spaCy/extra/DEVELOPER_DOCS/Language.md
Ines Montani 593a22cf2d
Add development docs for Language and code conventions (#8745)
* WIP: add dev docs for Language / config [ci skip]

* Add section on initialization [ci skip]

* Fix wording [ci skip]

* Add code conventions WIP [ci skip]

* Update code convention docs [ci skip]

* Update contributing guide and conventions [ci skip]

* Update Code Conventions.md [ci skip]

* Clarify sourced components + vectors

* Apply suggestions from code review [ci skip]

Co-authored-by: Sofie Van Landeghem <svlandeg@users.noreply.github.com>

* Update wording and add link [ci skip]

* restructure slightly + extended index

* remove paragraph that breaks flow and is repeated in more detail later

* fix anchors

Co-authored-by: Adriane Boyd <adrianeboyd@gmail.com>
Co-authored-by: Sofie Van Landeghem <svlandeg@users.noreply.github.com>
Co-authored-by: svlandeg <sofie.vanlandeghem@gmail.com>
2021-08-17 09:38:15 +02:00

20 KiB
Raw Blame History

Language

Reference: spacy/language.py

  1. Constructing the nlp object from a config
  2. Initialization

1. Constructing the nlp object from a config

1A. Overview

Most of the functions referenced in the config are regular functions with arbitrary arguments registered via the function registry. However, the pipeline components are a bit special: they don't only receive arguments passed in via the config file, but also the current nlp object and the string name of the individual component instance (so a user can have multiple components created with the same factory, e.g. ner_one and ner_two). This name can then be used by the components to add to the losses and scores. This special requirement means that pipeline components can't just be resolved via the config the "normal" way: we need to retrieve the component functions manually and pass them their arguments, plus the nlp and name.

The Language.from_config classmethod takes care of constructing the nlp object from a config. It's the single place where this happens and what spacy.load delegates to under the hood. Its main responsibilities are:

  • Load and validate the config, and optionally auto-fill all missing values that we either have defaults for in the config template or that registered function arguments define defaults for. This helps ensure backwards-compatibility, because we're able to add a new argument foo: str = "bar" to an existing function, without breaking configs that don't specity it.
  • Execute relevant callbacks for pipeline creation, e.g. optional functions called before and after creation of the nlp object and pipeline.
  • Initialize language subclass and create tokenizer. The from_config classmethod will always be called on a language subclass, e.g. English, not on Language directly. Initializing the subclass takes a callback to create the tokenizer.
  • Set up the pipeline components. Components can either refer to a component factory or a source, i.e. an existing pipeline that's loaded and that the component is then copied from. We also need to ensure that we update the information about which components are disabled.
  • Manage listeners. If sourced components "listen" to other components (tok2vec, transformer), we need to ensure that the references are valid. If the config specifies that listeners should be replaced by copies (e.g. to give the ner component its own tok2vec model instead of listening to the shared tok2vec component in the pipeline), we also need to take care of that.

Note that we only resolve and load selected sections in Language.from_config, i.e. only the parts that are relevant at runtime, which is [nlp] and [components]. We don't want to be resolving anything related to training or initialization, since this would mean loading and constructing unnecessary functions, including functions that require information that isn't necessarily available at runtime, like paths.train.

1B. How pipeline component factories work in the config

As opposed to regular registered functions that refer to a registry and function name (e.g. "@misc": "foo.v1"), pipeline components follow a different format and refer to their component factory name. This corresponds to the name defined via the @Language.component or @Language.factory decorator. We need this decorator to define additional meta information for the components, like their default config and score weights.

[components.my_component]
factory = "foo"
some_arg = "bar"
other_arg = ${paths.some_path}

This means that we need to create and resolve the config["components"] separately from the rest of the config. There are some important considerations and things we need to manage explicitly to avoid unexpected behavior:

Variable interpolation

When a config is resolved, references to variables are replaced, so that the functions receive the correct value instead of just the variable name. To interpolate a config, we need it in its entirety: we couldn't just interpolate a subsection that refers to variables defined in a different subsection. So we first interpolate the entire config.

However, the nlp.config should include the original config with variables intact otherwise, loading a pipeline and saving it to disk will destroy all logic implemented via variables and hard-code the values all over the place. This means that when we create the components, we need to keep two versions of the config: the interpolated config with the "real" values and the raw_config including the variable references.

Factory registry

Component factories are special and use the @Language.factory or @Language.component decorator to register themselves and their meta. When the decorator runs, it performs some basic validation, stores the meta information for the factory on the Language class (default config, scores etc.) and then adds the factory function to registry.factories. The component decorator can be used for registering simple functions that just take a Doc object and return it so in that case, we create the factory for the user automatically.

