spaCy/spacy/scorer.py

798 lines
31 KiB
Python
Raw Normal View History

2020-07-28 22:39:42 +03:00
from typing import Optional, Iterable, Dict, Any, Callable, Tuple, TYPE_CHECKING
Add textcat to train CLI (#4226) * Add doc.cats to spacy.gold at the paragraph level Support `doc.cats` as `"cats": [{"label": string, "value": number}]` in the spacy JSON training format at the paragraph level. * `spacy.gold.docs_to_json()` writes `docs.cats` * `GoldCorpus` reads in cats in each `GoldParse` * Update instances of gold_tuples to handle cats Update iteration over gold_tuples / gold_parses to handle addition of cats at the paragraph level. * Add textcat to train CLI * Add textcat options to train CLI * Add textcat labels in `TextCategorizer.begin_training()` * Add textcat evaluation to `Scorer`: * For binary exclusive classes with provided label: F1 for label * For 2+ exclusive classes: F1 macro average * For multilabel (not exclusive): ROC AUC macro average (currently relying on sklearn) * Provide user info on textcat evaluation settings, potential incompatibilities * Provide pipeline to Scorer in `Language.evaluate` for textcat config * Customize train CLI output to include only metrics relevant to current pipeline * Add textcat evaluation to evaluate CLI * Fix handling of unset arguments and config params Fix handling of unset arguments and model confiug parameters in Scorer initialization. * Temporarily add sklearn requirement * Remove sklearn version number * Improve Scorer handling of models without textcats * Fixing Scorer handling of models without textcats * Update Scorer output for python 2.7 * Modify inf in Scorer for python 2.7 * Auto-format Also make small adjustments to make auto-formatting with black easier and produce nicer results * Move error message to Errors * Update documentation * Add cats to annotation JSON format [ci skip] * Fix tpl flag and docs [ci skip] * Switch to internal roc_auc_score Switch to internal `roc_auc_score()` adapted from scikit-learn. * Add AUCROCScore tests and improve errors/warnings * Add tests for AUCROCScore and roc_auc_score * Add missing error for only positive/negative values * Remove unnecessary warnings and errors * Make reduced roc_auc_score functions private Because most of the checks and warnings have been stripped for the internal functions and access is only intended through `ROCAUCScore`, make the functions for roc_auc_score adapted from scikit-learn private. * Check that data corresponds with multilabel flag Check that the training instances correspond with the multilabel flag, adding the multilabel flag if required. * Add textcat score to early stopping check * Add more checks to debug-data for textcat * Add example training data for textcat * Add more checks to textcat train CLI * Check configuration when extending base model * Fix typos * Update textcat example data * Provide licensing details and licenses for data * Remove two labels with no positive instances from jigsaw-toxic-comment data. Co-authored-by: Ines Montani <ines@ines.io>
2019-09-15 23:31:31 +03:00
import numpy as np
2020-07-28 22:39:42 +03:00
from .gold import Example
2020-08-17 17:45:24 +03:00
from .tokens import Token, Doc, Span
Add textcat to train CLI (#4226) * Add doc.cats to spacy.gold at the paragraph level Support `doc.cats` as `"cats": [{"label": string, "value": number}]` in the spacy JSON training format at the paragraph level. * `spacy.gold.docs_to_json()` writes `docs.cats` * `GoldCorpus` reads in cats in each `GoldParse` * Update instances of gold_tuples to handle cats Update iteration over gold_tuples / gold_parses to handle addition of cats at the paragraph level. * Add textcat to train CLI * Add textcat options to train CLI * Add textcat labels in `TextCategorizer.begin_training()` * Add textcat evaluation to `Scorer`: * For binary exclusive classes with provided label: F1 for label * For 2+ exclusive classes: F1 macro average * For multilabel (not exclusive): ROC AUC macro average (currently relying on sklearn) * Provide user info on textcat evaluation settings, potential incompatibilities * Provide pipeline to Scorer in `Language.evaluate` for textcat config * Customize train CLI output to include only metrics relevant to current pipeline * Add textcat evaluation to evaluate CLI * Fix handling of unset arguments and config params Fix handling of unset arguments and model confiug parameters in Scorer initialization. * Temporarily add sklearn requirement * Remove sklearn version number * Improve Scorer handling of models without textcats * Fixing Scorer handling of models without textcats * Update Scorer output for python 2.7 * Modify inf in Scorer for python 2.7 * Auto-format Also make small adjustments to make auto-formatting with black easier and produce nicer results * Move error message to Errors * Update documentation * Add cats to annotation JSON format [ci skip] * Fix tpl flag and docs [ci skip] * Switch to internal roc_auc_score Switch to internal `roc_auc_score()` adapted from scikit-learn. * Add AUCROCScore tests and improve errors/warnings * Add tests for AUCROCScore and roc_auc_score * Add missing error for only positive/negative values * Remove unnecessary warnings and errors * Make reduced roc_auc_score functions private Because most of the checks and warnings have been stripped for the internal functions and access is only intended through `ROCAUCScore`, make the functions for roc_auc_score adapted from scikit-learn private. * Check that data corresponds with multilabel flag Check that the training instances correspond with the multilabel flag, adding the multilabel flag if required. * Add textcat score to early stopping check * Add more checks to debug-data for textcat * Add example training data for textcat * Add more checks to textcat train CLI * Check configuration when extending base model * Fix typos * Update textcat example data * Provide licensing details and licenses for data * Remove two labels with no positive instances from jigsaw-toxic-comment data. Co-authored-by: Ines Montani <ines@ines.io>
2019-09-15 23:31:31 +03:00
from .errors import Errors
Refactor the Scorer to improve flexibility (#5731) * Refactor the Scorer to improve flexibility Refactor the `Scorer` to improve flexibility for arbitrary pipeline components. * Individual pipeline components provide their own `evaluate` methods that score a list of `Example`s and return a dictionary of scores * `Scorer` is initialized either: * with a provided pipeline containing components to be scored * with a default pipeline containing the built-in statistical components (senter, tagger, morphologizer, parser, ner) * `Scorer.score` evaluates a list of `Example`s and returns a dictionary of scores referring to the scores provided by the components in the pipeline Significant differences: * `tags_acc` is renamed to `tag_acc` to be consistent with `token_acc` and the new `morph_acc`, `pos_acc`, and `lemma_acc` * Scoring is no longer cumulative: `Scorer.score` scores a list of examples rather than a single example and does not retain any state about previously scored examples * PRF values in the returned scores are no longer multiplied by 100 * Add kwargs to Morphologizer.evaluate * Create generalized scoring methods in Scorer * Generalized static scoring methods are added to `Scorer` * Methods require an attribute (either on Token or Doc) that is used to key the returned scores Naming differences: * `uas`, `las`, and `las_per_type` in the scores dict are renamed to `dep_uas`, `dep_las`, and `dep_las_per_type` Scoring differences: * `Doc.sents` is now scored as spans rather than on sentence-initial token positions so that `Doc.sents` and `Doc.ents` can be scored with the same method (this lowers scores since a single incorrect sentence start results in two incorrect spans) * Simplify / extend hasattr check for eval method * Add hasattr check to tokenizer scoring * Simplify to hasattr check for component scoring * Reset Example alignment if docs are set Reset the Example alignment if either doc is set in case the tokenization has changed. * Add PRF tokenization scoring for tokens as spans Add PRF scores for tokens as character spans. The scores are: * token_acc: # correct tokens / # gold tokens * token_p/r/f: PRF for (token.idx, token.idx + len(token)) * Add docstring to Scorer.score_tokenization * Rename component.evaluate() to component.score() * Update Scorer API docs * Update scoring for positive_label in textcat * Fix TextCategorizer.score kwargs * Update Language.evaluate docs * Update score names in default config
2020-07-25 13:53:02 +03:00
from .util import get_lang_class
from .morphology import Morphology
2015-05-27 04:18:16 +03:00
2020-07-28 22:39:42 +03:00
if TYPE_CHECKING:
# This lets us add type hints for mypy etc. without causing circular imports
from .language import Language # noqa: F401
DEFAULT_PIPELINE = ["senter", "tagger", "morphologizer", "parser", "ner", "textcat"]
2015-04-05 23:29:30 +03:00
2020-07-12 15:03:23 +03:00
class PRFScore:
2020-07-28 22:39:42 +03:00
"""A precision / recall / F score."""
2020-07-28 22:39:42 +03:00
def __init__(self) -> None:
self.tp = 0
self.fp = 0
self.fn = 0
2020-07-28 22:39:42 +03:00
def score_set(self, cand: set, gold: set) -> None:
self.tp += len(cand.intersection(gold))
self.fp += len(cand - gold)
self.fn += len(gold - cand)
@property
2020-07-28 22:39:42 +03:00
def precision(self) -> float:
return self.tp / (self.tp + self.fp + 1e-100)
@property
2020-07-28 22:39:42 +03:00
def recall(self) -> float:
return self.tp / (self.tp + self.fn + 1e-100)
@property
2020-07-28 22:39:42 +03:00
def fscore(self) -> float:
p = self.precision
r = self.recall
return 2 * ((p * r) / (p + r + 1e-100))
2020-07-28 22:39:42 +03:00
def to_dict(self) -> Dict[str, float]:
Refactor the Scorer to improve flexibility (#5731) * Refactor the Scorer to improve flexibility Refactor the `Scorer` to improve flexibility for arbitrary pipeline components. * Individual pipeline components provide their own `evaluate` methods that score a list of `Example`s and return a dictionary of scores * `Scorer` is initialized either: * with a provided pipeline containing components to be scored * with a default pipeline containing the built-in statistical components (senter, tagger, morphologizer, parser, ner) * `Scorer.score` evaluates a list of `Example`s and returns a dictionary of scores referring to the scores provided by the components in the pipeline Significant differences: * `tags_acc` is renamed to `tag_acc` to be consistent with `token_acc` and the new `morph_acc`, `pos_acc`, and `lemma_acc` * Scoring is no longer cumulative: `Scorer.score` scores a list of examples rather than a single example and does not retain any state about previously scored examples * PRF values in the returned scores are no longer multiplied by 100 * Add kwargs to Morphologizer.evaluate * Create generalized scoring methods in Scorer * Generalized static scoring methods are added to `Scorer` * Methods require an attribute (either on Token or Doc) that is used to key the returned scores Naming differences: * `uas`, `las`, and `las_per_type` in the scores dict are renamed to `dep_uas`, `dep_las`, and `dep_las_per_type` Scoring differences: * `Doc.sents` is now scored as spans rather than on sentence-initial token positions so that `Doc.sents` and `Doc.ents` can be scored with the same method (this lowers scores since a single incorrect sentence start results in two incorrect spans) * Simplify / extend hasattr check for eval method * Add hasattr check to tokenizer scoring * Simplify to hasattr check for component scoring * Reset Example alignment if docs are set Reset the Example alignment if either doc is set in case the tokenization has changed. * Add PRF tokenization scoring for tokens as spans Add PRF scores for tokens as character spans. The scores are: * token_acc: # correct tokens / # gold tokens * token_p/r/f: PRF for (token.idx, token.idx + len(token)) * Add docstring to Scorer.score_tokenization * Rename component.evaluate() to component.score() * Update Scorer API docs * Update scoring for positive_label in textcat * Fix TextCategorizer.score kwargs * Update Language.evaluate docs * Update score names in default config
2020-07-25 13:53:02 +03:00
return {"p": self.precision, "r": self.recall, "f": self.fscore}
2020-07-12 15:03:23 +03:00
class ROCAUCScore:
2020-07-28 22:39:42 +03:00
"""An AUC ROC score."""
Add textcat to train CLI (#4226) * Add doc.cats to spacy.gold at the paragraph level Support `doc.cats` as `"cats": [{"label": string, "value": number}]` in the spacy JSON training format at the paragraph level. * `spacy.gold.docs_to_json()` writes `docs.cats` * `GoldCorpus` reads in cats in each `GoldParse` * Update instances of gold_tuples to handle cats Update iteration over gold_tuples / gold_parses to handle addition of cats at the paragraph level. * Add textcat to train CLI * Add textcat options to train CLI * Add textcat labels in `TextCategorizer.begin_training()` * Add textcat evaluation to `Scorer`: * For binary exclusive classes with provided label: F1 for label * For 2+ exclusive classes: F1 macro average * For multilabel (not exclusive): ROC AUC macro average (currently relying on sklearn) * Provide user info on textcat evaluation settings, potential incompatibilities * Provide pipeline to Scorer in `Language.evaluate` for textcat config * Customize train CLI output to include only metrics relevant to current pipeline * Add textcat evaluation to evaluate CLI * Fix handling of unset arguments and config params Fix handling of unset arguments and model confiug parameters in Scorer initialization. * Temporarily add sklearn requirement * Remove sklearn version number * Improve Scorer handling of models without textcats * Fixing Scorer handling of models without textcats * Update Scorer output for python 2.7 * Modify inf in Scorer for python 2.7 * Auto-format Also make small adjustments to make auto-formatting with black easier and produce nicer results * Move error message to Errors * Update documentation * Add cats to annotation JSON format [ci skip] * Fix tpl flag and docs [ci skip] * Switch to internal roc_auc_score Switch to internal `roc_auc_score()` adapted from scikit-learn. * Add AUCROCScore tests and improve errors/warnings * Add tests for AUCROCScore and roc_auc_score * Add missing error for only positive/negative values * Remove unnecessary warnings and errors * Make reduced roc_auc_score functions private Because most of the checks and warnings have been stripped for the internal functions and access is only intended through `ROCAUCScore`, make the functions for roc_auc_score adapted from scikit-learn private. * Check that data corresponds with multilabel flag Check that the training instances correspond with the multilabel flag, adding the multilabel flag if required. * Add textcat score to early stopping check * Add more checks to debug-data for textcat * Add example training data for textcat * Add more checks to textcat train CLI * Check configuration when extending base model * Fix typos * Update textcat example data * Provide licensing details and licenses for data * Remove two labels with no positive instances from jigsaw-toxic-comment data. Co-authored-by: Ines Montani <ines@ines.io>
2019-09-15 23:31:31 +03:00
2020-07-29 11:39:33 +03:00
def __init__(self) -> None:
Add textcat to train CLI (#4226) * Add doc.cats to spacy.gold at the paragraph level Support `doc.cats` as `"cats": [{"label": string, "value": number}]` in the spacy JSON training format at the paragraph level. * `spacy.gold.docs_to_json()` writes `docs.cats` * `GoldCorpus` reads in cats in each `GoldParse` * Update instances of gold_tuples to handle cats Update iteration over gold_tuples / gold_parses to handle addition of cats at the paragraph level. * Add textcat to train CLI * Add textcat options to train CLI * Add textcat labels in `TextCategorizer.begin_training()` * Add textcat evaluation to `Scorer`: * For binary exclusive classes with provided label: F1 for label * For 2+ exclusive classes: F1 macro average * For multilabel (not exclusive): ROC AUC macro average (currently relying on sklearn) * Provide user info on textcat evaluation settings, potential incompatibilities * Provide pipeline to Scorer in `Language.evaluate` for textcat config * Customize train CLI output to include only metrics relevant to current pipeline * Add textcat evaluation to evaluate CLI * Fix handling of unset arguments and config params Fix handling of unset arguments and model confiug parameters in Scorer initialization. * Temporarily add sklearn requirement * Remove sklearn version number * Improve Scorer handling of models without textcats * Fixing Scorer handling of models without textcats * Update Scorer output for python 2.7 * Modify inf in Scorer for python 2.7 * Auto-format Also make small adjustments to make auto-formatting with black easier and produce nicer results * Move error message to Errors * Update documentation * Add cats to annotation JSON format [ci skip] * Fix tpl flag and docs [ci skip] * Switch to internal roc_auc_score Switch to internal `roc_auc_score()` adapted from scikit-learn. * Add AUCROCScore tests and improve errors/warnings * Add tests for AUCROCScore and roc_auc_score * Add missing error for only positive/negative values * Remove unnecessary warnings and errors * Make reduced roc_auc_score functions private Because most of the checks and warnings have been stripped for the internal functions and access is only intended through `ROCAUCScore`, make the functions for roc_auc_score adapted from scikit-learn private. * Check that data corresponds with multilabel flag Check that the training instances correspond with the multilabel flag, adding the multilabel flag if required. * Add textcat score to early stopping check * Add more checks to debug-data for textcat * Add example training data for textcat * Add more checks to textcat train CLI * Check configuration when extending base model * Fix typos * Update textcat example data * Provide licensing details and licenses for data * Remove two labels with no positive instances from jigsaw-toxic-comment data. Co-authored-by: Ines Montani <ines@ines.io>
2019-09-15 23:31:31 +03:00
self.golds = []
self.cands = []
self.saved_score = 0.0
self.saved_score_at_len = 0
2020-07-28 22:39:42 +03:00
def score_set(self, cand, gold) -> None:
Add textcat to train CLI (#4226) * Add doc.cats to spacy.gold at the paragraph level Support `doc.cats` as `"cats": [{"label": string, "value": number}]` in the spacy JSON training format at the paragraph level. * `spacy.gold.docs_to_json()` writes `docs.cats` * `GoldCorpus` reads in cats in each `GoldParse` * Update instances of gold_tuples to handle cats Update iteration over gold_tuples / gold_parses to handle addition of cats at the paragraph level. * Add textcat to train CLI * Add textcat options to train CLI * Add textcat labels in `TextCategorizer.begin_training()` * Add textcat evaluation to `Scorer`: * For binary exclusive classes with provided label: F1 for label * For 2+ exclusive classes: F1 macro average * For multilabel (not exclusive): ROC AUC macro average (currently relying on sklearn) * Provide user info on textcat evaluation settings, potential incompatibilities * Provide pipeline to Scorer in `Language.evaluate` for textcat config * Customize train CLI output to include only metrics relevant to current pipeline * Add textcat evaluation to evaluate CLI * Fix handling of unset arguments and config params Fix handling of unset arguments and model confiug parameters in Scorer initialization. * Temporarily add sklearn requirement * Remove sklearn version number * Improve Scorer handling of models without textcats * Fixing Scorer handling of models without textcats * Update Scorer output for python 2.7 * Modify inf in Scorer for python 2.7 * Auto-format Also make small adjustments to make auto-formatting with black easier and produce nicer results * Move error message to Errors * Update documentation * Add cats to annotation JSON format [ci skip] * Fix tpl flag and docs [ci skip] * Switch to internal roc_auc_score Switch to internal `roc_auc_score()` adapted from scikit-learn. * Add AUCROCScore tests and improve errors/warnings * Add tests for AUCROCScore and roc_auc_score * Add missing error for only positive/negative values * Remove unnecessary warnings and errors * Make reduced roc_auc_score functions private Because most of the checks and warnings have been stripped for the internal functions and access is only intended through `ROCAUCScore`, make the functions for roc_auc_score adapted from scikit-learn private. * Check that data corresponds with multilabel flag Check that the training instances correspond with the multilabel flag, adding the multilabel flag if required. * Add textcat score to early stopping check * Add more checks to debug-data for textcat * Add example training data for textcat * Add more checks to textcat train CLI * Check configuration when extending base model * Fix typos * Update textcat example data * Provide licensing details and licenses for data * Remove two labels with no positive instances from jigsaw-toxic-comment data. Co-authored-by: Ines Montani <ines@ines.io>
2019-09-15 23:31:31 +03:00
self.cands.append(cand)
self.golds.append(gold)
@property
def score(self):
if len(self.golds) == self.saved_score_at_len:
return self.saved_score
try:
self.saved_score = _roc_auc_score(self.golds, self.cands)
# catch ValueError: Only one class present in y_true.
# ROC AUC score is not defined in that case.
2019-09-18 20:57:08 +03:00
except ValueError:
Add textcat to train CLI (#4226) * Add doc.cats to spacy.gold at the paragraph level Support `doc.cats` as `"cats": [{"label": string, "value": number}]` in the spacy JSON training format at the paragraph level. * `spacy.gold.docs_to_json()` writes `docs.cats` * `GoldCorpus` reads in cats in each `GoldParse` * Update instances of gold_tuples to handle cats Update iteration over gold_tuples / gold_parses to handle addition of cats at the paragraph level. * Add textcat to train CLI * Add textcat options to train CLI * Add textcat labels in `TextCategorizer.begin_training()` * Add textcat evaluation to `Scorer`: * For binary exclusive classes with provided label: F1 for label * For 2+ exclusive classes: F1 macro average * For multilabel (not exclusive): ROC AUC macro average (currently relying on sklearn) * Provide user info on textcat evaluation settings, potential incompatibilities * Provide pipeline to Scorer in `Language.evaluate` for textcat config * Customize train CLI output to include only metrics relevant to current pipeline * Add textcat evaluation to evaluate CLI * Fix handling of unset arguments and config params Fix handling of unset arguments and model confiug parameters in Scorer initialization. * Temporarily add sklearn requirement * Remove sklearn version number * Improve Scorer handling of models without textcats * Fixing Scorer handling of models without textcats * Update Scorer output for python 2.7 * Modify inf in Scorer for python 2.7 * Auto-format Also make small adjustments to make auto-formatting with black easier and produce nicer results * Move error message to Errors * Update documentation * Add cats to annotation JSON format [ci skip] * Fix tpl flag and docs [ci skip] * Switch to internal roc_auc_score Switch to internal `roc_auc_score()` adapted from scikit-learn. * Add AUCROCScore tests and improve errors/warnings * Add tests for AUCROCScore and roc_auc_score * Add missing error for only positive/negative values * Remove unnecessary warnings and errors * Make reduced roc_auc_score functions private Because most of the checks and warnings have been stripped for the internal functions and access is only intended through `ROCAUCScore`, make the functions for roc_auc_score adapted from scikit-learn private. * Check that data corresponds with multilabel flag Check that the training instances correspond with the multilabel flag, adding the multilabel flag if required. * Add textcat score to early stopping check * Add more checks to debug-data for textcat * Add example training data for textcat * Add more checks to textcat train CLI * Check configuration when extending base model * Fix typos * Update textcat example data * Provide licensing details and licenses for data * Remove two labels with no positive instances from jigsaw-toxic-comment data. Co-authored-by: Ines Montani <ines@ines.io>
2019-09-15 23:31:31 +03:00
self.saved_score = -float("inf")
self.saved_score_at_len = len(self.golds)
return self.saved_score
2020-07-12 15:03:23 +03:00
class Scorer:
2019-05-24 15:06:04 +03:00
"""Compute evaluation scores."""
