipeline with arbitrary
| functions.
+code.
def arbitrary_fixup_rules(doc):
for token in doc:
if token.text == u'bill' and token.tag_ == u'NNP':
token.tag_ = u'NN'
def custom_pipeline(nlp):
return (nlp.tagger, arbitrary_fixup_rules, nlp.parser, nlp.entity)
nlp = spacy.load('en', create_pipeline=custom_pipeline)
p
| The easiest way to customise the pipeline is to pass a
| #[code create_pipeline] callback to the #[code spacy.load()] function.
p
| The callback you pass to #[code create_pipeline] should take a single
| argument, and return a sequence of callables. Each callable in the
| sequence should accept a #[code Doc] object and modify it in place.
p
| Instead of passing a callback, you can also write to the
| #[code .pipeline] attribute directly.
| So when you call this...
+code.
nlp = spacy.load('en')
nlp.pipeline = [nlp.tagger]
p
| ... the model tells spaCy to use the pipeline
| #[code ["vectorizer", "tagger", "parser", "ner"]]. spaCy will then look
| up each string in its internal factories registry and initialise the
| individual components. It'll then load #[code spacy.lang.en.English],
| pass it the path to the model's data directory, and return it for you
| to use as the #[code nlp] object.
p
| When you call #[code nlp] on a text, spaCy will #[strong tokenize] it and
| then #[strong call each component] on the #[code Doc], in order.
| Components all return the modified document, which is then processed by
| the component next in the pipeline.
+code("The pipeline under the hood").
doc = nlp.make_doc(u'This is a sentence')
for proc in nlp.pipeline:
doc = proc(doc)
+h(2, "creating") Creating pipeline components and factories
p
| spaCy lets you customise the pipeline with your own components. Components
| are functions that receive a #[code Doc] object, modify and return it.
| If your component is stateful, you'll want to create a new one for each
| pipeline. You can do that by defining and registering a factory which
| receives the shared #[code Vocab] object and returns a component.
+h(3, "creating-component") Creating a component
p
| A component receives a #[code Doc] object and
| #[strong performs the actual processing] for example, using the current
| weights to make a prediction and set some annotation on the document. By
| adding a component to the pipeline, you'll get access to the #[code Doc]
| at any point #[strong during] processing instead of only being able to
| modify it afterwards.
+aside-code("Example").
def my_component(doc):
# do something to the doc here
return doc
+table(["Argument", "Type", "Description"])
+row
+cell #[code doc]
+cell #[code Doc]
+cell The #[code Doc] object processed by the previous component.
+footrow
+cell returns
+cell #[code Doc]
+cell The #[code Doc] object processed by this pipeline component.
p
| When creating a new #[code Language] class, you can pass it a list of
| pipeline component functions to execute in that order. You can also
| add it to an existing pipeline by modifying #[code nlp.pipeline] just
| be careful not to overwrite a pipeline or its components by accident!
+code.
# Create a new Language object with a pipeline
from spacy.language import Language
nlp = Language(pipeline=[my_component])
# Modify an existing pipeline
nlp = spacy.load('en')
nlp.pipeline.append(my_component)
+h(3, "creating-factory") Creating a factory
p
| A factory is a #[strong function that returns a pipeline component].
| It's called with the #[code Vocab] object, to give it access to the
| shared data between components for example, the strings, morphology,
| vectors or annotation scheme. Factories are useful for creating
| #[strong stateful components], especially ones which
| #[strong depend on shared data].
+aside-code("Example").
def my_factory(vocab):
# load some state
def my_component(doc):
# process the doc
return doc
return my_component
+table(["Argument", "Type", "Description"])
+row
+cell #[code vocab]
+cell #[coce Vocab]
+cell
| Shared data between components, including strings, morphology,
| vectors etc.
+footrow
+cell returns
+cell callable
+cell The pipeline component.
p
| By creating a factory, you're essentially telling spaCy how to get the
| pipeline component #[strong once the vocab is available]. Factories need to
| be registered via #[+api("spacy#set_factory") #[code set_factory()]] and
| by assigning them a unique ID. This ID can be added to the pipeline as a
| string. When creating a pipeline, you're free to mix strings and
| callable components:
+code.
spacy.set_factory('my_factory', my_factory)
nlp = Language(pipeline=['my_factory', my_other_component])
p
| If spaCy comes across a string in the pipeline, it will try to resolve it
| by looking it up in the available factories. The factory will then be
| initialised with the #[code Vocab]. Providing factory names instead of
| callables also makes it easy to specify them in the model's
| #[+a("/docs/usage/saving-loading#models-generating") meta.json]. If you're
| training your own model and want to use one of spaCy's default components,
| you won't have to worry about finding and implementing it either to use
| the default tagger, simply add #[code "tagger"] to the pipeline, and
| #[strong spaCy will know what to do].
