spaCy/website/docs/api/data-formats.md
Sofie Van Landeghem e796aab4b3
Resizable textcat (#7862)
* implement textcat resizing for TextCatCNN

* resizing textcat in-place

* simplify code

* ensure predictions for old textcat labels remain the same after resizing (WIP)

* fix for softmax

* store softmax as attr

* fix ensemble weight copy and cleanup

* restructure slightly

* adjust documentation, update tests and quickstart templates to use latest versions

* extend unit test slightly

* revert unnecessary edits

* fix typo

* ensemble architecture won't be resizable for now

* use resizable layer (WIP)

* revert using resizable layer

* resizable container while avoid shape inference trouble

* cleanup

* ensure model continues training after resizing

* use fill_b parameter

* use fill_defaults

* resize_layer callback

* format

* bump thinc to 8.0.4

* bump spacy-legacy to 3.0.6
2021-06-16 11:45:00 +02:00

46 KiB

title teaser menu
Data formats Details on spaCy's input and output data formats
Training Config
config
Training Data
training
Vocabulary
vocab-jsonl
Pipeline Meta
meta

This section documents input and output formats of data used by spaCy, including the training config, training data and lexical vocabulary data. For an overview of label schemes used by the models, see the models directory. Each trained pipeline documents the label schemes used in its components, depending on the data it was trained on.

Training config

Config files define the training process and pipeline and can be passed to spacy train. They use Thinc's configuration system under the hood. For details on how to use training configs, see the usage documentation. To get started with the recommended settings for your use case, check out the quickstart widget or run the init config command.

What does the @ mean?

The @ syntax lets you refer to function names registered in the function registry. For example, @architectures = "spacy.HashEmbedCNN.v2" refers to a registered function of the name spacy.HashEmbedCNN.v2 and all other values defined in its block will be passed into that function as arguments. Those arguments depend on the registered function. See the usage guide on registered functions for details.

%%GITHUB_SPACY/spacy/default_config.cfg

Under the hood, spaCy's configs are powered by our machine learning library Thinc's config system, which uses pydantic for data validation based on type hints. See spacy/schemas.py for the schemas used to validate the default config. Arguments of registered functions are validated against their type annotations, if available. To debug your config and check that it's valid, you can run the spacy debug config command.

nlp

Example

[nlp]
lang = "en"
pipeline = ["tagger", "parser", "ner"]
before_creation = null
after_creation = null
after_pipeline_creation = null
batch_size = 1000

[nlp.tokenizer]
@tokenizers = "spacy.Tokenizer.v1"

Defines the nlp object, its tokenizer and processing pipeline component names.

Name Description
lang Pipeline language ISO code. Defaults to null. str
pipeline Names of pipeline components in order. Should correspond to sections in the [components] block, e.g. [components.ner]. See docs on defining components. Defaults to []. List[str]
disabled Names of pipeline components that are loaded but disabled by default and not run as part of the pipeline. Should correspond to components listed in pipeline. After a pipeline is loaded, disabled components can be enabled using Language.enable_pipe. List[str]
before_creation Optional callback to modify Language subclass before it's initialized. Defaults to null. Optional[CallableType[Language, Type[Language]]]
after_creation Optional callback to modify nlp object right after it's initialized. Defaults to null. Optional[CallableLanguage], Language
after_pipeline_creation Optional callback to modify nlp object after the pipeline components have been added. Defaults to null. Optional[CallableLanguage], Language
tokenizer The tokenizer to use. Defaults to Tokenizer. Callable[[str], Doc]
batch_size Default batch size for Language.pipe and Language.evaluate. int

components

Example

[components.textcat]
factory = "textcat"
labels = ["POSITIVE", "NEGATIVE"]

[components.textcat.model]
@architectures = "spacy.TextCatBOW.v2"
exclusive_classes = true
ngram_size = 1
no_output_layer = false

This section includes definitions of the pipeline components and their models, if available. Components in this section can be referenced in the pipeline of the [nlp] block. Component blocks need to specify either a factory (named function to use to create component) or a source (name of path of trained pipeline to copy components from). See the docs on defining pipeline components for details.

paths, system

These sections define variables that can be referenced across the other sections as variables. For example ${paths.train} uses the value of train defined in the block [paths]. If your config includes custom registered functions that need paths, you can define them here. All config values can also be overwritten on the CLI when you run spacy train, which is especially relevant for data paths that you don't want to hard-code in your config file.