2020-07-28 22:39:42 +03:00
def __init__(
self,
nlp: Optional["Language"] = None,
default_lang: str = "xx",
default_pipeline=DEFAULT_PIPELINE,
**cfg,
) -> None:
2019-05-24 15:06:04 +03:00
"""Initialize the Scorer.
DOCS: https://spacy.io/api/scorer#init
"""
Refactor the Scorer to improve flexibility (#5731) * Refactor the Scorer to improve flexibility Refactor the `Scorer` to improve flexibility for arbitrary pipeline components. * Individual pipeline components provide their own `evaluate` methods that score a list of `Example`s and return a dictionary of scores * `Scorer` is initialized either: * with a provided pipeline containing components to be scored * with a default pipeline containing the built-in statistical components (senter, tagger, morphologizer, parser, ner) * `Scorer.score` evaluates a list of `Example`s and returns a dictionary of scores referring to the scores provided by the components in the pipeline Significant differences: * `tags_acc` is renamed to `tag_acc` to be consistent with `token_acc` and the new `morph_acc`, `pos_acc`, and `lemma_acc` * Scoring is no longer cumulative: `Scorer.score` scores a list of examples rather than a single example and does not retain any state about previously scored examples * PRF values in the returned scores are no longer multiplied by 100 * Add kwargs to Morphologizer.evaluate * Create generalized scoring methods in Scorer * Generalized static scoring methods are added to `Scorer` * Methods require an attribute (either on Token or Doc) that is used to key the returned scores Naming differences: * `uas`, `las`, and `las_per_type` in the scores dict are renamed to `dep_uas`, `dep_las`, and `dep_las_per_type` Scoring differences: * `Doc.sents` is now scored as spans rather than on sentence-initial token positions so that `Doc.sents` and `Doc.ents` can be scored with the same method (this lowers scores since a single incorrect sentence start results in two incorrect spans) * Simplify / extend hasattr check for eval method * Add hasattr check to tokenizer scoring * Simplify to hasattr check for component scoring * Reset Example alignment if docs are set Reset the Example alignment if either doc is set in case the tokenization has changed. * Add PRF tokenization scoring for tokens as spans Add PRF scores for tokens as character spans. The scores are: * token_acc: # correct tokens / # gold tokens * token_p/r/f: PRF for (token.idx, token.idx + len(token)) * Add docstring to Scorer.score_tokenization * Rename component.evaluate() to component.score() * Update Scorer API docs * Update scoring for positive_label in textcat * Fix TextCategorizer.score kwargs * Update Language.evaluate docs * Update score names in default config
2020-07-25 13:53:02 +03:00
self.nlp = nlp
self.cfg = cfg
if not nlp:
2020-07-28 22:39:42 +03:00
nlp = get_lang_class(default_lang)()
for pipe in default_pipeline:
nlp.add_pipe(pipe)
Refactor the Scorer to improve flexibility (#5731) * Refactor the Scorer to improve flexibility Refactor the `Scorer` to improve flexibility for arbitrary pipeline components. * Individual pipeline components provide their own `evaluate` methods that score a list of `Example`s and return a dictionary of scores * `Scorer` is initialized either: * with a provided pipeline containing components to be scored * with a default pipeline containing the built-in statistical components (senter, tagger, morphologizer, parser, ner) * `Scorer.score` evaluates a list of `Example`s and returns a dictionary of scores referring to the scores provided by the components in the pipeline Significant differences: * `tags_acc` is renamed to `tag_acc` to be consistent with `token_acc` and the new `morph_acc`, `pos_acc`, and `lemma_acc` * Scoring is no longer cumulative: `Scorer.score` scores a list of examples rather than a single example and does not retain any state about previously scored examples * PRF values in the returned scores are no longer multiplied by 100 * Add kwargs to Morphologizer.evaluate * Create generalized scoring methods in Scorer * Generalized static scoring methods are added to `Scorer` * Methods require an attribute (either on Token or Doc) that is used to key the returned scores Naming differences: * `uas`, `las`, and `las_per_type` in the scores dict are renamed to `dep_uas`, `dep_las`, and `dep_las_per_type` Scoring differences: * `Doc.sents` is now scored as spans rather than on sentence-initial token positions so that `Doc.sents` and `Doc.ents` can be scored with the same method (this lowers scores since a single incorrect sentence start results in two incorrect spans) * Simplify / extend hasattr check for eval method * Add hasattr check to tokenizer scoring * Simplify to hasattr check for component scoring * Reset Example alignment if docs are set Reset the Example alignment if either doc is set in case the tokenization has changed. * Add PRF tokenization scoring for tokens as spans Add PRF scores for tokens as character spans. The scores are: * token_acc: # correct tokens / # gold tokens * token_p/r/f: PRF for (token.idx, token.idx + len(token)) * Add docstring to Scorer.score_tokenization * Rename component.evaluate() to component.score() * Update Scorer API docs * Update scoring for positive_label in textcat * Fix TextCategorizer.score kwargs * Update Language.evaluate docs * Update score names in default config
2020-07-25 13:53:02 +03:00
self.nlp = nlp
2020-07-28 22:39:42 +03:00
def score(self, examples: Iterable[Example]) -> Dict[str, Any]:
Refactor the Scorer to improve flexibility (#5731) * Refactor the Scorer to improve flexibility Refactor the `Scorer` to improve flexibility for arbitrary pipeline components. * Individual pipeline components provide their own `evaluate` methods that score a list of `Example`s and return a dictionary of scores * `Scorer` is initialized either: * with a provided pipeline containing components to be scored * with a default pipeline containing the built-in statistical components (senter, tagger, morphologizer, parser, ner) * `Scorer.score` evaluates a list of `Example`s and returns a dictionary of scores referring to the scores provided by the components in the pipeline Significant differences: * `tags_acc` is renamed to `tag_acc` to be consistent with `token_acc` and the new `morph_acc`, `pos_acc`, and `lemma_acc` * Scoring is no longer cumulative: `Scorer.score` scores a list of examples rather than a single example and does not retain any state about previously scored examples * PRF values in the returned scores are no longer multiplied by 100 * Add kwargs to Morphologizer.evaluate * Create generalized scoring methods in Scorer * Generalized static scoring methods are added to `Scorer` * Methods require an attribute (either on Token or Doc) that is used to key the returned scores Naming differences: * `uas`, `las`, and `las_per_type` in the scores dict are renamed to `dep_uas`, `dep_las`, and `dep_las_per_type` Scoring differences: * `Doc.sents` is now scored as spans rather than on sentence-initial token positions so that `Doc.sents` and `Doc.ents` can be scored with the same method (this lowers scores since a single incorrect sentence start results in two incorrect spans) * Simplify / extend hasattr check for eval method * Add hasattr check to tokenizer scoring * Simplify to hasattr check for component scoring * Reset Example alignment if docs are set Reset the Example alignment if either doc is set in case the tokenization has changed. * Add PRF tokenization scoring for tokens as spans Add PRF scores for tokens as character spans. The scores are: * token_acc: # correct tokens / # gold tokens * token_p/r/f: PRF for (token.idx, token.idx + len(token)) * Add docstring to Scorer.score_tokenization * Rename component.evaluate() to component.score() * Update Scorer API docs * Update scoring for positive_label in textcat * Fix TextCategorizer.score kwargs * Update Language.evaluate docs * Update score names in default config
2020-07-25 13:53:02 +03:00
"""Evaluate a list of Examples.
examples (Iterable[Example]): The predicted annotations + correct annotations.
RETURNS (Dict): A dictionary of scores.
2020-07-28 22:39:42 +03:00
Refactor the Scorer to improve flexibility (#5731) * Refactor the Scorer to improve flexibility Refactor the `Scorer` to improve flexibility for arbitrary pipeline components. * Individual pipeline components provide their own `evaluate` methods that score a list of `Example`s and return a dictionary of scores * `Scorer` is initialized either: * with a provided pipeline containing components to be scored * with a default pipeline containing the built-in statistical components (senter, tagger, morphologizer, parser, ner) * `Scorer.score` evaluates a list of `Example`s and returns a dictionary of scores referring to the scores provided by the components in the pipeline Significant differences: * `tags_acc` is renamed to `tag_acc` to be consistent with `token_acc` and the new `morph_acc`, `pos_acc`, and `lemma_acc` * Scoring is no longer cumulative: `Scorer.score` scores a list of examples rather than a single example and does not retain any state about previously scored examples * PRF values in the returned scores are no longer multiplied by 100 * Add kwargs to Morphologizer.evaluate * Create generalized scoring methods in Scorer * Generalized static scoring methods are added to `Scorer` * Methods require an attribute (either on Token or Doc) that is used to key the returned scores Naming differences: * `uas`, `las`, and `las_per_type` in the scores dict are renamed to `dep_uas`, `dep_las`, and `dep_las_per_type` Scoring differences: * `Doc.sents` is now scored as spans rather than on sentence-initial token positions so that `Doc.sents` and `Doc.ents` can be scored with the same method (this lowers scores since a single incorrect sentence start results in two incorrect spans) * Simplify / extend hasattr check for eval method * Add hasattr check to tokenizer scoring * Simplify to hasattr check for component scoring * Reset Example alignment if docs are set Reset the Example alignment if either doc is set in case the tokenization has changed. * Add PRF tokenization scoring for tokens as spans Add PRF scores for tokens as character spans. The scores are: * token_acc: # correct tokens / # gold tokens * token_p/r/f: PRF for (token.idx, token.idx + len(token)) * Add docstring to Scorer.score_tokenization * Rename component.evaluate() to component.score() * Update Scorer API docs * Update scoring for positive_label in textcat * Fix TextCategorizer.score kwargs * Update Language.evaluate docs * Update score names in default config
2020-07-25 13:53:02 +03:00
DOCS: https://spacy.io/api/scorer#score
"""
Refactor the Scorer to improve flexibility (#5731) * Refactor the Scorer to improve flexibility Refactor the `Scorer` to improve flexibility for arbitrary pipeline components. * Individual pipeline components provide their own `evaluate` methods that score a list of `Example`s and return a dictionary of scores * `Scorer` is initialized either: * with a provided pipeline containing components to be scored * with a default pipeline containing the built-in statistical components (senter, tagger, morphologizer, parser, ner) * `Scorer.score` evaluates a list of `Example`s and returns a dictionary of scores referring to the scores provided by the components in the pipeline Significant differences: * `tags_acc` is renamed to `tag_acc` to be consistent with `token_acc` and the new `morph_acc`, `pos_acc`, and `lemma_acc` * Scoring is no longer cumulative: `Scorer.score` scores a list of examples rather than a single example and does not retain any state about previously scored examples * PRF values in the returned scores are no longer multiplied by 100 * Add kwargs to Morphologizer.evaluate * Create generalized scoring methods in Scorer * Generalized static scoring methods are added to `Scorer` * Methods require an attribute (either on Token or Doc) that is used to key the returned scores Naming differences: * `uas`, `las`, and `las_per_type` in the scores dict are renamed to `dep_uas`, `dep_las`, and `dep_las_per_type` Scoring differences: * `Doc.sents` is now scored as spans rather than on sentence-initial token positions so that `Doc.sents` and `Doc.ents` can be scored with the same method (this lowers scores since a single incorrect sentence start results in two incorrect spans) * Simplify / extend hasattr check for eval method * Add hasattr check to tokenizer scoring * Simplify to hasattr check for component scoring * Reset Example alignment if docs are set Reset the Example alignment if either doc is set in case the tokenization has changed. * Add PRF tokenization scoring for tokens as spans Add PRF scores for tokens as character spans. The scores are: * token_acc: # correct tokens / # gold tokens * token_p/r/f: PRF for (token.idx, token.idx + len(token)) * Add docstring to Scorer.score_tokenization * Rename component.evaluate() to component.score() * Update Scorer API docs * Update scoring for positive_label in textcat * Fix TextCategorizer.score kwargs * Update Language.evaluate docs * Update score names in default config
2020-07-25 13:53:02 +03:00
scores = {}
if hasattr(self.nlp.tokenizer, "score"):
scores.update(self.nlp.tokenizer.score(examples, **self.cfg))
for name, component in self.nlp.pipeline:
if hasattr(component, "score"):
scores.update(component.score(examples, **self.cfg))
return scores
2015-03-11 04:07:03 +03:00
Refactor the Scorer to improve flexibility (#5731) * Refactor the Scorer to improve flexibility Refactor the `Scorer` to improve flexibility for arbitrary pipeline components. * Individual pipeline components provide their own `evaluate` methods that score a list of `Example`s and return a dictionary of scores * `Scorer` is initialized either: * with a provided pipeline containing components to be scored * with a default pipeline containing the built-in statistical components (senter, tagger, morphologizer, parser, ner) * `Scorer.score` evaluates a list of `Example`s and returns a dictionary of scores referring to the scores provided by the components in the pipeline Significant differences: * `tags_acc` is renamed to `tag_acc` to be consistent with `token_acc` and the new `morph_acc`, `pos_acc`, and `lemma_acc` * Scoring is no longer cumulative: `Scorer.score` scores a list of examples rather than a single example and does not retain any state about previously scored examples * PRF values in the returned scores are no longer multiplied by 100 * Add kwargs to Morphologizer.evaluate * Create generalized scoring methods in Scorer * Generalized static scoring methods are added to `Scorer` * Methods require an attribute (either on Token or Doc) that is used to key the returned scores Naming differences: * `uas`, `las`, and `las_per_type` in the scores dict are renamed to `dep_uas`, `dep_las`, and `dep_las_per_type` Scoring differences: * `Doc.sents` is now scored as spans rather than on sentence-initial token positions so that `Doc.sents` and `Doc.ents` can be scored with the same method (this lowers scores since a single incorrect sentence start results in two incorrect spans) * Simplify / extend hasattr check for eval method * Add hasattr check to tokenizer scoring * Simplify to hasattr check for component scoring * Reset Example alignment if docs are set Reset the Example alignment if either doc is set in case the tokenization has changed. * Add PRF tokenization scoring for tokens as spans Add PRF scores for tokens as character spans. The scores are: * token_acc: # correct tokens / # gold tokens * token_p/r/f: PRF for (token.idx, token.idx + len(token)) * Add docstring to Scorer.score_tokenization * Rename component.evaluate() to component.score() * Update Scorer API docs * Update scoring for positive_label in textcat * Fix TextCategorizer.score kwargs * Update Language.evaluate docs * Update score names in default config
2020-07-25 13:53:02 +03:00
@staticmethod
2020-07-28 22:39:42 +03:00
def score_tokenization(examples: Iterable[Example], **cfg) -> Dict[str, float]:
Refactor the Scorer to improve flexibility (#5731) * Refactor the Scorer to improve flexibility Refactor the `Scorer` to improve flexibility for arbitrary pipeline components. * Individual pipeline components provide their own `evaluate` methods that score a list of `Example`s and return a dictionary of scores * `Scorer` is initialized either: * with a provided pipeline containing components to be scored * with a default pipeline containing the built-in statistical components (senter, tagger, morphologizer, parser, ner) * `Scorer.score` evaluates a list of `Example`s and returns a dictionary of scores referring to the scores provided by the components in the pipeline Significant differences: * `tags_acc` is renamed to `tag_acc` to be consistent with `token_acc` and the new `morph_acc`, `pos_acc`, and `lemma_acc` * Scoring is no longer cumulative: `Scorer.score` scores a list of examples rather than a single example and does not retain any state about previously scored examples * PRF values in the returned scores are no longer multiplied by 100 * Add kwargs to Morphologizer.evaluate * Create generalized scoring methods in Scorer * Generalized static scoring methods are added to `Scorer` * Methods require an attribute (either on Token or Doc) that is used to key the returned scores Naming differences: * `uas`, `las`, and `las_per_type` in the scores dict are renamed to `dep_uas`, `dep_las`, and `dep_las_per_type` Scoring differences: * `Doc.sents` is now scored as spans rather than on sentence-initial token positions so that `Doc.sents` and `Doc.ents` can be scored with the same method (this lowers scores since a single incorrect sentence start results in two incorrect spans) * Simplify / extend hasattr check for eval method * Add hasattr check to tokenizer scoring * Simplify to hasattr check for component scoring * Reset Example alignment if docs are set Reset the Example alignment if either doc is set in case the tokenization has changed. * Add PRF tokenization scoring for tokens as spans Add PRF scores for tokens as character spans. The scores are: * token_acc: # correct tokens / # gold tokens * token_p/r/f: PRF for (token.idx, token.idx + len(token)) * Add docstring to Scorer.score_tokenization * Rename component.evaluate() to component.score() * Update Scorer API docs * Update scoring for positive_label in textcat * Fix TextCategorizer.score kwargs * Update Language.evaluate docs * Update score names in default config
2020-07-25 13:53:02 +03:00
"""Returns accuracy and PRF scores for tokenization.