There's one important detail to note about how factories are registered via entry points: A package that wants to expose spaCy components still needs to register them via the @Language decorators so we have the component meta information and can perform required checks. All we care about here is that the decorated function is loaded and imported. When it is, the @Language decorator takes care of everything, including actually registering the component factory.

Normally, adding to the registry via an entry point will just add the function to the registry under the given name. But for spacy_factories, we don't actually want that: all we care about is that the function decorated with @Language is imported so the decorator runs. So we only exploit Python's entry point system to automatically import the function, and the spacy_factories entry point group actually adds to a separate registry, registry._factories, under the hood. Its only purpose is that the functions are imported. The decorator then runs, creates the factory if needed and adds it to the registry.factories registry.

Language-specific factories

spaCy supports registering factories on the Language base class, as well as language-specific subclasses like English or German. This allows providing different factories depending on the language, e.g. a different default lemmatizer. The Language.get_factory_name classmethod constructs the factory name as {lang}.{name} if a language is available (i.e. if it's a subclass) and falls back to {name} otherwise. So @German.factory("foo") will add a factory de.foo under the hood. If you add nlp.add_pipe("foo"), we first check if there's a factory for {nlp.lang}.foo and if not, we fall back to checking for a factory foo.

Creating a pipeline component from a factory

Language.add_pipe takes care of adding a pipeline component, given its factory name, its config. If no source pipeline to copy the component from is provided, it delegates to Language.create_pipe, which sets up the actual component function.

  • Validate the config and make sure that the factory was registered via the decorator and that we have meta for it.
  • Update the component config with any defaults specified by the component's default_config, if available. This is done by merging the values we receive into the defaults. It ensures that you can still add a component without having to specify its entire config including more complex settings like model. If no model is defined, we use the default.
  • Check if we have a language-specific factory for the given nlp.lang and if not, fall back to the global factory.
  • Construct the component config, consisting of whatever arguments were provided, plus the current nlp object and name, which are default expected arguments of all factories. We also add a reference to the @factories registry, so we can resolve the config via the registry, like any other config. With the added nlp and name, it should now include all expected arguments of the given function.
  • Fill the config to make sure all unspecified defaults from the function arguments are added and update the raw_config (uninterpolated with variables intact) with that information, so the component config we store in nlp.config is up to date. We do this by adding the raw_config into the filled config otherwise, the references to variables would be overwritten.
  • Resolve the config and create all functions it refers to (e.g. model). This gives us the actual component function that we can insert into the pipeline.

1C. Sourcing a pipeline component

[components.ner]
source = "en_core_web_sm"

spaCy also allows "sourcing" a component, which will copy it over from an existing pipeline. In this case, Language.add_pipe will delegate to Language.create_pipe_from_source. In order to copy a component effectively and validate it, the source pipeline first needs to be loaded. This is done in Language.from_config, so a source pipeline only has to be loaded once if multiple components source from it. Sourcing a component will perform the following checks and modifications:

  • For each sourced pipeline component loaded in Language.from_config, a hash of the vectors data from the source pipeline is stored in the pipeline meta so we're able to check whether the vectors match and warn if not (since different vectors that are used as features in components can lead to degraded performance). Because the vectors are not loaded at the point when components are sourced, the check is postponed to init_vocab as part of Language.initialize.
  • If the sourced pipeline component is loaded through Language.add_pipe(source=), the vectors are already loaded and can be compared directly. The check compares the shape and keys first and finally falls back to comparing the actual byte representation of the vectors (which is slower).
  • Ensure that the component is available in the pipeline.
  • Interpolate the entire config of the source pipeline so all variables are replaced and the component's config that's copied over doesn't include references to variables that are not available in the destination config.
  • Add the source vocab.strings to the destination's vocab.strings so we don't end up with unavailable strings in the final pipeline (which would also include labels used by the sourced component).

Note that there may be other incompatibilities that we're currently not checking for and that could cause a sourced component to not work in the destination pipeline. We're interested in adding more checks here but there'll always be a small number of edge cases we'll never be able to catch, including a sourced component depending on other pipeline state that's not available in the destination pipeline.

1D. Tracking components as they're modified

The Language class implements methods for removing, replacing or renaming pipeline components. Whenever we make these changes, we need to update the information stored on the Language object to ensure that it matches the current state of the pipeline. If a user just writes to nlp.config manually, we obviously can't ensure that the config matches the reality but since we offer modification via the pipe methods, it's expected that spaCy keeps the config in sync under the hood. Otherwise, saving a modified pipeline to disk and loading it back wouldn't work. The internal attributes we need to keep in sync here are:

Attribute Type Description
Language._components List[Tuple[str, Callable]] All pipeline components as (name, func) tuples. This is used as the source of truth for Language.pipeline, Language.pipe_names and Language.components.
Language._pipe_meta Dict[str, FactoryMeta] The meta information of a component's factory, keyed by component name. This can include multiple components referring to the same factory meta.
Language._pipe_configs Dict[str, Config] The component's config, keyed by component name.
Language._disabled Set[str] Names of components that are currently disabled.
Language._config Config The underlying config. This is only internals and will be used as the basis for constructing the config in the Language.config property.