$ python -m spacy train config.cfg --paths.train ./corpus/train.spacy

corpora

Example

[corpora]

[corpora.train]
@readers = "spacy.Corpus.v1"
path = ${paths:train}

[corpora.dev]
@readers = "spacy.Corpus.v1"
path = ${paths:dev}

[corpora.pretrain]
@readers = "spacy.JsonlCorpus.v1"
path = ${paths.raw}

[corpora.my_custom_data]
@readers = "my_custom_reader.v1"

This section defines a dictionary mapping of string keys to functions. Each function takes an nlp object and yields Example objects. By default, the two keys train and dev are specified and each refer to a Corpus. When pretraining, an additional pretrain section is added that defaults to a JsonlCorpus. You can also register custom functions that return a callable.

Name Description
train Training data corpus, typically used in [training] block. CallableLanguage], Iterator[Example
dev Development data corpus, typically used in [training] block. CallableLanguage], Iterator[Example
pretrain Raw text for pretraining, typically used in [pretraining] block (if available). CallableLanguage], Iterator[Example
... Any custom or alternative corpora. CallableLanguage], Iterator[Example

Alternatively, the [corpora] block can refer to one function that returns a dictionary keyed by the corpus names. This can be useful if you want to load a single corpus once and then divide it up into train and dev partitions.

Example

[corpora]
@readers = "my_custom_reader.v1"
train_path = ${paths:train}
dev_path = ${paths:dev}
shuffle = true

Name Description
corpora A dictionary keyed by string names, mapped to corpus functions that receive the current nlp object and return an iterator of Example objects. Dict[str, CallableLanguage], Iterator[Example]

training

This section defines settings and controls for the training and evaluation process that are used when you run spacy train.

Name Description
accumulate_gradient Whether to divide the batch up into substeps. Defaults to 1. int
batcher Callable that takes an iterator of Doc objects and yields batches of Docs. Defaults to batch_by_words. CallableIterator[Doc], Iterator[List[Doc]]
before_to_disk Optional callback to modify nlp object right before it is saved to disk during and after training. Can be used to remove or reset config values or disable components. Defaults to null. Optional[CallableLanguage], Language
dev_corpus Dot notation of the config location defining the dev corpus. Defaults to corpora.dev. str
dropout The dropout rate. Defaults to 0.1. float
eval_frequency How often to evaluate during training (steps). Defaults to 200. int
frozen_components Pipeline component names that are "frozen" and shouldn't be initialized or updated during training. See here for details. Defaults to []. List[str]
annotating_components Pipeline component names that should set annotations on the predicted docs during training. See here for details. Defaults to []. List[str]
gpu_allocator Library for cupy to route GPU memory allocation to. Can be "pytorch" or "tensorflow". Defaults to variable ${system.gpu_allocator}. str
logger Callable that takes the nlp and stdout and stderr IO objects, sets up the logger, and returns two new callables to log a training step and to finalize the logger. Defaults to ConsoleLogger. CallableLanguage, IO, IO], [Tuple[Callable[[Dict[str, Any, None], Callable], None]]
max_epochs Maximum number of epochs to train for. 0 means an unlimited number of epochs. -1 means that the train corpus should be streamed rather than loaded into memory with no shuffling within the training loop. Defaults to 0. int
max_steps Maximum number of update steps to train for. 0 means an unlimited number of steps. Defaults to 20000. int
optimizer The optimizer. The learning rate schedule and other settings can be configured as part of the optimizer. Defaults to Adam. Optimizer
patience How many steps to continue without improvement in evaluation score. 0 disables early stopping. Defaults to 1600. int
score_weights Score names shown in metrics mapped to their weight towards the final weighted score. See here for details. Defaults to {}. Dict[str, float]
seed The random seed. Defaults to variable ${system.seed}. int
train_corpus Dot notation of the config location defining the train corpus. Defaults to corpora.train. str

pretraining

This section is optional and defines settings and controls for language model pretraining. It's used when you run spacy pretrain.

Name Description
max_epochs Maximum number of epochs. Defaults to 1000. int
dropout The dropout rate. Defaults to 0.2. float
n_save_every Saving frequency. Defaults to null. Optional[int]
objective The pretraining objective. Defaults to {"type": "characters", "n_characters": 4}. Dict[str, Any]
optimizer The optimizer. The learning rate schedule and other settings can be configured as part of the optimizer. Defaults to Adam. Optimizer
corpus Dot notation of the config location defining the corpus with raw text. Defaults to corpora.pretrain. str
batcher Callable that takes an iterator of Doc objects and yields batches of Docs. Defaults to batch_by_words. CallableIterator[Doc], Iterator[List[Doc]]
component Component name to identify the layer with the model to pretrain. Defaults to "tok2vec". str
layer The specific layer of the model to pretrain. If empty, the whole model will be used. str

initialize

This config block lets you define resources for initializing the pipeline. It's used by Language.initialize and typically called right before training (but not at runtime). The section allows you to specify local file paths or custom functions to load data resources from, without requiring them at runtime when you load the trained pipeline back in. Also see the usage guides on the config lifecycle and custom initialization.