* token_acc: # correct tokens / # gold tokens
* token_p/r/f: PRF for token character spans
2015-03-11 04:07:03 +03:00
Refactor the Scorer to improve flexibility (#5731) * Refactor the Scorer to improve flexibility Refactor the `Scorer` to improve flexibility for arbitrary pipeline components. * Individual pipeline components provide their own `evaluate` methods that score a list of `Example`s and return a dictionary of scores * `Scorer` is initialized either: * with a provided pipeline containing components to be scored * with a default pipeline containing the built-in statistical components (senter, tagger, morphologizer, parser, ner) * `Scorer.score` evaluates a list of `Example`s and returns a dictionary of scores referring to the scores provided by the components in the pipeline Significant differences: * `tags_acc` is renamed to `tag_acc` to be consistent with `token_acc` and the new `morph_acc`, `pos_acc`, and `lemma_acc` * Scoring is no longer cumulative: `Scorer.score` scores a list of examples rather than a single example and does not retain any state about previously scored examples * PRF values in the returned scores are no longer multiplied by 100 * Add kwargs to Morphologizer.evaluate * Create generalized scoring methods in Scorer * Generalized static scoring methods are added to `Scorer` * Methods require an attribute (either on Token or Doc) that is used to key the returned scores Naming differences: * `uas`, `las`, and `las_per_type` in the scores dict are renamed to `dep_uas`, `dep_las`, and `dep_las_per_type` Scoring differences: * `Doc.sents` is now scored as spans rather than on sentence-initial token positions so that `Doc.sents` and `Doc.ents` can be scored with the same method (this lowers scores since a single incorrect sentence start results in two incorrect spans) * Simplify / extend hasattr check for eval method * Add hasattr check to tokenizer scoring * Simplify to hasattr check for component scoring * Reset Example alignment if docs are set Reset the Example alignment if either doc is set in case the tokenization has changed. * Add PRF tokenization scoring for tokens as spans Add PRF scores for tokens as character spans. The scores are: * token_acc: # correct tokens / # gold tokens * token_p/r/f: PRF for (token.idx, token.idx + len(token)) * Add docstring to Scorer.score_tokenization * Rename component.evaluate() to component.score() * Update Scorer API docs * Update scoring for positive_label in textcat * Fix TextCategorizer.score kwargs * Update Language.evaluate docs * Update score names in default config
2020-07-25 13:53:02 +03:00
examples (Iterable[Example]): Examples to score
2020-07-28 22:39:42 +03:00
RETURNS (Dict[str, float]): A dictionary containing the scores
token_acc/p/r/f.
DOCS: https://spacy.io/api/scorer#score_tokenization
"""
Refactor the Scorer to improve flexibility (#5731) * Refactor the Scorer to improve flexibility Refactor the `Scorer` to improve flexibility for arbitrary pipeline components. * Individual pipeline components provide their own `evaluate` methods that score a list of `Example`s and return a dictionary of scores * `Scorer` is initialized either: * with a provided pipeline containing components to be scored * with a default pipeline containing the built-in statistical components (senter, tagger, morphologizer, parser, ner) * `Scorer.score` evaluates a list of `Example`s and returns a dictionary of scores referring to the scores provided by the components in the pipeline Significant differences: * `tags_acc` is renamed to `tag_acc` to be consistent with `token_acc` and the new `morph_acc`, `pos_acc`, and `lemma_acc` * Scoring is no longer cumulative: `Scorer.score` scores a list of examples rather than a single example and does not retain any state about previously scored examples * PRF values in the returned scores are no longer multiplied by 100 * Add kwargs to Morphologizer.evaluate * Create generalized scoring methods in Scorer * Generalized static scoring methods are added to `Scorer` * Methods require an attribute (either on Token or Doc) that is used to key the returned scores Naming differences: * `uas`, `las`, and `las_per_type` in the scores dict are renamed to `dep_uas`, `dep_las`, and `dep_las_per_type` Scoring differences: * `Doc.sents` is now scored as spans rather than on sentence-initial token positions so that `Doc.sents` and `Doc.ents` can be scored with the same method (this lowers scores since a single incorrect sentence start results in two incorrect spans) * Simplify / extend hasattr check for eval method * Add hasattr check to tokenizer scoring * Simplify to hasattr check for component scoring * Reset Example alignment if docs are set Reset the Example alignment if either doc is set in case the tokenization has changed. * Add PRF tokenization scoring for tokens as spans Add PRF scores for tokens as character spans. The scores are: * token_acc: # correct tokens / # gold tokens * token_p/r/f: PRF for (token.idx, token.idx + len(token)) * Add docstring to Scorer.score_tokenization * Rename component.evaluate() to component.score() * Update Scorer API docs * Update scoring for positive_label in textcat * Fix TextCategorizer.score kwargs * Update Language.evaluate docs * Update score names in default config
2020-07-25 13:53:02 +03:00
acc_score = PRFScore()
prf_score = PRFScore()
for example in examples:
gold_doc = example.reference
pred_doc = example.predicted
align = example.alignment
gold_spans = set()
pred_spans = set()
for token in gold_doc:
if token.orth_.isspace():
continue
gold_spans.add((token.idx, token.idx + len(token)))
for token in pred_doc:
if token.orth_.isspace():
continue
pred_spans.add((token.idx, token.idx + len(token)))
if align.x2y.lengths[token.i] != 1:
acc_score.fp += 1
else:
acc_score.tp += 1
prf_score.score_set(pred_spans, gold_spans)
return {
Refactor the Scorer to improve flexibility (#5731) * Refactor the Scorer to improve flexibility Refactor the `Scorer` to improve flexibility for arbitrary pipeline components. * Individual pipeline components provide their own `evaluate` methods that score a list of `Example`s and return a dictionary of scores * `Scorer` is initialized either: * with a provided pipeline containing components to be scored * with a default pipeline containing the built-in statistical components (senter, tagger, morphologizer, parser, ner) * `Scorer.score` evaluates a list of `Example`s and returns a dictionary of scores referring to the scores provided by the components in the pipeline Significant differences: * `tags_acc` is renamed to `tag_acc` to be consistent with `token_acc` and the new `morph_acc`, `pos_acc`, and `lemma_acc` * Scoring is no longer cumulative: `Scorer.score` scores a list of examples rather than a single example and does not retain any state about previously scored examples * PRF values in the returned scores are no longer multiplied by 100 * Add kwargs to Morphologizer.evaluate * Create generalized scoring methods in Scorer * Generalized static scoring methods are added to `Scorer` * Methods require an attribute (either on Token or Doc) that is used to key the returned scores Naming differences: * `uas`, `las`, and `las_per_type` in the scores dict are renamed to `dep_uas`, `dep_las`, and `dep_las_per_type` Scoring differences: * `Doc.sents` is now scored as spans rather than on sentence-initial token positions so that `Doc.sents` and `Doc.ents` can be scored with the same method (this lowers scores since a single incorrect sentence start results in two incorrect spans) * Simplify / extend hasattr check for eval method * Add hasattr check to tokenizer scoring * Simplify to hasattr check for component scoring * Reset Example alignment if docs are set Reset the Example alignment if either doc is set in case the tokenization has changed. * Add PRF tokenization scoring for tokens as spans Add PRF scores for tokens as character spans. The scores are: * token_acc: # correct tokens / # gold tokens * token_p/r/f: PRF for (token.idx, token.idx + len(token)) * Add docstring to Scorer.score_tokenization * Rename component.evaluate() to component.score() * Update Scorer API docs * Update scoring for positive_label in textcat * Fix TextCategorizer.score kwargs * Update Language.evaluate docs * Update score names in default config
2020-07-25 13:53:02 +03:00
"token_acc": acc_score.fscore,
"token_p": prf_score.precision,
"token_r": prf_score.recall,
"token_f": prf_score.fscore,
}
Refactor the Scorer to improve flexibility (#5731) * Refactor the Scorer to improve flexibility Refactor the `Scorer` to improve flexibility for arbitrary pipeline components. * Individual pipeline components provide their own `evaluate` methods that score a list of `Example`s and return a dictionary of scores * `Scorer` is initialized either: * with a provided pipeline containing components to be scored * with a default pipeline containing the built-in statistical components (senter, tagger, morphologizer, parser, ner) * `Scorer.score` evaluates a list of `Example`s and returns a dictionary of scores referring to the scores provided by the components in the pipeline Significant differences: * `tags_acc` is renamed to `tag_acc` to be consistent with `token_acc` and the new `morph_acc`, `pos_acc`, and `lemma_acc` * Scoring is no longer cumulative: `Scorer.score` scores a list of examples rather than a single example and does not retain any state about previously scored examples * PRF values in the returned scores are no longer multiplied by 100 * Add kwargs to Morphologizer.evaluate * Create generalized scoring methods in Scorer * Generalized static scoring methods are added to `Scorer` * Methods require an attribute (either on Token or Doc) that is used to key the returned scores Naming differences: * `uas`, `las`, and `las_per_type` in the scores dict are renamed to `dep_uas`, `dep_las`, and `dep_las_per_type` Scoring differences: * `Doc.sents` is now scored as spans rather than on sentence-initial token positions so that `Doc.sents` and `Doc.ents` can be scored with the same method (this lowers scores since a single incorrect sentence start results in two incorrect spans) * Simplify / extend hasattr check for eval method * Add hasattr check to tokenizer scoring * Simplify to hasattr check for component scoring * Reset Example alignment if docs are set Reset the Example alignment if either doc is set in case the tokenization has changed. * Add PRF tokenization scoring for tokens as spans Add PRF scores for tokens as character spans. The scores are: * token_acc: # correct tokens / # gold tokens * token_p/r/f: PRF for (token.idx, token.idx + len(token)) * Add docstring to Scorer.score_tokenization * Rename component.evaluate() to component.score() * Update Scorer API docs * Update scoring for positive_label in textcat * Fix TextCategorizer.score kwargs * Update Language.evaluate docs * Update score names in default config
2020-07-25 13:53:02 +03:00
@staticmethod
2020-07-28 22:39:42 +03:00
def score_token_attr(
examples: Iterable[Example],
attr: str,
*,
getter: Callable[[Token, str], Any] = getattr,
**cfg,
) -> Dict[str, float]:
Refactor the Scorer to improve flexibility (#5731) * Refactor the Scorer to improve flexibility Refactor the `Scorer` to improve flexibility for arbitrary pipeline components. * Individual pipeline components provide their own `evaluate` methods that score a list of `Example`s and return a dictionary of scores * `Scorer` is initialized either: * with a provided pipeline containing components to be scored * with a default pipeline containing the built-in statistical components (senter, tagger, morphologizer, parser, ner) * `Scorer.score` evaluates a list of `Example`s and returns a dictionary of scores referring to the scores provided by the components in the pipeline Significant differences: * `tags_acc` is renamed to `tag_acc` to be consistent with `token_acc` and the new `morph_acc`, `pos_acc`, and `lemma_acc` * Scoring is no longer cumulative: `Scorer.score` scores a list of examples rather than a single example and does not retain any state about previously scored examples * PRF values in the returned scores are no longer multiplied by 100 * Add kwargs to Morphologizer.evaluate * Create generalized scoring methods in Scorer * Generalized static scoring methods are added to `Scorer` * Methods require an attribute (either on Token or Doc) that is used to key the returned scores Naming differences: * `uas`, `las`, and `las_per_type` in the scores dict are renamed to `dep_uas`, `dep_las`, and `dep_las_per_type` Scoring differences: * `Doc.sents` is now scored as spans rather than on sentence-initial token positions so that `Doc.sents` and `Doc.ents` can be scored with the same method (this lowers scores since a single incorrect sentence start results in two incorrect spans) * Simplify / extend hasattr check for eval method * Add hasattr check to tokenizer scoring * Simplify to hasattr check for component scoring * Reset Example alignment if docs are set Reset the Example alignment if either doc is set in case the tokenization has changed. * Add PRF tokenization scoring for tokens as spans Add PRF scores for tokens as character spans. The scores are: * token_acc: # correct tokens / # gold tokens * token_p/r/f: PRF for (token.idx, token.idx + len(token)) * Add docstring to Scorer.score_tokenization * Rename component.evaluate() to component.score() * Update Scorer API docs * Update scoring for positive_label in textcat * Fix TextCategorizer.score kwargs * Update Language.evaluate docs * Update score names in default config
2020-07-25 13:53:02 +03:00
"""Returns an accuracy score for a token-level attribute.
examples (Iterable[Example]): Examples to score
attr (str): The attribute to score.
2020-07-28 22:39:42 +03:00
getter (Callable[[Token, str], Any]): Defaults to getattr. If provided,
Refactor the Scorer to improve flexibility (#5731) * Refactor the Scorer to improve flexibility Refactor the `Scorer` to improve flexibility for arbitrary pipeline components. * Individual pipeline components provide their own `evaluate` methods that score a list of `Example`s and return a dictionary of scores * `Scorer` is initialized either: * with a provided pipeline containing components to be scored * with a default pipeline containing the built-in statistical components (senter, tagger, morphologizer, parser, ner) * `Scorer.score` evaluates a list of `Example`s and returns a dictionary of scores referring to the scores provided by the components in the pipeline Significant differences: * `tags_acc` is renamed to `tag_acc` to be consistent with `token_acc` and the new `morph_acc`, `pos_acc`, and `lemma_acc` * Scoring is no longer cumulative: `Scorer.score` scores a list of examples rather than a single example and does not retain any state about previously scored examples * PRF values in the returned scores are no longer multiplied by 100 * Add kwargs to Morphologizer.evaluate * Create generalized scoring methods in Scorer * Generalized static scoring methods are added to `Scorer` * Methods require an attribute (either on Token or Doc) that is used to key the returned scores Naming differences: * `uas`, `las`, and `las_per_type` in the scores dict are renamed to `dep_uas`, `dep_las`, and `dep_las_per_type` Scoring differences: * `Doc.sents` is now scored as spans rather than on sentence-initial token positions so that `Doc.sents` and `Doc.ents` can be scored with the same method (this lowers scores since a single incorrect sentence start results in two incorrect spans) * Simplify / extend hasattr check for eval method * Add hasattr check to tokenizer scoring * Simplify to hasattr check for component scoring * Reset Example alignment if docs are set Reset the Example alignment if either doc is set in case the tokenization has changed. * Add PRF tokenization scoring for tokens as spans Add PRF scores for tokens as character spans. The scores are: * token_acc: # correct tokens / # gold tokens * token_p/r/f: PRF for (token.idx, token.idx + len(token)) * Add docstring to Scorer.score_tokenization * Rename component.evaluate() to component.score() * Update Scorer API docs * Update scoring for positive_label in textcat * Fix TextCategorizer.score kwargs * Update Language.evaluate docs * Update score names in default config
2020-07-25 13:53:02 +03:00
getter(token, attr) should return the value of the attribute for an
individual token.
2020-07-28 22:39:42 +03:00
RETURNS (Dict[str, float]): A dictionary containing the accuracy score
under the key attr_acc.
DOCS: https://spacy.io/api/scorer#score_token_attr
2019-07-10 12:19:28 +03:00
"""
Refactor the Scorer to improve flexibility (#5731) * Refactor the Scorer to improve flexibility Refactor the `Scorer` to improve flexibility for arbitrary pipeline components. * Individual pipeline components provide their own `evaluate` methods that score a list of `Example`s and return a dictionary of scores * `Scorer` is initialized either: * with a provided pipeline containing components to be scored * with a default pipeline containing the built-in statistical components (senter, tagger, morphologizer, parser, ner) * `Scorer.score` evaluates a list of `Example`s and returns a dictionary of scores referring to the scores provided by the components in the pipeline Significant differences: * `tags_acc` is renamed to `tag_acc` to be consistent with `token_acc` and the new `morph_acc`, `pos_acc`, and `lemma_acc` * Scoring is no longer cumulative: `Scorer.score` scores a list of examples rather than a single example and does not retain any state about previously scored examples * PRF values in the returned scores are no longer multiplied by 100 * Add kwargs to Morphologizer.evaluate * Create generalized scoring methods in Scorer * Generalized static scoring methods are added to `Scorer` * Methods require an attribute (either on Token or Doc) that is used to key the returned scores Naming differences: * `uas`, `las`, and `las_per_type` in the scores dict are renamed to `dep_uas`, `dep_las`, and `dep_las_per_type` Scoring differences: * `Doc.sents` is now scored as spans rather than on sentence-initial token positions so that `Doc.sents` and `Doc.ents` can be scored with the same method (this lowers scores since a single incorrect sentence start results in two incorrect spans) * Simplify / extend hasattr check for eval method * Add hasattr check to tokenizer scoring * Simplify to hasattr check for component scoring * Reset Example alignment if docs are set Reset the Example alignment if either doc is set in case the tokenization has changed. * Add PRF tokenization scoring for tokens as spans Add PRF scores for tokens as character spans. The scores are: * token_acc: # correct tokens / # gold tokens * token_p/r/f: PRF for (token.idx, token.idx + len(token)) * Add docstring to Scorer.score_tokenization * Rename component.evaluate() to component.score() * Update Scorer API docs * Update scoring for positive_label in textcat * Fix TextCategorizer.score kwargs * Update Language.evaluate docs * Update score names in default config
2020-07-25 13:53:02 +03:00
tag_score = PRFScore()
for example in examples:
gold_doc = example.reference
pred_doc = example.predicted
align = example.alignment
gold_tags = set()
for gold_i, token in enumerate(gold_doc):
gold_tags.add((gold_i, getter(token, attr)))
pred_tags = set()
for token in pred_doc:
if token.orth_.isspace():
continue
if align.x2y.lengths[token.i] == 1:
gold_i = align.x2y[token.i].dataXd[0, 0]
pred_tags.add((gold_i, getter(token, attr)))
tag_score.score_set(pred_tags, gold_tags)
2020-07-28 22:39:42 +03:00
return {f"{attr}_acc": tag_score.fscore}
2019-07-10 12:19:28 +03:00
Refactor the Scorer to improve flexibility (#5731) * Refactor the Scorer to improve flexibility Refactor the `Scorer` to improve flexibility for arbitrary pipeline components. * Individual pipeline components provide their own `evaluate` methods that score a list of `Example`s and return a dictionary of scores * `Scorer` is initialized either: * with a provided pipeline containing components to be scored * with a default pipeline containing the built-in statistical components (senter, tagger, morphologizer, parser, ner) * `Scorer.score` evaluates a list of `Example`s and returns a dictionary of scores referring to the scores provided by the components in the pipeline Significant differences: * `tags_acc` is renamed to `tag_acc` to be consistent with `token_acc` and the new `morph_acc`, `pos_acc`, and `lemma_acc` * Scoring is no longer cumulative: `Scorer.score` scores a list of examples rather than a single example and does not retain any state about previously scored examples * PRF values in the returned scores are no longer multiplied by 100 * Add kwargs to Morphologizer.evaluate * Create generalized scoring methods in Scorer * Generalized static scoring methods are added to `Scorer` * Methods require an attribute (either on Token or Doc) that is used to key the returned scores Naming differences: * `uas`, `las`, and `las_per_type` in the scores dict are renamed to `dep_uas`, `dep_las`, and `dep_las_per_type` Scoring differences: * `Doc.sents` is now scored as spans rather than on sentence-initial token positions so that `Doc.sents` and `Doc.ents` can be scored with the same method (this lowers scores since a single incorrect sentence start results in two incorrect spans) * Simplify / extend hasattr check for eval method * Add hasattr check to tokenizer scoring * Simplify to hasattr check for component scoring * Reset Example alignment if docs are set Reset the Example alignment if either doc is set in case the tokenization has changed. * Add PRF tokenization scoring for tokens as spans Add PRF scores for tokens as character spans. The scores are: * token_acc: # correct tokens / # gold tokens * token_p/r/f: PRF for (token.idx, token.idx + len(token)) * Add docstring to Scorer.score_tokenization * Rename component.evaluate() to component.score() * Update Scorer API docs * Update scoring for positive_label in textcat * Fix TextCategorizer.score kwargs * Update Language.evaluate docs * Update score names in default config
2020-07-25 13:53:02 +03:00
@staticmethod
2020-07-28 22:39:42 +03:00
def score_token_attr_per_feat(
examples: Iterable[Example],
attr: str,
*,
getter: Callable[[Token, str], Any] = getattr,
**cfg,
):
Refactor the Scorer to improve flexibility (#5731) * Refactor the Scorer to improve flexibility Refactor the `Scorer` to improve flexibility for arbitrary pipeline components. * Individual pipeline components provide their own `evaluate` methods that score a list of `Example`s and return a dictionary of scores * `Scorer` is initialized either: * with a provided pipeline containing components to be scored * with a default pipeline containing the built-in statistical components (senter, tagger, morphologizer, parser, ner) * `Scorer.score` evaluates a list of `Example`s and returns a dictionary of scores referring to the scores provided by the components in the pipeline Significant differences: * `tags_acc` is renamed to `tag_acc` to be consistent with `token_acc` and the new `morph_acc`, `pos_acc`, and `lemma_acc` * Scoring is no longer cumulative: `Scorer.score` scores a list of examples rather than a single example and does not retain any state about previously scored examples * PRF values in the returned scores are no longer multiplied by 100 * Add kwargs to Morphologizer.evaluate * Create generalized scoring methods in Scorer * Generalized static scoring methods are added to `Scorer` * Methods require an attribute (either on Token or Doc) that is used to key the returned scores Naming differences: * `uas`, `las`, and `las_per_type` in the scores dict are renamed to `dep_uas`, `dep_las`, and `dep_las_per_type` Scoring differences: * `Doc.sents` is now scored as spans rather than on sentence-initial token positions so that `Doc.sents` and `Doc.ents` can be scored with the same method (this lowers scores since a single incorrect sentence start results in two incorrect spans) * Simplify / extend hasattr check for eval method * Add hasattr check to tokenizer scoring * Simplify to hasattr check for component scoring * Reset Example alignment if docs are set Reset the Example alignment if either doc is set in case the tokenization has changed. * Add PRF tokenization scoring for tokens as spans Add PRF scores for tokens as character spans. The scores are: * token_acc: # correct tokens / # gold tokens * token_p/r/f: PRF for (token.idx, token.idx + len(token)) * Add docstring to Scorer.score_tokenization * Rename component.evaluate() to component.score() * Update Scorer API docs * Update scoring for positive_label in textcat * Fix TextCategorizer.score kwargs * Update Language.evaluate docs * Update score names in default config
2020-07-25 13:53:02 +03:00
"""Return PRF scores per feat for a token attribute in UFEATS format.
examples (Iterable[Example]): Examples to score
attr (str): The attribute to score.