+infobox("Important note")
| Because factories are #[strong resolved on initialisation] of the
| #[code Language] class, it's #[strong not possible] to add them to the
| pipeline afterwards, e.g. by modifying #[code nlp.pipeline]. This only
| works with individual component functions. To use factories, you need to
| create a new #[code Language] object, or generate a
| #[+a("/docs/usage/saving-loading#models-generating") model package] with
| a custom pipeline.
+h(2, "example1") Example: Custom sentence segmentation logic
+aside("Real-world examples")
| To see real-world examples of pipeline factories and components in action,
| you can have a look at the source of spaCy's built-in components, e.g.
| the #[+src(gh("spacy")) tagger], #[+src(gh("spacy")) parser] or
| #[+src(gh("spacy")) entity recognizer].
p
| Let's say you want to implement custom logic to improve spaCy's sentence
| boundary detection. Currently, sentence segmentation is based on the
| dependency parse, which doesn't always produce ideal results. The custom
| logic should therefore be applied #[strong after] tokenization, but
| #[strong before] the dependency parsing this way, the parser can also
| take advantage of the sentence boundaries.
+code.
def sbd_component(doc):
for i, token in enumerate(doc[:-2]):
# define sentence start if period + titlecase token
if token.text == '.' and doc[i+1].is_title:
doc[i+1].sent_start = True
return doc
p
| In this case, we simply want to add the component to the existing
| pipeline of the English model. We can do this by inserting it at index 0
| of #[code nlp.pipeline]:
+code.
nlp = spacy.load('en')
nlp.pipeline.insert(0, sbd_component)
p
| When you call #[code nlp] on some text, spaCy will tokenize it to create
| a #[code Doc] object, and first call #[code sbd_component] on it, followed
| by the model's default pipeline.
+h(2, "example2") Example: Sentiment model
p
| Let's say you have trained your own document sentiment model on English
| text. After tokenization, you want spaCy to first execute the
| #[strong default vectorizer], followed by a custom
| #[strong sentiment component] that adds a #[code .sentiment]
| property to the #[code Doc], containing your model's sentiment precition.
p
| Your component class will have a #[code from_disk()] method that spaCy
| calls to load the model data. When called, the component will compute
| the sentiment score, add it to the #[code Doc] and return the modified
| document. Optionally, the component can include an #[code update()] method
| to allow training the model.
+code.
import pickle
from pathlib import Path
class SentimentComponent(object):
def __init__(self, vocab):
self.weights = None
def __call__(self, doc):
doc.sentiment = sum(self.weights*doc.vector) # set sentiment property
return doc
def from_disk(self, path): # path = model path + factory ID ('sentiment')
self.weights = pickle.load(Path(path) / 'weights.bin') # load weights
return self
def update(self, doc, gold): # update weights allows training!
prediction = sum(self.weights*doc.vector)
self.weights -= 0.001*doc.vector*(prediction-gold.sentiment)
p
| The factory will initialise the component with the #[code Vocab] object.
| To be able to add it to your model's pipeline as #[code 'sentiment'],
| it also needs to be registered via
| #[+api("spacy#set_factory") #[code set_factory()]].
+code.
def sentiment_factory(vocab):
component = SentimentComponent(vocab) # initialise component
return component
spacy.set_factory('sentiment', sentiment_factory)
p
| The above code should be #[strong shipped with your model]. You can use
| the #[+api("cli#package") #[code package]] command to create all required
| files and directories. The model package will include an
| #[+src(gh("spacy-dev-resources", "templates/model/en_model_name/__init__.py")) __init__.py]
| with a #[code load()] method, that will initialise the language class with
| the model's pipeline and call the #[code from_disk()] method to load
| the model data.
p
| In the model package's meta.json, specify the language class and pipeline
| IDs in #[code setup]:
+code("meta.json (excerpt)", "json").
{
"name": "my_sentiment_model",
"version": "1.0.0",
"spacy_version": ">=2.0.0,<3.0.0",
"setup": {
"lang": "en",
"pipeline": ["vectorizer", "sentiment"]
}
}
p
| When you load your new model, spaCy will call the model's #[code load()]
| method. This will return a #[code Language] object with a pipeline
| containing the default vectorizer, and the sentiment component returned
| by your custom #[code "sentiment"] factory.
+code.
nlp = spacy.load('my_sentiment_model')
doc = nlp(u'I love pizza')
assert doc.sentiment
+infobox("Saving and loading models")
| For more information and a detailed guide on how to package your model,
| see the documentation on
| #[+a("/docs/usage/saving-loading#models") saving and loading models].

View File

@ -105,6 +105,8 @@ include _spacy-101/_word-vectors
+h(2, "pipelines") Pipelines
include _spacy-101/_pipelines
+h(2, "serialization") Serialization
include _spacy-101/_serialization