Example

[initialize]
vectors = "/path/to/vectors_nlp"
init_tok2vec = "/path/to/pretrain.bin"

[initialize_components]

[initialize.components.my_component]
data_path = "/path/to/component_data"
Name Description
after_init Optional callback to modify the nlp object after initialization. Optional[CallableLanguage], Language
before_init Optional callback to modify the nlp object before initialization. Optional[CallableLanguage], Language
components Additional arguments passed to the initialize method of a pipeline component, keyed by component name. If type annotations are available on the method, the config will be validated against them. The initialize methods will always receive the get_examples callback and the current nlp object. Dict[str, Dict[str, Any]]
init_tok2vec Optional path to pretrained tok2vec weights created with spacy pretrain. Defaults to variable ${paths.init_tok2vec}. Optional[str]
lookups Additional lexeme and vocab data from spacy-lookups-data. Defaults to null. Optional[Lookups]
tokenizer Additional arguments passed to the initialize method of the specified tokenizer. Can be used for languages like Chinese that depend on dictionaries or trained models for tokenization. If type annotations are available on the method, the config will be validated against them. The initialize method will always receive the get_examples callback and the current nlp object. Dict[str, Any]
vectors Name or path of pipeline containing pretrained word vectors to use, e.g. created with init vectors. Defaults to null. Optional[str]
vocab_data Path to JSONL-formatted vocabulary file to initialize vocabulary. Optional[str]

Training data

Binary training format

Example

from spacy.tokens import DocBin
from spacy.training import Corpus

doc_bin = DocBin(docs=docs)
doc_bin.to_disk("./data.spacy")
reader = Corpus("./data.spacy")

The main data format used in spaCy v3.0 is a binary format created by serializing a DocBin, which represents a collection of Doc objects. This means that you can train spaCy pipelines using the same format it outputs: annotated Doc objects. The binary format is extremely efficient in storage, especially when packing multiple documents together.

Typically, the extension for these binary files is .spacy, and they are used as input format for specifying a training corpus and for spaCy's CLI train command. The built-in convert command helps you convert spaCy's previous JSON format to the new binary format. It also supports conversion of the .conllu format used by the Universal Dependencies corpora.

JSON training format

As of v3.0, the JSON input format is deprecated and is replaced by the binary format. Instead of converting Doc objects to JSON, you can now serialize them directly using the DocBin container and then use them as input data.

spacy convert lets you convert your JSON data to the new .spacy format:

$ python -m spacy convert ./data.json ./output.spacy

Annotating entities

Named entities are provided in the BILUO notation. Tokens outside an entity are set to "O" and tokens that are part of an entity are set to the entity label, prefixed by the BILUO marker. For example "B-ORG" describes the first token of a multi-token ORG entity and "U-PERSON" a single token representing a PERSON entity. The offsets_to_biluo_tags function can help you convert entity offsets to the right format.

### Example structure
[{
    "id": int,                      # ID of the document within the corpus
    "paragraphs": [{                # list of paragraphs in the corpus
        "raw": string,              # raw text of the paragraph
        "sentences": [{             # list of sentences in the paragraph
            "tokens": [{            # list of tokens in the sentence
                "id": int,          # index of the token in the document
                "dep": string,      # dependency label
                "head": int,        # offset of token head relative to token index
                "tag": string,      # part-of-speech tag
                "orth": string,     # verbatim text of the token
                "ner": string       # BILUO label, e.g. "O" or "B-ORG"
            }],
            "brackets": [{          # phrase structure (NOT USED by current models)
                "first": int,       # index of first token
                "last": int,        # index of last token
                "label": string     # phrase label
            }]
        }],
        "cats": [{                  # new in v2.2: categories for text classifier
            "label": string,        # text category label
            "value": float / bool   # label applies (1.0/true) or not (0.0/false)
        }]
    }]
}]

Here's an example of dependencies, part-of-speech tags and named entities, taken from the English Wall Street Journal portion of the Penn Treebank:

https://github.com/explosion/spaCy/blob/v2.3.x/examples/training/training-data.json

Annotation format for creating training examples

An Example object holds the information for one training instance. It stores two Doc objects: one for holding the gold-standard reference data, and one for holding the predictions of the pipeline. Examples can be created using the Example.from_dict method with a reference Doc and a dictionary of gold-standard annotations.