2020-07-28 22:39:42 +03:00
getter (Callable[[Token, str], Any]): Defaults to getattr. If provided,
Refactor the Scorer to improve flexibility (#5731) * Refactor the Scorer to improve flexibility Refactor the `Scorer` to improve flexibility for arbitrary pipeline components. * Individual pipeline components provide their own `evaluate` methods that score a list of `Example`s and return a dictionary of scores * `Scorer` is initialized either: * with a provided pipeline containing components to be scored * with a default pipeline containing the built-in statistical components (senter, tagger, morphologizer, parser, ner) * `Scorer.score` evaluates a list of `Example`s and returns a dictionary of scores referring to the scores provided by the components in the pipeline Significant differences: * `tags_acc` is renamed to `tag_acc` to be consistent with `token_acc` and the new `morph_acc`, `pos_acc`, and `lemma_acc` * Scoring is no longer cumulative: `Scorer.score` scores a list of examples rather than a single example and does not retain any state about previously scored examples * PRF values in the returned scores are no longer multiplied by 100 * Add kwargs to Morphologizer.evaluate * Create generalized scoring methods in Scorer * Generalized static scoring methods are added to `Scorer` * Methods require an attribute (either on Token or Doc) that is used to key the returned scores Naming differences: * `uas`, `las`, and `las_per_type` in the scores dict are renamed to `dep_uas`, `dep_las`, and `dep_las_per_type` Scoring differences: * `Doc.sents` is now scored as spans rather than on sentence-initial token positions so that `Doc.sents` and `Doc.ents` can be scored with the same method (this lowers scores since a single incorrect sentence start results in two incorrect spans) * Simplify / extend hasattr check for eval method * Add hasattr check to tokenizer scoring * Simplify to hasattr check for component scoring * Reset Example alignment if docs are set Reset the Example alignment if either doc is set in case the tokenization has changed. * Add PRF tokenization scoring for tokens as spans Add PRF scores for tokens as character spans. The scores are: * token_acc: # correct tokens / # gold tokens * token_p/r/f: PRF for (token.idx, token.idx + len(token)) * Add docstring to Scorer.score_tokenization * Rename component.evaluate() to component.score() * Update Scorer API docs * Update scoring for positive_label in textcat * Fix TextCategorizer.score kwargs * Update Language.evaluate docs * Update score names in default config
2020-07-25 13:53:02 +03:00
getter(token, attr) should return the value of the attribute for an
individual token.
RETURNS (dict): A dictionary containing the per-feat PRF scores unders
the key attr_per_feat.
Add textcat to train CLI (#4226) * Add doc.cats to spacy.gold at the paragraph level Support `doc.cats` as `"cats": [{"label": string, "value": number}]` in the spacy JSON training format at the paragraph level. * `spacy.gold.docs_to_json()` writes `docs.cats` * `GoldCorpus` reads in cats in each `GoldParse` * Update instances of gold_tuples to handle cats Update iteration over gold_tuples / gold_parses to handle addition of cats at the paragraph level. * Add textcat to train CLI * Add textcat options to train CLI * Add textcat labels in `TextCategorizer.begin_training()` * Add textcat evaluation to `Scorer`: * For binary exclusive classes with provided label: F1 for label * For 2+ exclusive classes: F1 macro average * For multilabel (not exclusive): ROC AUC macro average (currently relying on sklearn) * Provide user info on textcat evaluation settings, potential incompatibilities * Provide pipeline to Scorer in `Language.evaluate` for textcat config * Customize train CLI output to include only metrics relevant to current pipeline * Add textcat evaluation to evaluate CLI * Fix handling of unset arguments and config params Fix handling of unset arguments and model confiug parameters in Scorer initialization. * Temporarily add sklearn requirement * Remove sklearn version number * Improve Scorer handling of models without textcats * Fixing Scorer handling of models without textcats * Update Scorer output for python 2.7 * Modify inf in Scorer for python 2.7 * Auto-format Also make small adjustments to make auto-formatting with black easier and produce nicer results * Move error message to Errors * Update documentation * Add cats to annotation JSON format [ci skip] * Fix tpl flag and docs [ci skip] * Switch to internal roc_auc_score Switch to internal `roc_auc_score()` adapted from scikit-learn. * Add AUCROCScore tests and improve errors/warnings * Add tests for AUCROCScore and roc_auc_score * Add missing error for only positive/negative values * Remove unnecessary warnings and errors * Make reduced roc_auc_score functions private Because most of the checks and warnings have been stripped for the internal functions and access is only intended through `ROCAUCScore`, make the functions for roc_auc_score adapted from scikit-learn private. * Check that data corresponds with multilabel flag Check that the training instances correspond with the multilabel flag, adding the multilabel flag if required. * Add textcat score to early stopping check * Add more checks to debug-data for textcat * Add example training data for textcat * Add more checks to textcat train CLI * Check configuration when extending base model * Fix typos * Update textcat example data * Provide licensing details and licenses for data * Remove two labels with no positive instances from jigsaw-toxic-comment data. Co-authored-by: Ines Montani <ines@ines.io>
2019-09-15 23:31:31 +03:00
"""
Refactor the Scorer to improve flexibility (#5731) * Refactor the Scorer to improve flexibility Refactor the `Scorer` to improve flexibility for arbitrary pipeline components. * Individual pipeline components provide their own `evaluate` methods that score a list of `Example`s and return a dictionary of scores * `Scorer` is initialized either: * with a provided pipeline containing components to be scored * with a default pipeline containing the built-in statistical components (senter, tagger, morphologizer, parser, ner) * `Scorer.score` evaluates a list of `Example`s and returns a dictionary of scores referring to the scores provided by the components in the pipeline Significant differences: * `tags_acc` is renamed to `tag_acc` to be consistent with `token_acc` and the new `morph_acc`, `pos_acc`, and `lemma_acc` * Scoring is no longer cumulative: `Scorer.score` scores a list of examples rather than a single example and does not retain any state about previously scored examples * PRF values in the returned scores are no longer multiplied by 100 * Add kwargs to Morphologizer.evaluate * Create generalized scoring methods in Scorer * Generalized static scoring methods are added to `Scorer` * Methods require an attribute (either on Token or Doc) that is used to key the returned scores Naming differences: * `uas`, `las`, and `las_per_type` in the scores dict are renamed to `dep_uas`, `dep_las`, and `dep_las_per_type` Scoring differences: * `Doc.sents` is now scored as spans rather than on sentence-initial token positions so that `Doc.sents` and `Doc.ents` can be scored with the same method (this lowers scores since a single incorrect sentence start results in two incorrect spans) * Simplify / extend hasattr check for eval method * Add hasattr check to tokenizer scoring * Simplify to hasattr check for component scoring * Reset Example alignment if docs are set Reset the Example alignment if either doc is set in case the tokenization has changed. * Add PRF tokenization scoring for tokens as spans Add PRF scores for tokens as character spans. The scores are: * token_acc: # correct tokens / # gold tokens * token_p/r/f: PRF for (token.idx, token.idx + len(token)) * Add docstring to Scorer.score_tokenization * Rename component.evaluate() to component.score() * Update Scorer API docs * Update scoring for positive_label in textcat * Fix TextCategorizer.score kwargs * Update Language.evaluate docs * Update score names in default config
2020-07-25 13:53:02 +03:00
per_feat = {}
for example in examples:
pred_doc = example.predicted
gold_doc = example.reference
align = example.alignment
gold_per_feat = {}
for gold_i, token in enumerate(gold_doc):
morph = str(getter(token, attr))
if morph:
for feat in morph.split(Morphology.FEATURE_SEP):
field, values = feat.split(Morphology.FIELD_SEP)
if field not in per_feat:
per_feat[field] = PRFScore()
if field not in gold_per_feat:
gold_per_feat[field] = set()
gold_per_feat[field].add((gold_i, feat))
pred_per_feat = {}
for token in pred_doc:
if token.orth_.isspace():
continue
if align.x2y.lengths[token.i] == 1:
gold_i = align.x2y[token.i].dataXd[0, 0]
morph = str(getter(token, attr))
if morph:
for feat in morph.split("|"):
field, values = feat.split("=")
if field not in per_feat:
per_feat[field] = PRFScore()
if field not in pred_per_feat:
pred_per_feat[field] = set()
pred_per_feat[field].add((gold_i, feat))
for field in per_feat:
per_feat[field].score_set(
pred_per_feat.get(field, set()), gold_per_feat.get(field, set()),
)
result = {k: v.to_dict() for k, v in per_feat.items()}
return {f"{attr}_per_feat": result}
Refactor the Scorer to improve flexibility (#5731) * Refactor the Scorer to improve flexibility Refactor the `Scorer` to improve flexibility for arbitrary pipeline components. * Individual pipeline components provide their own `evaluate` methods that score a list of `Example`s and return a dictionary of scores * `Scorer` is initialized either: * with a provided pipeline containing components to be scored * with a default pipeline containing the built-in statistical components (senter, tagger, morphologizer, parser, ner) * `Scorer.score` evaluates a list of `Example`s and returns a dictionary of scores referring to the scores provided by the components in the pipeline Significant differences: * `tags_acc` is renamed to `tag_acc` to be consistent with `token_acc` and the new `morph_acc`, `pos_acc`, and `lemma_acc` * Scoring is no longer cumulative: `Scorer.score` scores a list of examples rather than a single example and does not retain any state about previously scored examples * PRF values in the returned scores are no longer multiplied by 100 * Add kwargs to Morphologizer.evaluate * Create generalized scoring methods in Scorer * Generalized static scoring methods are added to `Scorer` * Methods require an attribute (either on Token or Doc) that is used to key the returned scores Naming differences: * `uas`, `las`, and `las_per_type` in the scores dict are renamed to `dep_uas`, `dep_las`, and `dep_las_per_type` Scoring differences: * `Doc.sents` is now scored as spans rather than on sentence-initial token positions so that `Doc.sents` and `Doc.ents` can be scored with the same method (this lowers scores since a single incorrect sentence start results in two incorrect spans) * Simplify / extend hasattr check for eval method * Add hasattr check to tokenizer scoring * Simplify to hasattr check for component scoring * Reset Example alignment if docs are set Reset the Example alignment if either doc is set in case the tokenization has changed. * Add PRF tokenization scoring for tokens as spans Add PRF scores for tokens as character spans. The scores are: * token_acc: # correct tokens / # gold tokens * token_p/r/f: PRF for (token.idx, token.idx + len(token)) * Add docstring to Scorer.score_tokenization * Rename component.evaluate() to component.score() * Update Scorer API docs * Update scoring for positive_label in textcat * Fix TextCategorizer.score kwargs * Update Language.evaluate docs * Update score names in default config
2020-07-25 13:53:02 +03:00
@staticmethod
2020-07-28 22:39:42 +03:00
def score_spans(
examples: Iterable[Example],
attr: str,
*,
2020-08-17 17:45:24 +03:00
getter: Callable[[Doc, str], Iterable[Span]] = getattr,
2020-07-28 22:39:42 +03:00
**cfg,
) -> Dict[str, Any]:
Refactor the Scorer to improve flexibility (#5731) * Refactor the Scorer to improve flexibility Refactor the `Scorer` to improve flexibility for arbitrary pipeline components. * Individual pipeline components provide their own `evaluate` methods that score a list of `Example`s and return a dictionary of scores * `Scorer` is initialized either: * with a provided pipeline containing components to be scored * with a default pipeline containing the built-in statistical components (senter, tagger, morphologizer, parser, ner) * `Scorer.score` evaluates a list of `Example`s and returns a dictionary of scores referring to the scores provided by the components in the pipeline Significant differences: * `tags_acc` is renamed to `tag_acc` to be consistent with `token_acc` and the new `morph_acc`, `pos_acc`, and `lemma_acc` * Scoring is no longer cumulative: `Scorer.score` scores a list of examples rather than a single example and does not retain any state about previously scored examples * PRF values in the returned scores are no longer multiplied by 100 * Add kwargs to Morphologizer.evaluate * Create generalized scoring methods in Scorer * Generalized static scoring methods are added to `Scorer` * Methods require an attribute (either on Token or Doc) that is used to key the returned scores Naming differences: * `uas`, `las`, and `las_per_type` in the scores dict are renamed to `dep_uas`, `dep_las`, and `dep_las_per_type` Scoring differences: * `Doc.sents` is now scored as spans rather than on sentence-initial token positions so that `Doc.sents` and `Doc.ents` can be scored with the same method (this lowers scores since a single incorrect sentence start results in two incorrect spans) * Simplify / extend hasattr check for eval method * Add hasattr check to tokenizer scoring * Simplify to hasattr check for component scoring * Reset Example alignment if docs are set Reset the Example alignment if either doc is set in case the tokenization has changed. * Add PRF tokenization scoring for tokens as spans Add PRF scores for tokens as character spans. The scores are: * token_acc: # correct tokens / # gold tokens * token_p/r/f: PRF for (token.idx, token.idx + len(token)) * Add docstring to Scorer.score_tokenization * Rename component.evaluate() to component.score() * Update Scorer API docs * Update scoring for positive_label in textcat * Fix TextCategorizer.score kwargs * Update Language.evaluate docs * Update score names in default config
2020-07-25 13:53:02 +03:00
"""Returns PRF scores for labeled spans.
Add textcat to train CLI (#4226) * Add doc.cats to spacy.gold at the paragraph level Support `doc.cats` as `"cats": [{"label": string, "value": number}]` in the spacy JSON training format at the paragraph level. * `spacy.gold.docs_to_json()` writes `docs.cats` * `GoldCorpus` reads in cats in each `GoldParse` * Update instances of gold_tuples to handle cats Update iteration over gold_tuples / gold_parses to handle addition of cats at the paragraph level. * Add textcat to train CLI * Add textcat options to train CLI * Add textcat labels in `TextCategorizer.begin_training()` * Add textcat evaluation to `Scorer`: * For binary exclusive classes with provided label: F1 for label * For 2+ exclusive classes: F1 macro average * For multilabel (not exclusive): ROC AUC macro average (currently relying on sklearn) * Provide user info on textcat evaluation settings, potential incompatibilities * Provide pipeline to Scorer in `Language.evaluate` for textcat config * Customize train CLI output to include only metrics relevant to current pipeline * Add textcat evaluation to evaluate CLI * Fix handling of unset arguments and config params Fix handling of unset arguments and model confiug parameters in Scorer initialization. * Temporarily add sklearn requirement * Remove sklearn version number * Improve Scorer handling of models without textcats * Fixing Scorer handling of models without textcats * Update Scorer output for python 2.7 * Modify inf in Scorer for python 2.7 * Auto-format Also make small adjustments to make auto-formatting with black easier and produce nicer results * Move error message to Errors * Update documentation * Add cats to annotation JSON format [ci skip] * Fix tpl flag and docs [ci skip] * Switch to internal roc_auc_score Switch to internal `roc_auc_score()` adapted from scikit-learn. * Add AUCROCScore tests and improve errors/warnings * Add tests for AUCROCScore and roc_auc_score * Add missing error for only positive/negative values * Remove unnecessary warnings and errors * Make reduced roc_auc_score functions private Because most of the checks and warnings have been stripped for the internal functions and access is only intended through `ROCAUCScore`, make the functions for roc_auc_score adapted from scikit-learn private. * Check that data corresponds with multilabel flag Check that the training instances correspond with the multilabel flag, adding the multilabel flag if required. * Add textcat score to early stopping check * Add more checks to debug-data for textcat * Add example training data for textcat * Add more checks to textcat train CLI * Check configuration when extending base model * Fix typos * Update textcat example data * Provide licensing details and licenses for data * Remove two labels with no positive instances from jigsaw-toxic-comment data. Co-authored-by: Ines Montani <ines@ines.io>
2019-09-15 23:31:31 +03:00
Refactor the Scorer to improve flexibility (#5731) * Refactor the Scorer to improve flexibility Refactor the `Scorer` to improve flexibility for arbitrary pipeline components. * Individual pipeline components provide their own `evaluate` methods that score a list of `Example`s and return a dictionary of scores * `Scorer` is initialized either: * with a provided pipeline containing components to be scored * with a default pipeline containing the built-in statistical components (senter, tagger, morphologizer, parser, ner) * `Scorer.score` evaluates a list of `Example`s and returns a dictionary of scores referring to the scores provided by the components in the pipeline Significant differences: * `tags_acc` is renamed to `tag_acc` to be consistent with `token_acc` and the new `morph_acc`, `pos_acc`, and `lemma_acc` * Scoring is no longer cumulative: `Scorer.score` scores a list of examples rather than a single example and does not retain any state about previously scored examples * PRF values in the returned scores are no longer multiplied by 100 * Add kwargs to Morphologizer.evaluate * Create generalized scoring methods in Scorer * Generalized static scoring methods are added to `Scorer` * Methods require an attribute (either on Token or Doc) that is used to key the returned scores Naming differences: * `uas`, `las`, and `las_per_type` in the scores dict are renamed to `dep_uas`, `dep_las`, and `dep_las_per_type` Scoring differences: * `Doc.sents` is now scored as spans rather than on sentence-initial token positions so that `Doc.sents` and `Doc.ents` can be scored with the same method (this lowers scores since a single incorrect sentence start results in two incorrect spans) * Simplify / extend hasattr check for eval method * Add hasattr check to tokenizer scoring * Simplify to hasattr check for component scoring * Reset Example alignment if docs are set Reset the Example alignment if either doc is set in case the tokenization has changed. * Add PRF tokenization scoring for tokens as spans Add PRF scores for tokens as character spans. The scores are: * token_acc: # correct tokens / # gold tokens * token_p/r/f: PRF for (token.idx, token.idx + len(token)) * Add docstring to Scorer.score_tokenization * Rename component.evaluate() to component.score() * Update Scorer API docs * Update scoring for positive_label in textcat * Fix TextCategorizer.score kwargs * Update Language.evaluate docs * Update score names in default config
2020-07-25 13:53:02 +03:00
examples (Iterable[Example]): Examples to score
attr (str): The attribute to score.