In addition to the actual component settings in [components], the config also allows specifying component-specific arguments via the [initialize.components] block, which are passed to the component's initialize method during initialization if it's available. So we also need to keep this in sync in the underlying config.

1E. spaCy's config utility functions

When working with configs in spaCy, make sure to use the utility functions provided by spaCy if available, instead of calling the respective Config methods. The utilities take care of providing spaCy-specific error messages and ensure a consistent order of config sections by setting the section_order argument. This ensures that exported configs always have the same consistent format.

  • util.load_config: load a config from a file
  • util.load_config_from_str: load a confirm from a string representation
  • util.copy_config: deepcopy a config

2. Initialization

Initialization is a separate step of the config lifecycle that's not performed at runtime. It's implemented via the training.initialize.init_nlp helper and calls into Language.initialize method, which sets up the pipeline and component models before training. The initialize method takes a callback that returns a sample of examples, which is used to initialize the component models, add all required labels and perform shape inference if applicable.

Components can also define custom initialization setting via the [initialize.components] block, e.g. if they require external data like lookup tables to be loaded in. All config settings defined here will be passed to the component's initialize method, if it implements one. Components are expected to handle their own serialization after they're initialized so that any data or settings they require are saved with the pipeline and will be available from disk when the pipeline is loaded back at runtime.

2A. Initialization for training

The init_nlp function is called before training and returns an initialized nlp object that can be updated with the examples. It only needs the config and does the following:

  • Load and validate the config. In order to validate certain settings like the seed, we also interpolate the config to get the final value (because in theory, a user could provide this via a variable).
  • Set up the GPU allocation, if required.
  • Create the nlp object from the raw, uninterpolated config, which delegates to Language.from_config. Since this method may modify and auto-fill the config and pipeline component settings, we then use the interpolated version of nlp.config going forward, to ensure that what we're training with is up to date.
  • Resolve the [training] block of the config and perform validation, e.g. to check that the corpora are available.
  • Determine the components that should be frozen (not updated during training) or resumed (sourced components from a different pipeline that should be updated from the examples and not reset and re-initialized). To resume training, we can call the nlp.resume_training method.
  • Initialize the nlp object via nlp.initialize and pass it a get_examples callback that returns the training corpus (used for shape inference, setting up labels etc.). If the training corpus is streamed, we only provide a small sample of the data, which can potentially be infinite. nlp.initialize will delegate to the components as well and pass the data sample forward.
  • Check the listeners and warn about components dependencies, e.g. if a frozen component listens to a component that is retrained, or vice versa (which can degrade results).

2B. Initializing the nlp object

The Language.initialize method does the following:

  • Resolve the config defined in the [initialize] block separately (since everything else is already available in the loaded nlp object), based on the fully interpolated config.
  • Execute callbacks, i.e. before_init and after_init, if they're defined.
  • Initialize the vocab, including vocab data, lookup tables and vectors.
  • Initialize the tokenizer if it implements an initialize method. This is not the case for the default tokenizers, but it allows custom tokenizers to depend on external data resources that are loaded in on initialization.
  • Initialize all pipeline components if they implement an initialize method and pass them the get_examples callback, the current nlp object as well as well additional initialization config settings provided in the component-specific block.
  • Initialize pretraining if a [pretraining] block is available in the config. This allows loading pretrained tok2vec weights in spacy pretrain.
  • Register listeners if token-to-vector embedding layers of a component model "listen" to a previous component (tok2vec, transformer) in the pipeline.
  • Create an optimizer on the Language class, either by adding the optimizer passed as sgd to initialize, or by creating the optimizer defined in the config's training settings.

2C. Initializing the vocab

Vocab initialization is handled in the training.initialize.init_vocab helper. It takes the relevant loaded functions and values from the config and takes care of the following:

  • Add lookup tables defined in the config initialization, e.g. custom lemmatization tables. Those will be added to nlp.vocab.lookups from where they can be accessed by components.
  • Add JSONL-formatted vocabulary data to pre-populate the lexical attributes.
  • Load vectors into the pipeline. Vectors are defined as a name or path to a saved nlp object containing the vectors, e.g. en_vectors_web_lg. It's loaded and the vectors are ported over, while ensuring that all source strings are available in the destination strings. We also warn if there's a mismatch between sourced vectors, since this can lead to problems.