Example

example = Example.from_dict(doc, gold_dict)

Example objects are used as part of the internal training API and they're expected when you call nlp.update. However, for most use cases, you shouldn't have to write your own training scripts. It's recommended to train your pipelines via the spacy train command with a config file to keep track of your settings and hyperparameters and your own registered functions to customize the setup.

Example

{
   "text": str,
   "words": List[str],
   "lemmas": List[str],
   "spaces": List[bool],
   "tags": List[str],
   "pos": List[str],
   "morphs": List[str],
   "sent_starts": List[Optional[bool]],
   "deps": List[string],
   "heads": List[int],
   "entities": List[str],
   "entities": List[(int, int, str)],
   "cats": Dict[str, float],
   "links": Dict[(int, int), dict],
}
Name Description
text Raw text. str
words List of gold-standard tokens. List[str]
lemmas List of lemmas. List[str]
spaces List of boolean values indicating whether the corresponding tokens is followed by a space or not. List[bool]
tags List of fine-grained POS tags. List[str]
pos List of coarse-grained POS tags. List[str]
morphs List of morphological features. List[str]
sent_starts List of boolean values indicating whether each token is the first of a sentence or not. List[bool]
deps List of string values indicating the dependency relation of a token to its head. List[str]
heads List of integer values indicating the dependency head of each token, referring to the absolute index of each token in the text. List[int]
entities Option 1: List of BILUO tags per token of the format "{action}-{label}", or None for unannotated tokens. List[str]
entities Option 2: List of "(start, end, label)" tuples defining all entities in the text. List[Tuple[int, int, str]]
cats Dictionary of label/value pairs indicating how relevant a certain text category is for the text. Dict[str, float]
links Dictionary of offset/dict pairs defining named entity links. The character offsets are linked to a dictionary of relevant knowledge base IDs. Dict[Tuple[int, int], Dict]
  • Multiple formats are possible for the "entities" entry, but you have to pick one.
  • Any values for sentence starts will be ignored if there are annotations for dependency relations.
  • If the dictionary contains values for "text" and "words", but not "spaces", the latter are inferred automatically. If "words" is not provided either, the values are inferred from the Doc argument.
### Examples
# Training data for a part-of-speech tagger
doc = Doc(vocab, words=["I", "like", "stuff"])
gold_dict = {"tags": ["NOUN", "VERB", "NOUN"]}
example = Example.from_dict(doc, gold_dict)

# Training data for an entity recognizer (option 1)
doc = nlp("Laura flew to Silicon Valley.")
gold_dict = {"entities": ["U-PERS", "O", "O", "B-LOC", "L-LOC"]}
example = Example.from_dict(doc, gold_dict)

# Training data for an entity recognizer (option 2)
doc = nlp("Laura flew to Silicon Valley.")
gold_dict = {"entities": [(0, 5, "PERSON"), (14, 28, "LOC")]}
example = Example.from_dict(doc, gold_dict)

# Training data for text categorization
doc = nlp("I'm pretty happy about that!")
gold_dict = {"cats": {"POSITIVE": 1.0, "NEGATIVE": 0.0}}
example = Example.from_dict(doc, gold_dict)

# Training data for an Entity Linking component (also requires entities & sentences)
doc = nlp("Russ Cochran his reprints include EC Comics.")
gold_dict = {"entities": [(0, 12, "PERSON")],
             "links": {(0, 12): {"Q7381115": 1.0, "Q2146908": 0.0}},
             "sent_starts": [1, -1, -1, -1, -1, -1, -1, -1]}
example = Example.from_dict(doc, gold_dict)

Lexical data for vocabulary

This data file can be provided via the vocab_data setting in the [initialize] block of the training config to pre-define the lexical data to initialize the nlp object's vocabulary with. The file should contain one lexical entry per line. The first line defines the language and vocabulary settings. All other lines are expected to be JSON objects describing an individual lexeme. The lexical attributes will be then set as attributes on spaCy's Lexeme object.