2020-08-17 17:45:24 +03:00
getter (Callable[[Doc, str], Iterable[Span]]): Defaults to getattr. If
provided, getter(doc, attr) should return the spans for the
individual doc.
2020-07-28 22:39:42 +03:00
RETURNS (Dict[str, Any]): A dictionary containing the PRF scores under
the keys attr_p/r/f and the per-type PRF scores under attr_per_type.
DOCS: https://spacy.io/api/scorer#score_spans
2019-05-24 15:06:04 +03:00
"""
Refactor the Scorer to improve flexibility (#5731) * Refactor the Scorer to improve flexibility Refactor the `Scorer` to improve flexibility for arbitrary pipeline components. * Individual pipeline components provide their own `evaluate` methods that score a list of `Example`s and return a dictionary of scores * `Scorer` is initialized either: * with a provided pipeline containing components to be scored * with a default pipeline containing the built-in statistical components (senter, tagger, morphologizer, parser, ner) * `Scorer.score` evaluates a list of `Example`s and returns a dictionary of scores referring to the scores provided by the components in the pipeline Significant differences: * `tags_acc` is renamed to `tag_acc` to be consistent with `token_acc` and the new `morph_acc`, `pos_acc`, and `lemma_acc` * Scoring is no longer cumulative: `Scorer.score` scores a list of examples rather than a single example and does not retain any state about previously scored examples * PRF values in the returned scores are no longer multiplied by 100 * Add kwargs to Morphologizer.evaluate * Create generalized scoring methods in Scorer * Generalized static scoring methods are added to `Scorer` * Methods require an attribute (either on Token or Doc) that is used to key the returned scores Naming differences: * `uas`, `las`, and `las_per_type` in the scores dict are renamed to `dep_uas`, `dep_las`, and `dep_las_per_type` Scoring differences: * `Doc.sents` is now scored as spans rather than on sentence-initial token positions so that `Doc.sents` and `Doc.ents` can be scored with the same method (this lowers scores since a single incorrect sentence start results in two incorrect spans) * Simplify / extend hasattr check for eval method * Add hasattr check to tokenizer scoring * Simplify to hasattr check for component scoring * Reset Example alignment if docs are set Reset the Example alignment if either doc is set in case the tokenization has changed. * Add PRF tokenization scoring for tokens as spans Add PRF scores for tokens as character spans. The scores are: * token_acc: # correct tokens / # gold tokens * token_p/r/f: PRF for (token.idx, token.idx + len(token)) * Add docstring to Scorer.score_tokenization * Rename component.evaluate() to component.score() * Update Scorer API docs * Update scoring for positive_label in textcat * Fix TextCategorizer.score kwargs * Update Language.evaluate docs * Update score names in default config
2020-07-25 13:53:02 +03:00
score = PRFScore()
score_per_type = dict()
for example in examples:
pred_doc = example.predicted
gold_doc = example.reference
# Find all labels in gold and doc
labels = set(
[k.label_ for k in getter(gold_doc, attr)]
+ [k.label_ for k in getter(pred_doc, attr)]
)
# Set up all labels for per type scoring and prepare gold per type
gold_per_type = {label: set() for label in labels}
for label in labels:
if label not in score_per_type:
score_per_type[label] = PRFScore()
# Find all predidate labels, for all and per type
gold_spans = set()
pred_spans = set()
# Special case for ents:
# If we have missing values in the gold, we can't easily tell
# whether our NER predictions are true.
# It seems bad but it's what we've always done.
if attr == "ents" and not all(token.ent_iob != 0 for token in gold_doc):
continue
for span in getter(gold_doc, attr):
gold_span = (span.label_, span.start, span.end - 1)
gold_spans.add(gold_span)
gold_per_type[span.label_].add((span.label_, span.start, span.end - 1))
pred_per_type = {label: set() for label in labels}
for span in example.get_aligned_spans_x2y(getter(pred_doc, attr)):
pred_spans.add((span.label_, span.start, span.end - 1))
pred_per_type[span.label_].add((span.label_, span.start, span.end - 1))
# Scores per label
for k, v in score_per_type.items():
if k in pred_per_type:
v.score_set(pred_per_type[k], gold_per_type[k])
# Score for all labels
score.score_set(pred_spans, gold_spans)
results = {
2020-07-28 22:39:42 +03:00
f"{attr}_p": score.precision,
f"{attr}_r": score.recall,
f"{attr}_f": score.fscore,
f"{attr}_per_type": {k: v.to_dict() for k, v in score_per_type.items()},
Refactor the Scorer to improve flexibility (#5731) * Refactor the Scorer to improve flexibility Refactor the `Scorer` to improve flexibility for arbitrary pipeline components. * Individual pipeline components provide their own `evaluate` methods that score a list of `Example`s and return a dictionary of scores * `Scorer` is initialized either: * with a provided pipeline containing components to be scored * with a default pipeline containing the built-in statistical components (senter, tagger, morphologizer, parser, ner) * `Scorer.score` evaluates a list of `Example`s and returns a dictionary of scores referring to the scores provided by the components in the pipeline Significant differences: * `tags_acc` is renamed to `tag_acc` to be consistent with `token_acc` and the new `morph_acc`, `pos_acc`, and `lemma_acc` * Scoring is no longer cumulative: `Scorer.score` scores a list of examples rather than a single example and does not retain any state about previously scored examples * PRF values in the returned scores are no longer multiplied by 100 * Add kwargs to Morphologizer.evaluate * Create generalized scoring methods in Scorer * Generalized static scoring methods are added to `Scorer` * Methods require an attribute (either on Token or Doc) that is used to key the returned scores Naming differences: * `uas`, `las`, and `las_per_type` in the scores dict are renamed to `dep_uas`, `dep_las`, and `dep_las_per_type` Scoring differences: * `Doc.sents` is now scored as spans rather than on sentence-initial token positions so that `Doc.sents` and `Doc.ents` can be scored with the same method (this lowers scores since a single incorrect sentence start results in two incorrect spans) * Simplify / extend hasattr check for eval method * Add hasattr check to tokenizer scoring * Simplify to hasattr check for component scoring * Reset Example alignment if docs are set Reset the Example alignment if either doc is set in case the tokenization has changed. * Add PRF tokenization scoring for tokens as spans Add PRF scores for tokens as character spans. The scores are: * token_acc: # correct tokens / # gold tokens * token_p/r/f: PRF for (token.idx, token.idx + len(token)) * Add docstring to Scorer.score_tokenization * Rename component.evaluate() to component.score() * Update Scorer API docs * Update scoring for positive_label in textcat * Fix TextCategorizer.score kwargs * Update Language.evaluate docs * Update score names in default config
2020-07-25 13:53:02 +03:00
}
return results
@staticmethod
def score_cats(
2020-07-28 22:39:42 +03:00
examples: Iterable[Example],
attr: str,
*,
2020-07-29 11:39:33 +03:00
getter: Callable[[Doc, str], Any] = getattr,
2020-07-28 22:39:42 +03:00
labels: Iterable[str] = tuple(),
multi_label: bool = True,
positive_label: Optional[str] = None,
threshold: Optional[float] = None,
2020-07-28 22:39:42 +03:00
**cfg,
) -> Dict[str, Any]:
Refactor the Scorer to improve flexibility (#5731) * Refactor the Scorer to improve flexibility Refactor the `Scorer` to improve flexibility for arbitrary pipeline components. * Individual pipeline components provide their own `evaluate` methods that score a list of `Example`s and return a dictionary of scores * `Scorer` is initialized either: * with a provided pipeline containing components to be scored * with a default pipeline containing the built-in statistical components (senter, tagger, morphologizer, parser, ner) * `Scorer.score` evaluates a list of `Example`s and returns a dictionary of scores referring to the scores provided by the components in the pipeline Significant differences: * `tags_acc` is renamed to `tag_acc` to be consistent with `token_acc` and the new `morph_acc`, `pos_acc`, and `lemma_acc` * Scoring is no longer cumulative: `Scorer.score` scores a list of examples rather than a single example and does not retain any state about previously scored examples * PRF values in the returned scores are no longer multiplied by 100 * Add kwargs to Morphologizer.evaluate * Create generalized scoring methods in Scorer * Generalized static scoring methods are added to `Scorer` * Methods require an attribute (either on Token or Doc) that is used to key the returned scores Naming differences: * `uas`, `las`, and `las_per_type` in the scores dict are renamed to `dep_uas`, `dep_las`, and `dep_las_per_type` Scoring differences: * `Doc.sents` is now scored as spans rather than on sentence-initial token positions so that `Doc.sents` and `Doc.ents` can be scored with the same method (this lowers scores since a single incorrect sentence start results in two incorrect spans) * Simplify / extend hasattr check for eval method * Add hasattr check to tokenizer scoring * Simplify to hasattr check for component scoring * Reset Example alignment if docs are set Reset the Example alignment if either doc is set in case the tokenization has changed. * Add PRF tokenization scoring for tokens as spans Add PRF scores for tokens as character spans. The scores are: * token_acc: # correct tokens / # gold tokens * token_p/r/f: PRF for (token.idx, token.idx + len(token)) * Add docstring to Scorer.score_tokenization * Rename component.evaluate() to component.score() * Update Scorer API docs * Update scoring for positive_label in textcat * Fix TextCategorizer.score kwargs * Update Language.evaluate docs * Update score names in default config
2020-07-25 13:53:02 +03:00
"""Returns PRF and ROC AUC scores for a doc-level attribute with a
dict with scores for each label like Doc.cats. The reported overall
score depends on the scorer settings.
Refactor the Scorer to improve flexibility (#5731) * Refactor the Scorer to improve flexibility Refactor the `Scorer` to improve flexibility for arbitrary pipeline components. * Individual pipeline components provide their own `evaluate` methods that score a list of `Example`s and return a dictionary of scores * `Scorer` is initialized either: * with a provided pipeline containing components to be scored * with a default pipeline containing the built-in statistical components (senter, tagger, morphologizer, parser, ner) * `Scorer.score` evaluates a list of `Example`s and returns a dictionary of scores referring to the scores provided by the components in the pipeline Significant differences: * `tags_acc` is renamed to `tag_acc` to be consistent with `token_acc` and the new `morph_acc`, `pos_acc`, and `lemma_acc` * Scoring is no longer cumulative: `Scorer.score` scores a list of examples rather than a single example and does not retain any state about previously scored examples * PRF values in the returned scores are no longer multiplied by 100 * Add kwargs to Morphologizer.evaluate * Create generalized scoring methods in Scorer * Generalized static scoring methods are added to `Scorer` * Methods require an attribute (either on Token or Doc) that is used to key the returned scores Naming differences: * `uas`, `las`, and `las_per_type` in the scores dict are renamed to `dep_uas`, `dep_las`, and `dep_las_per_type` Scoring differences: * `Doc.sents` is now scored as spans rather than on sentence-initial token positions so that `Doc.sents` and `Doc.ents` can be scored with the same method (this lowers scores since a single incorrect sentence start results in two incorrect spans) * Simplify / extend hasattr check for eval method * Add hasattr check to tokenizer scoring * Simplify to hasattr check for component scoring * Reset Example alignment if docs are set Reset the Example alignment if either doc is set in case the tokenization has changed. * Add PRF tokenization scoring for tokens as spans Add PRF scores for tokens as character spans. The scores are: * token_acc: # correct tokens / # gold tokens * token_p/r/f: PRF for (token.idx, token.idx + len(token)) * Add docstring to Scorer.score_tokenization * Rename component.evaluate() to component.score() * Update Scorer API docs * Update scoring for positive_label in textcat * Fix TextCategorizer.score kwargs * Update Language.evaluate docs * Update score names in default config
2020-07-25 13:53:02 +03:00
examples (Iterable[Example]): Examples to score
attr (str): The attribute to score.
2020-07-29 11:39:33 +03:00
getter (Callable[[Doc, str], Any]): Defaults to getattr. If provided,
Refactor the Scorer to improve flexibility (#5731) * Refactor the Scorer to improve flexibility Refactor the `Scorer` to improve flexibility for arbitrary pipeline components. * Individual pipeline components provide their own `evaluate` methods that score a list of `Example`s and return a dictionary of scores * `Scorer` is initialized either: * with a provided pipeline containing components to be scored * with a default pipeline containing the built-in statistical components (senter, tagger, morphologizer, parser, ner) * `Scorer.score` evaluates a list of `Example`s and returns a dictionary of scores referring to the scores provided by the components in the pipeline Significant differences: * `tags_acc` is renamed to `tag_acc` to be consistent with `token_acc` and the new `morph_acc`, `pos_acc`, and `lemma_acc` * Scoring is no longer cumulative: `Scorer.score` scores a list of examples rather than a single example and does not retain any state about previously scored examples * PRF values in the returned scores are no longer multiplied by 100 * Add kwargs to Morphologizer.evaluate * Create generalized scoring methods in Scorer * Generalized static scoring methods are added to `Scorer` * Methods require an attribute (either on Token or Doc) that is used to key the returned scores Naming differences: * `uas`, `las`, and `las_per_type` in the scores dict are renamed to `dep_uas`, `dep_las`, and `dep_las_per_type` Scoring differences: * `Doc.sents` is now scored as spans rather than on sentence-initial token positions so that `Doc.sents` and `Doc.ents` can be scored with the same method (this lowers scores since a single incorrect sentence start results in two incorrect spans) * Simplify / extend hasattr check for eval method * Add hasattr check to tokenizer scoring * Simplify to hasattr check for component scoring * Reset Example alignment if docs are set Reset the Example alignment if either doc is set in case the tokenization has changed. * Add PRF tokenization scoring for tokens as spans Add PRF scores for tokens as character spans. The scores are: * token_acc: # correct tokens / # gold tokens * token_p/r/f: PRF for (token.idx, token.idx + len(token)) * Add docstring to Scorer.score_tokenization * Rename component.evaluate() to component.score() * Update Scorer API docs * Update scoring for positive_label in textcat * Fix TextCategorizer.score kwargs * Update Language.evaluate docs * Update score names in default config
2020-07-25 13:53:02 +03:00
getter(doc, attr) should return the values for the individual doc.
labels (Iterable[str]): The set of possible labels. Defaults to [].
multi_label (bool): Whether the attribute allows multiple labels.
Defaults to True.
positive_label (str): The positive label for a binary task with
exclusive classes. Defaults to None.
threshold (float): Cutoff to consider a prediction "positive". Defaults
to 0.5 for multi-label, and 0.0 (i.e. whatever's highest scoring)
otherwise.
2020-07-28 22:39:42 +03:00
RETURNS (Dict[str, Any]): A dictionary containing the scores, with
inapplicable scores as None:
for all:
attr_score (one of attr_micro_f / attr_macro_f / attr_macro_auc),
attr_score_desc (text description of the overall score),
attr_micro_f,
attr_macro_f,
attr_auc,
attr_f_per_type,
attr_auc_per_type
2020-07-28 22:39:42 +03:00
DOCS: https://spacy.io/api/scorer#score_cats
2019-05-24 15:06:04 +03:00
"""
if threshold is None:
threshold = 0.5 if multi_label else 0.0
f_per_type = {label: PRFScore() for label in labels}
auc_per_type = {label: ROCAUCScore() for label in labels}
labels = set(labels)
if labels:
for eg in examples:
labels.update(eg.predicted.cats.keys())
labels.update(eg.reference.cats.keys())
Refactor the Scorer to improve flexibility (#5731) * Refactor the Scorer to improve flexibility Refactor the `Scorer` to improve flexibility for arbitrary pipeline components. * Individual pipeline components provide their own `evaluate` methods that score a list of `Example`s and return a dictionary of scores * `Scorer` is initialized either: * with a provided pipeline containing components to be scored * with a default pipeline containing the built-in statistical components (senter, tagger, morphologizer, parser, ner) * `Scorer.score` evaluates a list of `Example`s and returns a dictionary of scores referring to the scores provided by the components in the pipeline Significant differences: * `tags_acc` is renamed to `tag_acc` to be consistent with `token_acc` and the new `morph_acc`, `pos_acc`, and `lemma_acc` * Scoring is no longer cumulative: `Scorer.score` scores a list of examples rather than a single example and does not retain any state about previously scored examples * PRF values in the returned scores are no longer multiplied by 100 * Add kwargs to Morphologizer.evaluate * Create generalized scoring methods in Scorer * Generalized static scoring methods are added to `Scorer` * Methods require an attribute (either on Token or Doc) that is used to key the returned scores Naming differences: * `uas`, `las`, and `las_per_type` in the scores dict are renamed to `dep_uas`, `dep_las`, and `dep_las_per_type` Scoring differences: * `Doc.sents` is now scored as spans rather than on sentence-initial token positions so that `Doc.sents` and `Doc.ents` can be scored with the same method (this lowers scores since a single incorrect sentence start results in two incorrect spans) * Simplify / extend hasattr check for eval method * Add hasattr check to tokenizer scoring * Simplify to hasattr check for component scoring * Reset Example alignment if docs are set Reset the Example alignment if either doc is set in case the tokenization has changed. * Add PRF tokenization scoring for tokens as spans Add PRF scores for tokens as character spans. The scores are: * token_acc: # correct tokens / # gold tokens * token_p/r/f: PRF for (token.idx, token.idx + len(token)) * Add docstring to Scorer.score_tokenization * Rename component.evaluate() to component.score() * Update Scorer API docs * Update scoring for positive_label in textcat * Fix TextCategorizer.score kwargs * Update Language.evaluate docs * Update score names in default config
2020-07-25 13:53:02 +03:00
for example in examples:
# Through this loop, None in the gold_cats indicates missing label.
pred_cats = getter(example.predicted, attr)
gold_cats = getter(example.reference, attr)
# I think the AUC metric is applicable regardless of whether we're
# doing multi-label classification? Unsure. If not, move this into
# the elif pred_cats and gold_cats block below.
for label in labels:
pred_score = pred_cats.get(label, 0.0)
gold_score = gold_cats.get(label, 0.0)
if gold_score is not None:
auc_per_type[label].score_set(pred_score, gold_score)
if multi_label:
for label in labels:
pred_score = pred_cats.get(label, 0.0)
gold_score = gold_cats.get(label, 0.0)
if gold_score is not None:
if pred_score >= threshold and gold_score > 0:
f_per_type[label].tp += 1
elif pred_score >= threshold and gold_score == 0:
f_per_type[label].fp += 1
elif pred_score < threshold and gold_score > 0:
f_per_type[label].fn += 1
elif pred_cats and gold_cats:
# Get the highest-scoring for each.
pred_label, pred_score = max(pred_cats.items(), key=lambda it: it[1])
gold_label, gold_score = max(gold_cats.items(), key=lambda it: it[1])
if gold_score is not None:
if pred_label == gold_label and pred_score >= threshold:
f_per_type[pred_label].tp += 1
else:
f_per_type[gold_label].fn += 1
if pred_score >= threshold:
f_per_type[pred_label].fp += 1
elif gold_cats:
gold_label, gold_score = max(gold_cats, key=lambda it: it[1])
if gold_score is not None and gold_score > 0:
f_per_type[gold_label].fn += 1
else:
pred_label, pred_score = max(pred_cats, key=lambda it: it[1])
if pred_score >= threshold:
f_per_type[pred_label].fp += 1
micro_prf = PRFScore()
for label_prf in f_per_type.values():
micro_prf.tp = label_prf.tp
micro_prf.fn = label_prf.fn
micro_prf.fp = label_prf.fp
n_cats = len(f_per_type) + 1e-100
macro_p = sum(prf.precision for prf in f_per_type.values()) / n_cats
macro_r = sum(prf.recall for prf in f_per_type.values()) / n_cats
macro_f = sum(prf.fscore for prf in f_per_type.values()) / n_cats
Refactor the Scorer to improve flexibility (#5731) * Refactor the Scorer to improve flexibility Refactor the `Scorer` to improve flexibility for arbitrary pipeline components. * Individual pipeline components provide their own `evaluate` methods that score a list of `Example`s and return a dictionary of scores * `Scorer` is initialized either: * with a provided pipeline containing components to be scored * with a default pipeline containing the built-in statistical components (senter, tagger, morphologizer, parser, ner) * `Scorer.score` evaluates a list of `Example`s and returns a dictionary of scores referring to the scores provided by the components in the pipeline Significant differences: * `tags_acc` is renamed to `tag_acc` to be consistent with `token_acc` and the new `morph_acc`, `pos_acc`, and `lemma_acc` * Scoring is no longer cumulative: `Scorer.score` scores a list of examples rather than a single example and does not retain any state about previously scored examples * PRF values in the returned scores are no longer multiplied by 100 * Add kwargs to Morphologizer.evaluate * Create generalized scoring methods in Scorer * Generalized static scoring methods are added to `Scorer` * Methods require an attribute (either on Token or Doc) that is used to key the returned scores Naming differences: * `uas`, `las`, and `las_per_type` in the scores dict are renamed to `dep_uas`, `dep_las`, and `dep_las_per_type` Scoring differences: * `Doc.sents` is now scored as spans rather than on sentence-initial token positions so that `Doc.sents` and `Doc.ents` can be scored with the same method (this lowers scores since a single incorrect sentence start results in two incorrect spans) * Simplify / extend hasattr check for eval method * Add hasattr check to tokenizer scoring * Simplify to hasattr check for component scoring * Reset Example alignment if docs are set Reset the Example alignment if either doc is set in case the tokenization has changed. * Add PRF tokenization scoring for tokens as spans Add PRF scores for tokens as character spans. The scores are: * token_acc: # correct tokens / # gold tokens * token_p/r/f: PRF for (token.idx, token.idx + len(token)) * Add docstring to Scorer.score_tokenization * Rename component.evaluate() to component.score() * Update Scorer API docs * Update scoring for positive_label in textcat * Fix TextCategorizer.score kwargs * Update Language.evaluate docs * Update score names in default config
2020-07-25 13:53:02 +03:00
results = {
2020-07-28 22:39:42 +03:00
f"{attr}_score": None,
f"{attr}_score_desc": None,
f"{attr}_micro_p": micro_prf.precision,
f"{attr}_micro_r": micro_prf.recall,
f"{attr}_micro_f": micro_prf.fscore,
f"{attr}_macro_p": macro_p,
f"{attr}_macro_r": macro_r,
f"{attr}_macro_f": macro_f,
2020-07-28 22:39:42 +03:00
f"{attr}_macro_auc": None,
f"{attr}_f_per_type": {k: v.to_dict() for k, v in f_per_type.items()},
f"{attr}_auc_per_type": {k: v.score for k, v in auc_per_type.items()},
Refactor the Scorer to improve flexibility (#5731) * Refactor the Scorer to improve flexibility Refactor the `Scorer` to improve flexibility for arbitrary pipeline components. * Individual pipeline components provide their own `evaluate` methods that score a list of `Example`s and return a dictionary of scores * `Scorer` is initialized either: * with a provided pipeline containing components to be scored * with a default pipeline containing the built-in statistical components (senter, tagger, morphologizer, parser, ner) * `Scorer.score` evaluates a list of `Example`s and returns a dictionary of scores referring to the scores provided by the components in the pipeline Significant differences: * `tags_acc` is renamed to `tag_acc` to be consistent with `token_acc` and the new `morph_acc`, `pos_acc`, and `lemma_acc` * Scoring is no longer cumulative: `Scorer.score` scores a list of examples rather than a single example and does not retain any state about previously scored examples * PRF values in the returned scores are no longer multiplied by 100 * Add kwargs to Morphologizer.evaluate * Create generalized scoring methods in Scorer * Generalized static scoring methods are added to `Scorer` * Methods require an attribute (either on Token or Doc) that is used to key the returned scores Naming differences: * `uas`, `las`, and `las_per_type` in the scores dict are renamed to `dep_uas`, `dep_las`, and `dep_las_per_type` Scoring differences: * `Doc.sents` is now scored as spans rather than on sentence-initial token positions so that `Doc.sents` and `Doc.ents` can be scored with the same method (this lowers scores since a single incorrect sentence start results in two incorrect spans) * Simplify / extend hasattr check for eval method * Add hasattr check to tokenizer scoring * Simplify to hasattr check for component scoring * Reset Example alignment if docs are set Reset the Example alignment if either doc is set in case the tokenization has changed. * Add PRF tokenization scoring for tokens as spans Add PRF scores for tokens as character spans. The scores are: * token_acc: # correct tokens / # gold tokens * token_p/r/f: PRF for (token.idx, token.idx + len(token)) * Add docstring to Scorer.score_tokenization * Rename component.evaluate() to component.score() * Update Scorer API docs * Update scoring for positive_label in textcat * Fix TextCategorizer.score kwargs * Update Language.evaluate docs * Update score names in default config
2020-07-25 13:53:02 +03:00
}
if len(labels) == 2 and not multi_label and positive_label:
2020-08-09 23:36:23 +03:00
positive_label_f = results[f"{attr}_f_per_type"][positive_label]["f"]
results[f"{attr}_score"] = positive_label_f
2020-07-28 22:39:42 +03:00
results[f"{attr}_score_desc"] = f"F ({positive_label})"
Refactor the Scorer to improve flexibility (#5731) * Refactor the Scorer to improve flexibility Refactor the `Scorer` to improve flexibility for arbitrary pipeline components. * Individual pipeline components provide their own `evaluate` methods that score a list of `Example`s and return a dictionary of scores * `Scorer` is initialized either: * with a provided pipeline containing components to be scored * with a default pipeline containing the built-in statistical components (senter, tagger, morphologizer, parser, ner) * `Scorer.score` evaluates a list of `Example`s and returns a dictionary of scores referring to the scores provided by the components in the pipeline Significant differences: * `tags_acc` is renamed to `tag_acc` to be consistent with `token_acc` and the new `morph_acc`, `pos_acc`, and `lemma_acc` * Scoring is no longer cumulative: `Scorer.score` scores a list of examples rather than a single example and does not retain any state about previously scored examples * PRF values in the returned scores are no longer multiplied by 100 * Add kwargs to Morphologizer.evaluate * Create generalized scoring methods in Scorer * Generalized static scoring methods are added to `Scorer` * Methods require an attribute (either on Token or Doc) that is used to key the returned scores Naming differences: * `uas`, `las`, and `las_per_type` in the scores dict are renamed to `dep_uas`, `dep_las`, and `dep_las_per_type` Scoring differences: * `Doc.sents` is now scored as spans rather than on sentence-initial token positions so that `Doc.sents` and `Doc.ents` can be scored with the same method (this lowers scores since a single incorrect sentence start results in two incorrect spans) * Simplify / extend hasattr check for eval method * Add hasattr check to tokenizer scoring * Simplify to hasattr check for component scoring * Reset Example alignment if docs are set Reset the Example alignment if either doc is set in case the tokenization has changed. * Add PRF tokenization scoring for tokens as spans Add PRF scores for tokens as character spans. The scores are: * token_acc: # correct tokens / # gold tokens * token_p/r/f: PRF for (token.idx, token.idx + len(token)) * Add docstring to Scorer.score_tokenization * Rename component.evaluate() to component.score() * Update Scorer API docs * Update scoring for positive_label in textcat * Fix TextCategorizer.score kwargs * Update Language.evaluate docs * Update score names in default config
2020-07-25 13:53:02 +03:00
elif not multi_label:
2020-07-28 22:39:42 +03:00
results[f"{attr}_score"] = results[f"{attr}_macro_f"]
results[f"{attr}_score_desc"] = "macro F"
Refactor the Scorer to improve flexibility (#5731) * Refactor the Scorer to improve flexibility Refactor the `Scorer` to improve flexibility for arbitrary pipeline components. * Individual pipeline components provide their own `evaluate` methods that score a list of `Example`s and return a dictionary of scores * `Scorer` is initialized either: * with a provided pipeline containing components to be scored * with a default pipeline containing the built-in statistical components (senter, tagger, morphologizer, parser, ner) * `Scorer.score` evaluates a list of `Example`s and returns a dictionary of scores referring to the scores provided by the components in the pipeline Significant differences: * `tags_acc` is renamed to `tag_acc` to be consistent with `token_acc` and the new `morph_acc`, `pos_acc`, and `lemma_acc` * Scoring is no longer cumulative: `Scorer.score` scores a list of examples rather than a single example and does not retain any state about previously scored examples * PRF values in the returned scores are no longer multiplied by 100 * Add kwargs to Morphologizer.evaluate * Create generalized scoring methods in Scorer * Generalized static scoring methods are added to `Scorer` * Methods require an attribute (either on Token or Doc) that is used to key the returned scores Naming differences: * `uas`, `las`, and `las_per_type` in the scores dict are renamed to `dep_uas`, `dep_las`, and `dep_las_per_type` Scoring differences: * `Doc.sents` is now scored as spans rather than on sentence-initial token positions so that `Doc.sents` and `Doc.ents` can be scored with the same method (this lowers scores since a single incorrect sentence start results in two incorrect spans) * Simplify / extend hasattr check for eval method * Add hasattr check to tokenizer scoring * Simplify to hasattr check for component scoring * Reset Example alignment if docs are set Reset the Example alignment if either doc is set in case the tokenization has changed. * Add PRF tokenization scoring for tokens as spans Add PRF scores for tokens as character spans. The scores are: * token_acc: # correct tokens / # gold tokens * token_p/r/f: PRF for (token.idx, token.idx + len(token)) * Add docstring to Scorer.score_tokenization * Rename component.evaluate() to component.score() * Update Scorer API docs * Update scoring for positive_label in textcat * Fix TextCategorizer.score kwargs * Update Language.evaluate docs * Update score names in default config
2020-07-25 13:53:02 +03:00
else:
2020-07-28 22:39:42 +03:00
results[f"{attr}_score"] = results[f"{attr}_macro_auc"]
results[f"{attr}_score_desc"] = "macro AUC"
Refactor the Scorer to improve flexibility (#5731) * Refactor the Scorer to improve flexibility Refactor the `Scorer` to improve flexibility for arbitrary pipeline components. * Individual pipeline components provide their own `evaluate` methods that score a list of `Example`s and return a dictionary of scores * `Scorer` is initialized either: * with a provided pipeline containing components to be scored * with a default pipeline containing the built-in statistical components (senter, tagger, morphologizer, parser, ner) * `Scorer.score` evaluates a list of `Example`s and returns a dictionary of scores referring to the scores provided by the components in the pipeline Significant differences: * `tags_acc` is renamed to `tag_acc` to be consistent with `token_acc` and the new `morph_acc`, `pos_acc`, and `lemma_acc` * Scoring is no longer cumulative: `Scorer.score` scores a list of examples rather than a single example and does not retain any state about previously scored examples * PRF values in the returned scores are no longer multiplied by 100 * Add kwargs to Morphologizer.evaluate * Create generalized scoring methods in Scorer * Generalized static scoring methods are added to `Scorer` * Methods require an attribute (either on Token or Doc) that is used to key the returned scores Naming differences: * `uas`, `las`, and `las_per_type` in the scores dict are renamed to `dep_uas`, `dep_las`, and `dep_las_per_type` Scoring differences: * `Doc.sents` is now scored as spans rather than on sentence-initial token positions so that `Doc.sents` and `Doc.ents` can be scored with the same method (this lowers scores since a single incorrect sentence start results in two incorrect spans) * Simplify / extend hasattr check for eval method * Add hasattr check to tokenizer scoring * Simplify to hasattr check for component scoring * Reset Example alignment if docs are set Reset the Example alignment if either doc is set in case the tokenization has changed. * Add PRF tokenization scoring for tokens as spans Add PRF scores for tokens as character spans. The scores are: * token_acc: # correct tokens / # gold tokens * token_p/r/f: PRF for (token.idx, token.idx + len(token)) * Add docstring to Scorer.score_tokenization * Rename component.evaluate() to component.score() * Update Scorer API docs * Update scoring for positive_label in textcat * Fix TextCategorizer.score kwargs * Update Language.evaluate docs * Update score names in default config
2020-07-25 13:53:02 +03:00
return results
@staticmethod
def score_deps(
2020-07-28 22:39:42 +03:00
examples: Iterable[Example],
attr: str,
*,
getter: Callable[[Token, str], Any] = getattr,
head_attr: str = "head",
2020-08-17 17:45:24 +03:00
head_getter: Callable[[Token, str], Token] = getattr,
2020-07-28 22:39:42 +03:00
ignore_labels: Tuple[str] = tuple(),
**cfg,
) -> Dict[str, Any]:
Refactor the Scorer to improve flexibility (#5731) * Refactor the Scorer to improve flexibility Refactor the `Scorer` to improve flexibility for arbitrary pipeline components. * Individual pipeline components provide their own `evaluate` methods that score a list of `Example`s and return a dictionary of scores * `Scorer` is initialized either: * with a provided pipeline containing components to be scored * with a default pipeline containing the built-in statistical components (senter, tagger, morphologizer, parser, ner) * `Scorer.score` evaluates a list of `Example`s and returns a dictionary of scores referring to the scores provided by the components in the pipeline Significant differences: * `tags_acc` is renamed to `tag_acc` to be consistent with `token_acc` and the new `morph_acc`, `pos_acc`, and `lemma_acc` * Scoring is no longer cumulative: `Scorer.score` scores a list of examples rather than a single example and does not retain any state about previously scored examples * PRF values in the returned scores are no longer multiplied by 100 * Add kwargs to Morphologizer.evaluate * Create generalized scoring methods in Scorer * Generalized static scoring methods are added to `Scorer` * Methods require an attribute (either on Token or Doc) that is used to key the returned scores Naming differences: * `uas`, `las`, and `las_per_type` in the scores dict are renamed to `dep_uas`, `dep_las`, and `dep_las_per_type` Scoring differences: * `Doc.sents` is now scored as spans rather than on sentence-initial token positions so that `Doc.sents` and `Doc.ents` can be scored with the same method (this lowers scores since a single incorrect sentence start results in two incorrect spans) * Simplify / extend hasattr check for eval method * Add hasattr check to tokenizer scoring * Simplify to hasattr check for component scoring * Reset Example alignment if docs are set Reset the Example alignment if either doc is set in case the tokenization has changed. * Add PRF tokenization scoring for tokens as spans Add PRF scores for tokens as character spans. The scores are: * token_acc: # correct tokens / # gold tokens * token_p/r/f: PRF for (token.idx, token.idx + len(token)) * Add docstring to Scorer.score_tokenization * Rename component.evaluate() to component.score() * Update Scorer API docs * Update scoring for positive_label in textcat * Fix TextCategorizer.score kwargs * Update Language.evaluate docs * Update score names in default config
2020-07-25 13:53:02 +03:00
"""Returns the UAS, LAS, and LAS per type scores for dependency
parses.
examples (Iterable[Example]): Examples to score
attr (str): The attribute containing the dependency label.
2020-07-28 22:39:42 +03:00
getter (Callable[[Token, str], Any]): Defaults to getattr. If provided,
Refactor the Scorer to improve flexibility (#5731) * Refactor the Scorer to improve flexibility Refactor the `Scorer` to improve flexibility for arbitrary pipeline components. * Individual pipeline components provide their own `evaluate` methods that score a list of `Example`s and return a dictionary of scores * `Scorer` is initialized either: * with a provided pipeline containing components to be scored * with a default pipeline containing the built-in statistical components (senter, tagger, morphologizer, parser, ner) * `Scorer.score` evaluates a list of `Example`s and returns a dictionary of scores referring to the scores provided by the components in the pipeline Significant differences: * `tags_acc` is renamed to `tag_acc` to be consistent with `token_acc` and the new `morph_acc`, `pos_acc`, and `lemma_acc` * Scoring is no longer cumulative: `Scorer.score` scores a list of examples rather than a single example and does not retain any state about previously scored examples * PRF values in the returned scores are no longer multiplied by 100 * Add kwargs to Morphologizer.evaluate * Create generalized scoring methods in Scorer * Generalized static scoring methods are added to `Scorer` * Methods require an attribute (either on Token or Doc) that is used to key the returned scores Naming differences: * `uas`, `las`, and `las_per_type` in the scores dict are renamed to `dep_uas`, `dep_las`, and `dep_las_per_type` Scoring differences: * `Doc.sents` is now scored as spans rather than on sentence-initial token positions so that `Doc.sents` and `Doc.ents` can be scored with the same method (this lowers scores since a single incorrect sentence start results in two incorrect spans) * Simplify / extend hasattr check for eval method * Add hasattr check to tokenizer scoring * Simplify to hasattr check for component scoring * Reset Example alignment if docs are set Reset the Example alignment if either doc is set in case the tokenization has changed. * Add PRF tokenization scoring for tokens as spans Add PRF scores for tokens as character spans. The scores are: * token_acc: # correct tokens / # gold tokens * token_p/r/f: PRF for (token.idx, token.idx + len(token)) * Add docstring to Scorer.score_tokenization * Rename component.evaluate() to component.score() * Update Scorer API docs * Update scoring for positive_label in textcat * Fix TextCategorizer.score kwargs * Update Language.evaluate docs * Update score names in default config
2020-07-25 13:53:02 +03:00
getter(token, attr) should return the value of the attribute for an
individual token.
head_attr (str): The attribute containing the head token. Defaults to
'head'.
2020-08-17 17:45:24 +03:00
head_getter (Callable[[Token, str], Token]): Defaults to getattr. If provided,
Refactor the Scorer to improve flexibility (#5731) * Refactor the Scorer to improve flexibility Refactor the `Scorer` to improve flexibility for arbitrary pipeline components. * Individual pipeline components provide their own `evaluate` methods that score a list of `Example`s and return a dictionary of scores * `Scorer` is initialized either: * with a provided pipeline containing components to be scored * with a default pipeline containing the built-in statistical components (senter, tagger, morphologizer, parser, ner) * `Scorer.score` evaluates a list of `Example`s and returns a dictionary of scores referring to the scores provided by the components in the pipeline Significant differences: * `tags_acc` is renamed to `tag_acc` to be consistent with `token_acc` and the new `morph_acc`, `pos_acc`, and `lemma_acc` * Scoring is no longer cumulative: `Scorer.score` scores a list of examples rather than a single example and does not retain any state about previously scored examples * PRF values in the returned scores are no longer multiplied by 100 * Add kwargs to Morphologizer.evaluate * Create generalized scoring methods in Scorer * Generalized static scoring methods are added to `Scorer` * Methods require an attribute (either on Token or Doc) that is used to key the returned scores Naming differences: * `uas`, `las`, and `las_per_type` in the scores dict are renamed to `dep_uas`, `dep_las`, and `dep_las_per_type` Scoring differences: * `Doc.sents` is now scored as spans rather than on sentence-initial token positions so that `Doc.sents` and `Doc.ents` can be scored with the same method (this lowers scores since a single incorrect sentence start results in two incorrect spans) * Simplify / extend hasattr check for eval method * Add hasattr check to tokenizer scoring * Simplify to hasattr check for component scoring * Reset Example alignment if docs are set Reset the Example alignment if either doc is set in case the tokenization has changed. * Add PRF tokenization scoring for tokens as spans Add PRF scores for tokens as character spans. The scores are: * token_acc: # correct tokens / # gold tokens * token_p/r/f: PRF for (token.idx, token.idx + len(token)) * Add docstring to Scorer.score_tokenization * Rename component.evaluate() to component.score() * Update Scorer API docs * Update scoring for positive_label in textcat * Fix TextCategorizer.score kwargs * Update Language.evaluate docs * Update score names in default config
2020-07-25 13:53:02 +03:00
head_getter(token, attr) should return the value of the head for an
individual token.
ignore_labels (Tuple): Labels to ignore while scoring (e.g., punct).