Example config

[initialize]
vocab_data = "/path/to/vocab-data.jsonl"
### First line
{"lang": "en", "settings": {"oov_prob": -20.502029418945312}}
### Entry structure
{
    "orth": string,     # the word text
    "id": int,          # can correspond to row in vectors table
    "lower": string,
    "norm": string,
    "shape": string
    "prefix": string,
    "suffix": string,
    "length": int,
    "cluster": string,
    "prob": float,
    "is_alpha": bool,
    "is_ascii": bool,
    "is_digit": bool,
    "is_lower": bool,
    "is_punct": bool,
    "is_space": bool,
    "is_title": bool,
    "is_upper": bool,
    "like_url": bool,
    "like_num": bool,
    "like_email": bool,
    "is_stop": bool,
    "is_oov": bool,
    "is_quote": bool,
    "is_left_punct": bool,
    "is_right_punct": bool
}

Here's an example of the 20 most frequent lexemes in the English training data:

%%GITHUB_SPACY/extra/example_data/vocab-data.jsonl

Pipeline meta

The pipeline meta is available as the file meta.json and exported automatically when you save an nlp object to disk. Its contents are available as nlp.meta.

As of spaCy v3.0, the meta.json isn't used to construct the language class and pipeline anymore and only contains meta information for reference and for creating a Python package with spacy package. How to set up the nlp object is now defined in the config.cfg, which includes detailed information about the pipeline components and their model architectures, and all other settings and hyperparameters used to train the pipeline. It's the single source of truth used for loading a pipeline.

Example

{
  "name": "example_pipeline",
  "lang": "en",
  "version": "1.0.0",
  "spacy_version": ">=3.0.0,<3.1.0",
  "parent_package": "spacy",
  "requirements": ["spacy-transformers>=1.0.0,<1.1.0"],
  "description": "Example pipeline for spaCy",
  "author": "You",
  "email": "you@example.com",
  "url": "https://example.com",
  "license": "CC BY-SA 3.0",
  "sources": [{ "name": "My Corpus", "license": "MIT" }],
  "vectors": { "width": 0, "vectors": 0, "keys": 0, "name": null },
  "pipeline": ["tok2vec", "ner", "textcat"],
  "labels": {
    "ner": ["PERSON", "ORG", "PRODUCT"],
    "textcat": ["POSITIVE", "NEGATIVE"]
  },
  "performance": {
    "ents_f": 82.7300930714,
    "ents_p": 82.135523614,
    "ents_r": 83.3333333333,
    "textcat_score": 88.364323811
  },
  "speed": { "cpu": 7667.8, "gpu": null, "nwords": 10329 },
  "spacy_git_version": "61dfdd9fb"
}
Name Description
lang Pipeline language ISO code. Defaults to "en". str
name Pipeline name, e.g. "core_web_sm". The final package name will be {lang}_{name}. Defaults to "pipeline". str
version Pipeline version. Will be used to version a Python package created with spacy package. Defaults to "0.0.0". str
spacy_version spaCy version range the package is compatible with. Defaults to the spaCy version used to create the pipeline, up to next minor version, which is the default compatibility for the available trained pipelines. For instance, a pipeline trained with v3.0.0 will have the version range ">=3.0.0,<3.1.0". str
parent_package Name of the spaCy package. Typically "spacy" or "spacy_nightly". Defaults to "spacy". str
requirements Python package requirements that the pipeline depends on. Will be used for the Python package setup in spacy package. Should be a list of package names with optional version specifiers, just like you'd define them in a setup.cfg or requirements.txt. Defaults to []. List[str]
description Pipeline description. Also used for Python package. Defaults to "". str
author Pipeline author name. Also used for Python package. Defaults to "". str
email Pipeline author email. Also used for Python package. Defaults to "". str
url Pipeline author URL. Also used for Python package. Defaults to "". str
license Pipeline license. Also used for Python package. Defaults to "". str
sources Data sources used to train the pipeline. Typically a list of dicts with the keys "name", "url", "author" and "license". See here for examples. Defaults to None. Optional[List[Dict[str, str]]]
vectors Information about the word vectors included with the pipeline. Typically a dict with the keys "width", "vectors" (number of vectors), "keys" and "name". Dict[str, Any]
pipeline Names of pipeline component names, in order. Corresponds to nlp.pipe_names. Only exists for reference and is not used to create the components. This information is defined in the config.cfg. Defaults to []. List[str]
labels Label schemes of the trained pipeline components, keyed by component name. Corresponds to nlp.pipe_labels. See here for examples. Defaults to {}. Dict[str, Dict[str, List[str]]]
performance Training accuracy, added automatically by spacy train. Dictionary of score names mapped to scores. Defaults to {}. Dict[str, Union[float, Dict[str, float]]]
speed Inference speed, added automatically by spacy train. Typically a dictionary with the keys "cpu", "gpu" and "nwords" (words per second). Defaults to {}. Dict[str, Optional[Union[float, str]]]
spacy_git_version 3 Git commit of spacy used to create pipeline. str
other Any other custom meta information you want to add. The data is preserved in nlp.meta. Any