2020-07-28 22:39:42 +03:00
RETURNS (Dict[str, Any]): A dictionary containing the scores:
Refactor the Scorer to improve flexibility (#5731) * Refactor the Scorer to improve flexibility Refactor the `Scorer` to improve flexibility for arbitrary pipeline components. * Individual pipeline components provide their own `evaluate` methods that score a list of `Example`s and return a dictionary of scores * `Scorer` is initialized either: * with a provided pipeline containing components to be scored * with a default pipeline containing the built-in statistical components (senter, tagger, morphologizer, parser, ner) * `Scorer.score` evaluates a list of `Example`s and returns a dictionary of scores referring to the scores provided by the components in the pipeline Significant differences: * `tags_acc` is renamed to `tag_acc` to be consistent with `token_acc` and the new `morph_acc`, `pos_acc`, and `lemma_acc` * Scoring is no longer cumulative: `Scorer.score` scores a list of examples rather than a single example and does not retain any state about previously scored examples * PRF values in the returned scores are no longer multiplied by 100 * Add kwargs to Morphologizer.evaluate * Create generalized scoring methods in Scorer * Generalized static scoring methods are added to `Scorer` * Methods require an attribute (either on Token or Doc) that is used to key the returned scores Naming differences: * `uas`, `las`, and `las_per_type` in the scores dict are renamed to `dep_uas`, `dep_las`, and `dep_las_per_type` Scoring differences: * `Doc.sents` is now scored as spans rather than on sentence-initial token positions so that `Doc.sents` and `Doc.ents` can be scored with the same method (this lowers scores since a single incorrect sentence start results in two incorrect spans) * Simplify / extend hasattr check for eval method * Add hasattr check to tokenizer scoring * Simplify to hasattr check for component scoring * Reset Example alignment if docs are set Reset the Example alignment if either doc is set in case the tokenization has changed. * Add PRF tokenization scoring for tokens as spans Add PRF scores for tokens as character spans. The scores are: * token_acc: # correct tokens / # gold tokens * token_p/r/f: PRF for (token.idx, token.idx + len(token)) * Add docstring to Scorer.score_tokenization * Rename component.evaluate() to component.score() * Update Scorer API docs * Update scoring for positive_label in textcat * Fix TextCategorizer.score kwargs * Update Language.evaluate docs * Update score names in default config
2020-07-25 13:53:02 +03:00
attr_uas, attr_las, and attr_las_per_type.
2020-07-28 22:39:42 +03:00
DOCS: https://spacy.io/api/scorer#score_deps
Refactor the Scorer to improve flexibility (#5731) * Refactor the Scorer to improve flexibility Refactor the `Scorer` to improve flexibility for arbitrary pipeline components. * Individual pipeline components provide their own `evaluate` methods that score a list of `Example`s and return a dictionary of scores * `Scorer` is initialized either: * with a provided pipeline containing components to be scored * with a default pipeline containing the built-in statistical components (senter, tagger, morphologizer, parser, ner) * `Scorer.score` evaluates a list of `Example`s and returns a dictionary of scores referring to the scores provided by the components in the pipeline Significant differences: * `tags_acc` is renamed to `tag_acc` to be consistent with `token_acc` and the new `morph_acc`, `pos_acc`, and `lemma_acc` * Scoring is no longer cumulative: `Scorer.score` scores a list of examples rather than a single example and does not retain any state about previously scored examples * PRF values in the returned scores are no longer multiplied by 100 * Add kwargs to Morphologizer.evaluate * Create generalized scoring methods in Scorer * Generalized static scoring methods are added to `Scorer` * Methods require an attribute (either on Token or Doc) that is used to key the returned scores Naming differences: * `uas`, `las`, and `las_per_type` in the scores dict are renamed to `dep_uas`, `dep_las`, and `dep_las_per_type` Scoring differences: * `Doc.sents` is now scored as spans rather than on sentence-initial token positions so that `Doc.sents` and `Doc.ents` can be scored with the same method (this lowers scores since a single incorrect sentence start results in two incorrect spans) * Simplify / extend hasattr check for eval method * Add hasattr check to tokenizer scoring * Simplify to hasattr check for component scoring * Reset Example alignment if docs are set Reset the Example alignment if either doc is set in case the tokenization has changed. * Add PRF tokenization scoring for tokens as spans Add PRF scores for tokens as character spans. The scores are: * token_acc: # correct tokens / # gold tokens * token_p/r/f: PRF for (token.idx, token.idx + len(token)) * Add docstring to Scorer.score_tokenization * Rename component.evaluate() to component.score() * Update Scorer API docs * Update scoring for positive_label in textcat * Fix TextCategorizer.score kwargs * Update Language.evaluate docs * Update score names in default config
2020-07-25 13:53:02 +03:00
"""
unlabelled = PRFScore()
labelled = PRFScore()
labelled_per_dep = dict()
for example in examples:
gold_doc = example.reference
pred_doc = example.predicted
align = example.alignment
gold_deps = set()
gold_deps_per_dep = {}
for gold_i, token in enumerate(gold_doc):
dep = getter(token, attr)
head = head_getter(token, head_attr)
if dep not in ignore_labels:
gold_deps.add((gold_i, head.i, dep))
if dep not in labelled_per_dep:
labelled_per_dep[dep] = PRFScore()
if dep not in gold_deps_per_dep:
gold_deps_per_dep[dep] = set()
gold_deps_per_dep[dep].add((gold_i, head.i, dep))
pred_deps = set()
pred_deps_per_dep = {}
for token in pred_doc:
if token.orth_.isspace():
continue
if align.x2y.lengths[token.i] != 1:
gold_i = None
else:
gold_i = align.x2y[token.i].dataXd[0, 0]
dep = getter(token, attr)
head = head_getter(token, head_attr)
if dep not in ignore_labels and token.orth_.strip():
if align.x2y.lengths[head.i] == 1:
gold_head = align.x2y[head.i].dataXd[0, 0]
else:
gold_head = None
# None is indistinct, so we can't just add it to the set
# Multiple (None, None) deps are possible
if gold_i is None or gold_head is None:
unlabelled.fp += 1
labelled.fp += 1
else:
pred_deps.add((gold_i, gold_head, dep))
if dep not in labelled_per_dep:
labelled_per_dep[dep] = PRFScore()
if dep not in pred_deps_per_dep:
pred_deps_per_dep[dep] = set()
pred_deps_per_dep[dep].add((gold_i, gold_head, dep))
labelled.score_set(pred_deps, gold_deps)
for dep in labelled_per_dep:
labelled_per_dep[dep].score_set(
pred_deps_per_dep.get(dep, set()), gold_deps_per_dep.get(dep, set())
)
unlabelled.score_set(
set(item[:2] for item in pred_deps), set(item[:2] for item in gold_deps)
Add textcat to train CLI (#4226) * Add doc.cats to spacy.gold at the paragraph level Support `doc.cats` as `"cats": [{"label": string, "value": number}]` in the spacy JSON training format at the paragraph level. * `spacy.gold.docs_to_json()` writes `docs.cats` * `GoldCorpus` reads in cats in each `GoldParse` * Update instances of gold_tuples to handle cats Update iteration over gold_tuples / gold_parses to handle addition of cats at the paragraph level. * Add textcat to train CLI * Add textcat options to train CLI * Add textcat labels in `TextCategorizer.begin_training()` * Add textcat evaluation to `Scorer`: * For binary exclusive classes with provided label: F1 for label * For 2+ exclusive classes: F1 macro average * For multilabel (not exclusive): ROC AUC macro average (currently relying on sklearn) * Provide user info on textcat evaluation settings, potential incompatibilities * Provide pipeline to Scorer in `Language.evaluate` for textcat config * Customize train CLI output to include only metrics relevant to current pipeline * Add textcat evaluation to evaluate CLI * Fix handling of unset arguments and config params Fix handling of unset arguments and model confiug parameters in Scorer initialization. * Temporarily add sklearn requirement * Remove sklearn version number * Improve Scorer handling of models without textcats * Fixing Scorer handling of models without textcats * Update Scorer output for python 2.7 * Modify inf in Scorer for python 2.7 * Auto-format Also make small adjustments to make auto-formatting with black easier and produce nicer results * Move error message to Errors * Update documentation * Add cats to annotation JSON format [ci skip] * Fix tpl flag and docs [ci skip] * Switch to internal roc_auc_score Switch to internal `roc_auc_score()` adapted from scikit-learn. * Add AUCROCScore tests and improve errors/warnings * Add tests for AUCROCScore and roc_auc_score * Add missing error for only positive/negative values * Remove unnecessary warnings and errors * Make reduced roc_auc_score functions private Because most of the checks and warnings have been stripped for the internal functions and access is only intended through `ROCAUCScore`, make the functions for roc_auc_score adapted from scikit-learn private. * Check that data corresponds with multilabel flag Check that the training instances correspond with the multilabel flag, adding the multilabel flag if required. * Add textcat score to early stopping check * Add more checks to debug-data for textcat * Add example training data for textcat * Add more checks to textcat train CLI * Check configuration when extending base model * Fix typos * Update textcat example data * Provide licensing details and licenses for data * Remove two labels with no positive instances from jigsaw-toxic-comment data. Co-authored-by: Ines Montani <ines@ines.io>
2019-09-15 23:31:31 +03:00
)
Refactor the Scorer to improve flexibility (#5731) * Refactor the Scorer to improve flexibility Refactor the `Scorer` to improve flexibility for arbitrary pipeline components. * Individual pipeline components provide their own `evaluate` methods that score a list of `Example`s and return a dictionary of scores * `Scorer` is initialized either: * with a provided pipeline containing components to be scored * with a default pipeline containing the built-in statistical components (senter, tagger, morphologizer, parser, ner) * `Scorer.score` evaluates a list of `Example`s and returns a dictionary of scores referring to the scores provided by the components in the pipeline Significant differences: * `tags_acc` is renamed to `tag_acc` to be consistent with `token_acc` and the new `morph_acc`, `pos_acc`, and `lemma_acc` * Scoring is no longer cumulative: `Scorer.score` scores a list of examples rather than a single example and does not retain any state about previously scored examples * PRF values in the returned scores are no longer multiplied by 100 * Add kwargs to Morphologizer.evaluate * Create generalized scoring methods in Scorer * Generalized static scoring methods are added to `Scorer` * Methods require an attribute (either on Token or Doc) that is used to key the returned scores Naming differences: * `uas`, `las`, and `las_per_type` in the scores dict are renamed to `dep_uas`, `dep_las`, and `dep_las_per_type` Scoring differences: * `Doc.sents` is now scored as spans rather than on sentence-initial token positions so that `Doc.sents` and `Doc.ents` can be scored with the same method (this lowers scores since a single incorrect sentence start results in two incorrect spans) * Simplify / extend hasattr check for eval method * Add hasattr check to tokenizer scoring * Simplify to hasattr check for component scoring * Reset Example alignment if docs are set Reset the Example alignment if either doc is set in case the tokenization has changed. * Add PRF tokenization scoring for tokens as spans Add PRF scores for tokens as character spans. The scores are: * token_acc: # correct tokens / # gold tokens * token_p/r/f: PRF for (token.idx, token.idx + len(token)) * Add docstring to Scorer.score_tokenization * Rename component.evaluate() to component.score() * Update Scorer API docs * Update scoring for positive_label in textcat * Fix TextCategorizer.score kwargs * Update Language.evaluate docs * Update score names in default config
2020-07-25 13:53:02 +03:00
return {
2020-07-28 22:39:42 +03:00
f"{attr}_uas": unlabelled.fscore,
f"{attr}_las": labelled.fscore,
f"{attr}_las_per_type": {
k: v.to_dict() for k, v in labelled_per_dep.items()
},
Refactor the Scorer to improve flexibility (#5731) * Refactor the Scorer to improve flexibility Refactor the `Scorer` to improve flexibility for arbitrary pipeline components. * Individual pipeline components provide their own `evaluate` methods that score a list of `Example`s and return a dictionary of scores * `Scorer` is initialized either: * with a provided pipeline containing components to be scored * with a default pipeline containing the built-in statistical components (senter, tagger, morphologizer, parser, ner) * `Scorer.score` evaluates a list of `Example`s and returns a dictionary of scores referring to the scores provided by the components in the pipeline Significant differences: * `tags_acc` is renamed to `tag_acc` to be consistent with `token_acc` and the new `morph_acc`, `pos_acc`, and `lemma_acc` * Scoring is no longer cumulative: `Scorer.score` scores a list of examples rather than a single example and does not retain any state about previously scored examples * PRF values in the returned scores are no longer multiplied by 100 * Add kwargs to Morphologizer.evaluate * Create generalized scoring methods in Scorer * Generalized static scoring methods are added to `Scorer` * Methods require an attribute (either on Token or Doc) that is used to key the returned scores Naming differences: * `uas`, `las`, and `las_per_type` in the scores dict are renamed to `dep_uas`, `dep_las`, and `dep_las_per_type` Scoring differences: * `Doc.sents` is now scored as spans rather than on sentence-initial token positions so that `Doc.sents` and `Doc.ents` can be scored with the same method (this lowers scores since a single incorrect sentence start results in two incorrect spans) * Simplify / extend hasattr check for eval method * Add hasattr check to tokenizer scoring * Simplify to hasattr check for component scoring * Reset Example alignment if docs are set Reset the Example alignment if either doc is set in case the tokenization has changed. * Add PRF tokenization scoring for tokens as spans Add PRF scores for tokens as character spans. The scores are: * token_acc: # correct tokens / # gold tokens * token_p/r/f: PRF for (token.idx, token.idx + len(token)) * Add docstring to Scorer.score_tokenization * Rename component.evaluate() to component.score() * Update Scorer API docs * Update scoring for positive_label in textcat * Fix TextCategorizer.score kwargs * Update Language.evaluate docs * Update score names in default config
2020-07-25 13:53:02 +03:00
}
Add textcat to train CLI (#4226) * Add doc.cats to spacy.gold at the paragraph level Support `doc.cats` as `"cats": [{"label": string, "value": number}]` in the spacy JSON training format at the paragraph level. * `spacy.gold.docs_to_json()` writes `docs.cats` * `GoldCorpus` reads in cats in each `GoldParse` * Update instances of gold_tuples to handle cats Update iteration over gold_tuples / gold_parses to handle addition of cats at the paragraph level. * Add textcat to train CLI * Add textcat options to train CLI * Add textcat labels in `TextCategorizer.begin_training()` * Add textcat evaluation to `Scorer`: * For binary exclusive classes with provided label: F1 for label * For 2+ exclusive classes: F1 macro average * For multilabel (not exclusive): ROC AUC macro average (currently relying on sklearn) * Provide user info on textcat evaluation settings, potential incompatibilities * Provide pipeline to Scorer in `Language.evaluate` for textcat config * Customize train CLI output to include only metrics relevant to current pipeline * Add textcat evaluation to evaluate CLI * Fix handling of unset arguments and config params Fix handling of unset arguments and model confiug parameters in Scorer initialization. * Temporarily add sklearn requirement * Remove sklearn version number * Improve Scorer handling of models without textcats * Fixing Scorer handling of models without textcats * Update Scorer output for python 2.7 * Modify inf in Scorer for python 2.7 * Auto-format Also make small adjustments to make auto-formatting with black easier and produce nicer results * Move error message to Errors * Update documentation * Add cats to annotation JSON format [ci skip] * Fix tpl flag and docs [ci skip] * Switch to internal roc_auc_score Switch to internal `roc_auc_score()` adapted from scikit-learn. * Add AUCROCScore tests and improve errors/warnings * Add tests for AUCROCScore and roc_auc_score * Add missing error for only positive/negative values * Remove unnecessary warnings and errors * Make reduced roc_auc_score functions private Because most of the checks and warnings have been stripped for the internal functions and access is only intended through `ROCAUCScore`, make the functions for roc_auc_score adapted from scikit-learn private. * Check that data corresponds with multilabel flag Check that the training instances correspond with the multilabel flag, adding the multilabel flag if required. * Add textcat score to early stopping check * Add more checks to debug-data for textcat * Add example training data for textcat * Add more checks to textcat train CLI * Check configuration when extending base model * Fix typos * Update textcat example data * Provide licensing details and licenses for data * Remove two labels with no positive instances from jigsaw-toxic-comment data. Co-authored-by: Ines Montani <ines@ines.io>
2019-09-15 23:31:31 +03:00
#############################################################################
#
# The following implementation of roc_auc_score() is adapted from
# scikit-learn, which is distributed under the following license:
#
# New BSD License
#
# Copyright (c) 20072019 The scikit-learn developers.
# All rights reserved.
#
#
# Redistribution and use in source and binary forms, with or without
# modification, are permitted provided that the following conditions are met:
#
# a. Redistributions of source code must retain the above copyright notice,
# this list of conditions and the following disclaimer.
# b. Redistributions in binary form must reproduce the above copyright
# notice, this list of conditions and the following disclaimer in the
# documentation and/or other materials provided with the distribution.
# c. Neither the name of the Scikit-learn Developers nor the names of
# its contributors may be used to endorse or promote products
# derived from this software without specific prior written
2019-09-18 20:56:55 +03:00
# permission.
Add textcat to train CLI (#4226) * Add doc.cats to spacy.gold at the paragraph level Support `doc.cats` as `"cats": [{"label": string, "value": number}]` in the spacy JSON training format at the paragraph level. * `spacy.gold.docs_to_json()` writes `docs.cats` * `GoldCorpus` reads in cats in each `GoldParse` * Update instances of gold_tuples to handle cats Update iteration over gold_tuples / gold_parses to handle addition of cats at the paragraph level. * Add textcat to train CLI * Add textcat options to train CLI * Add textcat labels in `TextCategorizer.begin_training()` * Add textcat evaluation to `Scorer`: * For binary exclusive classes with provided label: F1 for label * For 2+ exclusive classes: F1 macro average * For multilabel (not exclusive): ROC AUC macro average (currently relying on sklearn) * Provide user info on textcat evaluation settings, potential incompatibilities * Provide pipeline to Scorer in `Language.evaluate` for textcat config * Customize train CLI output to include only metrics relevant to current pipeline * Add textcat evaluation to evaluate CLI * Fix handling of unset arguments and config params Fix handling of unset arguments and model confiug parameters in Scorer initialization. * Temporarily add sklearn requirement * Remove sklearn version number * Improve Scorer handling of models without textcats * Fixing Scorer handling of models without textcats * Update Scorer output for python 2.7 * Modify inf in Scorer for python 2.7 * Auto-format Also make small adjustments to make auto-formatting with black easier and produce nicer results * Move error message to Errors * Update documentation * Add cats to annotation JSON format [ci skip] * Fix tpl flag and docs [ci skip] * Switch to internal roc_auc_score Switch to internal `roc_auc_score()` adapted from scikit-learn. * Add AUCROCScore tests and improve errors/warnings * Add tests for AUCROCScore and roc_auc_score * Add missing error for only positive/negative values * Remove unnecessary warnings and errors * Make reduced roc_auc_score functions private Because most of the checks and warnings have been stripped for the internal functions and access is only intended through `ROCAUCScore`, make the functions for roc_auc_score adapted from scikit-learn private. * Check that data corresponds with multilabel flag Check that the training instances correspond with the multilabel flag, adding the multilabel flag if required. * Add textcat score to early stopping check * Add more checks to debug-data for textcat * Add example training data for textcat * Add more checks to textcat train CLI * Check configuration when extending base model * Fix typos * Update textcat example data * Provide licensing details and licenses for data * Remove two labels with no positive instances from jigsaw-toxic-comment data. Co-authored-by: Ines Montani <ines@ines.io>
2019-09-15 23:31:31 +03:00
#
#
# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
# AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
# IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
# ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE FOR
# ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
# DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
# SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
# CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
# LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
# OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH
# DAMAGE.
2019-09-18 20:56:55 +03:00
Add textcat to train CLI (#4226) * Add doc.cats to spacy.gold at the paragraph level Support `doc.cats` as `"cats": [{"label": string, "value": number}]` in the spacy JSON training format at the paragraph level. * `spacy.gold.docs_to_json()` writes `docs.cats` * `GoldCorpus` reads in cats in each `GoldParse` * Update instances of gold_tuples to handle cats Update iteration over gold_tuples / gold_parses to handle addition of cats at the paragraph level. * Add textcat to train CLI * Add textcat options to train CLI * Add textcat labels in `TextCategorizer.begin_training()` * Add textcat evaluation to `Scorer`: * For binary exclusive classes with provided label: F1 for label * For 2+ exclusive classes: F1 macro average * For multilabel (not exclusive): ROC AUC macro average (currently relying on sklearn) * Provide user info on textcat evaluation settings, potential incompatibilities * Provide pipeline to Scorer in `Language.evaluate` for textcat config * Customize train CLI output to include only metrics relevant to current pipeline * Add textcat evaluation to evaluate CLI * Fix handling of unset arguments and config params Fix handling of unset arguments and model confiug parameters in Scorer initialization. * Temporarily add sklearn requirement * Remove sklearn version number * Improve Scorer handling of models without textcats * Fixing Scorer handling of models without textcats * Update Scorer output for python 2.7 * Modify inf in Scorer for python 2.7 * Auto-format Also make small adjustments to make auto-formatting with black easier and produce nicer results * Move error message to Errors * Update documentation * Add cats to annotation JSON format [ci skip] * Fix tpl flag and docs [ci skip] * Switch to internal roc_auc_score Switch to internal `roc_auc_score()` adapted from scikit-learn. * Add AUCROCScore tests and improve errors/warnings * Add tests for AUCROCScore and roc_auc_score * Add missing error for only positive/negative values * Remove unnecessary warnings and errors * Make reduced roc_auc_score functions private Because most of the checks and warnings have been stripped for the internal functions and access is only intended through `ROCAUCScore`, make the functions for roc_auc_score adapted from scikit-learn private. * Check that data corresponds with multilabel flag Check that the training instances correspond with the multilabel flag, adding the multilabel flag if required. * Add textcat score to early stopping check * Add more checks to debug-data for textcat * Add example training data for textcat * Add more checks to textcat train CLI * Check configuration when extending base model * Fix typos * Update textcat example data * Provide licensing details and licenses for data * Remove two labels with no positive instances from jigsaw-toxic-comment data. Co-authored-by: Ines Montani <ines@ines.io>
2019-09-15 23:31:31 +03:00
def _roc_auc_score(y_true, y_score):
"""Compute Area Under the Receiver Operating Characteristic Curve (ROC AUC)
from prediction scores.
Note: this implementation is restricted to the binary classification task
Parameters
----------
y_true : array, shape = [n_samples] or [n_samples, n_classes]
True binary labels or binary label indicators.
The multiclass case expects shape = [n_samples] and labels
with values in ``range(n_classes)``.
y_score : array, shape = [n_samples] or [n_samples, n_classes]
Target scores, can either be probability estimates of the positive
class, confidence values, or non-thresholded measure of decisions
(as returned by "decision_function" on some classifiers). For binary
y_true, y_score is supposed to be the score of the class with greater
label. The multiclass case expects shape = [n_samples, n_classes]
where the scores correspond to probability estimates.
Returns
-------
auc : float
References
----------
.. [1] `Wikipedia entry for the Receiver operating characteristic
<https://en.wikipedia.org/wiki/Receiver_operating_characteristic>`_
.. [2] Fawcett T. An introduction to ROC analysis[J]. Pattern Recognition
Letters, 2006, 27(8):861-874.
.. [3] `Analyzing a portion of the ROC curve. McClish, 1989
<https://www.ncbi.nlm.nih.gov/pubmed/2668680>`_
"""
if len(np.unique(y_true)) != 2:
raise ValueError(Errors.E165)
fpr, tpr, _ = _roc_curve(y_true, y_score)
return _auc(fpr, tpr)
def _roc_curve(y_true, y_score):
"""Compute Receiver operating characteristic (ROC)
Note: this implementation is restricted to the binary classification task.
Parameters
----------
y_true : array, shape = [n_samples]
True binary labels. If labels are not either {-1, 1} or {0, 1}, then
pos_label should be explicitly given.
y_score : array, shape = [n_samples]
Target scores, can either be probability estimates of the positive
class, confidence values, or non-thresholded measure of decisions
(as returned by "decision_function" on some classifiers).
Returns
-------
fpr : array, shape = [>2]
Increasing false positive rates such that element i is the false
positive rate of predictions with score >= thresholds[i].
tpr : array, shape = [>2]
Increasing true positive rates such that element i is the true
positive rate of predictions with score >= thresholds[i].
thresholds : array, shape = [n_thresholds]
Decreasing thresholds on the decision function used to compute
fpr and tpr. `thresholds[0]` represents no instances being predicted
and is arbitrarily set to `max(y_score) + 1`.
Notes
-----
Since the thresholds are sorted from low to high values, they
are reversed upon returning them to ensure they correspond to both ``fpr``
and ``tpr``, which are sorted in reversed order during their calculation.
References
----------
.. [1] `Wikipedia entry for the Receiver operating characteristic
<https://en.wikipedia.org/wiki/Receiver_operating_characteristic>`_
.. [2] Fawcett T. An introduction to ROC analysis[J]. Pattern Recognition
Letters, 2006, 27(8):861-874.
"""
fps, tps, thresholds = _binary_clf_curve(y_true, y_score)
# Add an extra threshold position
# to make sure that the curve starts at (0, 0)
tps = np.r_[0, tps]
fps = np.r_[0, fps]
thresholds = np.r_[thresholds[0] + 1, thresholds]
if fps[-1] <= 0:
fpr = np.repeat(np.nan, fps.shape)
else:
fpr = fps / fps[-1]
if tps[-1] <= 0:
tpr = np.repeat(np.nan, tps.shape)
else:
tpr = tps / tps[-1]
return fpr, tpr, thresholds
def _binary_clf_curve(y_true, y_score):
"""Calculate true and false positives per binary classification threshold.
Parameters
----------
y_true : array, shape = [n_samples]
True targets of binary classification
y_score : array, shape = [n_samples]
Estimated probabilities or decision function
Returns
-------
fps : array, shape = [n_thresholds]
A count of false positives, at index i being the number of negative
samples assigned a score >= thresholds[i]. The total number of
negative samples is equal to fps[-1] (thus true negatives are given by
fps[-1] - fps).
tps : array, shape = [n_thresholds <= len(np.unique(y_score))]
An increasing count of true positives, at index i being the number
of positive samples assigned a score >= thresholds[i]. The total
number of positive samples is equal to tps[-1] (thus false negatives
are given by tps[-1] - tps).
thresholds : array, shape = [n_thresholds]
Decreasing score values.
"""
2019-09-18 20:56:55 +03:00
pos_label = 1.0
Add textcat to train CLI (#4226) * Add doc.cats to spacy.gold at the paragraph level Support `doc.cats` as `"cats": [{"label": string, "value": number}]` in the spacy JSON training format at the paragraph level. * `spacy.gold.docs_to_json()` writes `docs.cats` * `GoldCorpus` reads in cats in each `GoldParse` * Update instances of gold_tuples to handle cats Update iteration over gold_tuples / gold_parses to handle addition of cats at the paragraph level. * Add textcat to train CLI * Add textcat options to train CLI * Add textcat labels in `TextCategorizer.begin_training()` * Add textcat evaluation to `Scorer`: * For binary exclusive classes with provided label: F1 for label * For 2+ exclusive classes: F1 macro average * For multilabel (not exclusive): ROC AUC macro average (currently relying on sklearn) * Provide user info on textcat evaluation settings, potential incompatibilities * Provide pipeline to Scorer in `Language.evaluate` for textcat config * Customize train CLI output to include only metrics relevant to current pipeline * Add textcat evaluation to evaluate CLI * Fix handling of unset arguments and config params Fix handling of unset arguments and model confiug parameters in Scorer initialization. * Temporarily add sklearn requirement * Remove sklearn version number * Improve Scorer handling of models without textcats * Fixing Scorer handling of models without textcats * Update Scorer output for python 2.7 * Modify inf in Scorer for python 2.7 * Auto-format Also make small adjustments to make auto-formatting with black easier and produce nicer results * Move error message to Errors * Update documentation * Add cats to annotation JSON format [ci skip] * Fix tpl flag and docs [ci skip] * Switch to internal roc_auc_score Switch to internal `roc_auc_score()` adapted from scikit-learn. * Add AUCROCScore tests and improve errors/warnings * Add tests for AUCROCScore and roc_auc_score * Add missing error for only positive/negative values * Remove unnecessary warnings and errors * Make reduced roc_auc_score functions private Because most of the checks and warnings have been stripped for the internal functions and access is only intended through `ROCAUCScore`, make the functions for roc_auc_score adapted from scikit-learn private. * Check that data corresponds with multilabel flag Check that the training instances correspond with the multilabel flag, adding the multilabel flag if required. * Add textcat score to early stopping check * Add more checks to debug-data for textcat * Add example training data for textcat * Add more checks to textcat train CLI * Check configuration when extending base model * Fix typos * Update textcat example data * Provide licensing details and licenses for data * Remove two labels with no positive instances from jigsaw-toxic-comment data. Co-authored-by: Ines Montani <ines@ines.io>
2019-09-15 23:31:31 +03:00
y_true = np.ravel(y_true)
y_score = np.ravel(y_score)
# make y_true a boolean vector
2019-09-18 20:56:55 +03:00
y_true = y_true == pos_label
Add textcat to train CLI (#4226) * Add doc.cats to spacy.gold at the paragraph level Support `doc.cats` as `"cats": [{"label": string, "value": number}]` in the spacy JSON training format at the paragraph level. * `spacy.gold.docs_to_json()` writes `docs.cats` * `GoldCorpus` reads in cats in each `GoldParse` * Update instances of gold_tuples to handle cats Update iteration over gold_tuples / gold_parses to handle addition of cats at the paragraph level. * Add textcat to train CLI * Add textcat options to train CLI * Add textcat labels in `TextCategorizer.begin_training()` * Add textcat evaluation to `Scorer`: * For binary exclusive classes with provided label: F1 for label * For 2+ exclusive classes: F1 macro average * For multilabel (not exclusive): ROC AUC macro average (currently relying on sklearn) * Provide user info on textcat evaluation settings, potential incompatibilities * Provide pipeline to Scorer in `Language.evaluate` for textcat config * Customize train CLI output to include only metrics relevant to current pipeline * Add textcat evaluation to evaluate CLI * Fix handling of unset arguments and config params Fix handling of unset arguments and model confiug parameters in Scorer initialization. * Temporarily add sklearn requirement * Remove sklearn version number * Improve Scorer handling of models without textcats * Fixing Scorer handling of models without textcats * Update Scorer output for python 2.7 * Modify inf in Scorer for python 2.7 * Auto-format Also make small adjustments to make auto-formatting with black easier and produce nicer results * Move error message to Errors * Update documentation * Add cats to annotation JSON format [ci skip] * Fix tpl flag and docs [ci skip] * Switch to internal roc_auc_score Switch to internal `roc_auc_score()` adapted from scikit-learn. * Add AUCROCScore tests and improve errors/warnings * Add tests for AUCROCScore and roc_auc_score * Add missing error for only positive/negative values * Remove unnecessary warnings and errors * Make reduced roc_auc_score functions private Because most of the checks and warnings have been stripped for the internal functions and access is only intended through `ROCAUCScore`, make the functions for roc_auc_score adapted from scikit-learn private. * Check that data corresponds with multilabel flag Check that the training instances correspond with the multilabel flag, adding the multilabel flag if required. * Add textcat score to early stopping check * Add more checks to debug-data for textcat * Add example training data for textcat * Add more checks to textcat train CLI * Check configuration when extending base model * Fix typos * Update textcat example data * Provide licensing details and licenses for data * Remove two labels with no positive instances from jigsaw-toxic-comment data. Co-authored-by: Ines Montani <ines@ines.io>
2019-09-15 23:31:31 +03:00
# sort scores and corresponding truth values
desc_score_indices = np.argsort(y_score, kind="mergesort")[::-1]
y_score = y_score[desc_score_indices]
y_true = y_true[desc_score_indices]
2019-09-18 20:56:55 +03:00
weight = 1.0
Add textcat to train CLI (#4226) * Add doc.cats to spacy.gold at the paragraph level Support `doc.cats` as `"cats": [{"label": string, "value": number}]` in the spacy JSON training format at the paragraph level. * `spacy.gold.docs_to_json()` writes `docs.cats` * `GoldCorpus` reads in cats in each `GoldParse` * Update instances of gold_tuples to handle cats Update iteration over gold_tuples / gold_parses to handle addition of cats at the paragraph level. * Add textcat to train CLI * Add textcat options to train CLI * Add textcat labels in `TextCategorizer.begin_training()` * Add textcat evaluation to `Scorer`: * For binary exclusive classes with provided label: F1 for label * For 2+ exclusive classes: F1 macro average * For multilabel (not exclusive): ROC AUC macro average (currently relying on sklearn) * Provide user info on textcat evaluation settings, potential incompatibilities * Provide pipeline to Scorer in `Language.evaluate` for textcat config * Customize train CLI output to include only metrics relevant to current pipeline * Add textcat evaluation to evaluate CLI * Fix handling of unset arguments and config params Fix handling of unset arguments and model confiug parameters in Scorer initialization. * Temporarily add sklearn requirement * Remove sklearn version number * Improve Scorer handling of models without textcats * Fixing Scorer handling of models without textcats * Update Scorer output for python 2.7 * Modify inf in Scorer for python 2.7 * Auto-format Also make small adjustments to make auto-formatting with black easier and produce nicer results * Move error message to Errors * Update documentation * Add cats to annotation JSON format [ci skip] * Fix tpl flag and docs [ci skip] * Switch to internal roc_auc_score Switch to internal `roc_auc_score()` adapted from scikit-learn. * Add AUCROCScore tests and improve errors/warnings * Add tests for AUCROCScore and roc_auc_score * Add missing error for only positive/negative values * Remove unnecessary warnings and errors * Make reduced roc_auc_score functions private Because most of the checks and warnings have been stripped for the internal functions and access is only intended through `ROCAUCScore`, make the functions for roc_auc_score adapted from scikit-learn private. * Check that data corresponds with multilabel flag Check that the training instances correspond with the multilabel flag, adding the multilabel flag if required. * Add textcat score to early stopping check * Add more checks to debug-data for textcat * Add example training data for textcat * Add more checks to textcat train CLI * Check configuration when extending base model * Fix typos * Update textcat example data * Provide licensing details and licenses for data * Remove two labels with no positive instances from jigsaw-toxic-comment data. Co-authored-by: Ines Montani <ines@ines.io>
2019-09-15 23:31:31 +03:00
# y_score typically has many tied values. Here we extract
# the indices associated with the distinct values. We also
# concatenate a value for the end of the curve.
distinct_value_indices = np.where(np.diff(y_score))[0]
threshold_idxs = np.r_[distinct_value_indices, y_true.size - 1]
# accumulate the true positives with decreasing threshold
tps = _stable_cumsum(y_true * weight)[threshold_idxs]
fps = 1 + threshold_idxs - tps
return fps, tps, y_score[threshold_idxs]
def _stable_cumsum(arr, axis=None, rtol=1e-05, atol=1e-08):
"""Use high precision for cumsum and check that final value matches sum
Parameters
----------
arr : array-like
To be cumulatively summed as flat
axis : int, optional
Axis along which the cumulative sum is computed.
The default (None) is to compute the cumsum over the flattened array.
rtol : float
Relative tolerance, see ``np.allclose``
atol : float
Absolute tolerance, see ``np.allclose``
"""
out = np.cumsum(arr, axis=axis, dtype=np.float64)
expected = np.sum(arr, axis=axis, dtype=np.float64)
2019-09-18 20:56:55 +03:00
if not np.all(
np.isclose(
out.take(-1, axis=axis), expected, rtol=rtol, atol=atol, equal_nan=True
)
):
Add textcat to train CLI (#4226) * Add doc.cats to spacy.gold at the paragraph level Support `doc.cats` as `"cats": [{"label": string, "value": number}]` in the spacy JSON training format at the paragraph level. * `spacy.gold.docs_to_json()` writes `docs.cats` * `GoldCorpus` reads in cats in each `GoldParse` * Update instances of gold_tuples to handle cats Update iteration over gold_tuples / gold_parses to handle addition of cats at the paragraph level. * Add textcat to train CLI * Add textcat options to train CLI * Add textcat labels in `TextCategorizer.begin_training()` * Add textcat evaluation to `Scorer`: * For binary exclusive classes with provided label: F1 for label * For 2+ exclusive classes: F1 macro average * For multilabel (not exclusive): ROC AUC macro average (currently relying on sklearn) * Provide user info on textcat evaluation settings, potential incompatibilities * Provide pipeline to Scorer in `Language.evaluate` for textcat config * Customize train CLI output to include only metrics relevant to current pipeline * Add textcat evaluation to evaluate CLI * Fix handling of unset arguments and config params Fix handling of unset arguments and model confiug parameters in Scorer initialization. * Temporarily add sklearn requirement * Remove sklearn version number * Improve Scorer handling of models without textcats * Fixing Scorer handling of models without textcats * Update Scorer output for python 2.7 * Modify inf in Scorer for python 2.7 * Auto-format Also make small adjustments to make auto-formatting with black easier and produce nicer results * Move error message to Errors * Update documentation * Add cats to annotation JSON format [ci skip] * Fix tpl flag and docs [ci skip] * Switch to internal roc_auc_score Switch to internal `roc_auc_score()` adapted from scikit-learn. * Add AUCROCScore tests and improve errors/warnings * Add tests for AUCROCScore and roc_auc_score * Add missing error for only positive/negative values * Remove unnecessary warnings and errors * Make reduced roc_auc_score functions private Because most of the checks and warnings have been stripped for the internal functions and access is only intended through `ROCAUCScore`, make the functions for roc_auc_score adapted from scikit-learn private. * Check that data corresponds with multilabel flag Check that the training instances correspond with the multilabel flag, adding the multilabel flag if required. * Add textcat score to early stopping check * Add more checks to debug-data for textcat * Add example training data for textcat * Add more checks to textcat train CLI * Check configuration when extending base model * Fix typos * Update textcat example data * Provide licensing details and licenses for data * Remove two labels with no positive instances from jigsaw-toxic-comment data. Co-authored-by: Ines Montani <ines@ines.io>
2019-09-15 23:31:31 +03:00
raise ValueError(Errors.E163)
return out
def _auc(x, y):
"""Compute Area Under the Curve (AUC) using the trapezoidal rule
This is a general function, given points on a curve. For computing the
area under the ROC-curve, see :func:`roc_auc_score`.
Parameters
----------
x : array, shape = [n]
x coordinates. These must be either monotonic increasing or monotonic
decreasing.
y : array, shape = [n]
y coordinates.
Returns
-------
auc : float
"""
x = np.ravel(x)
y = np.ravel(y)
direction = 1
dx = np.diff(x)
if np.any(dx < 0):
if np.all(dx <= 0):
direction = -1
else:
raise ValueError(Errors.E164.format(x))
area = direction * np.trapz(y, x)
if isinstance(area, np.memmap):
# Reductions such as .sum used internally in np.trapz do not return a
# scalar by default for numpy.memmap instances contrary to
# regular numpy.ndarray instances.
area = area.dtype.type(area)
return area