explosion/spaCy
1
1
Fork 0
mirror of https://github.com/explosion/spaCy.git synced 2025-08-08 14:14:57 +03:00
Code Issues Packages Projects Releases Wiki Activity

Add beta parameter to Scorer and PRFScore.

Browse Source
This commit is contained in:
Raphael Mitsch 2022-09-01 16:52:50 +02:00
parent 51863cd267
commit ea9737a664
3 changed files with 50 additions and 32 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

21
spacy/cli/find_threshold.py
View File

@ -7,6 +7,8 @@ from typing import Optional, Tuple, Any, Dict, List
import numpy import numpy
import wasabi.tables import wasabi.tables
from ..pipeline import TrainablePipe, Pipe
from ..errors import Errors
from ..training import Corpus from ..training import Corpus
from ._util import app, Arg, Opt, import_code, setup_gpu from ._util import app, Arg, Opt, import_code, setup_gpu
from .. import util from .. import util
@ -47,7 +49,7 @@ def find_threshold_cli(
threshold_key (str): Key of threshold attribute in component's configuration. threshold_key (str): Key of threshold attribute in component's configuration.
scores_key (str): Name of score to metric to optimize. scores_key (str): Name of score to metric to optimize.
n_trials (int): Number of trials to determine optimal thresholds n_trials (int): Number of trials to determine optimal thresholds
beta (float): Beta for F1 calculation. beta (float): Beta for F-score calculation.
code_path (Optional[Path]): Path to Python file with additional code (registered functions) to be imported. code_path (Optional[Path]): Path to Python file with additional code (registered functions) to be imported.
use_gpu (int): GPU ID or -1 for CPU. use_gpu (int): GPU ID or -1 for CPU.
gold_preproc (bool): Whether to use gold preprocessing. Gold preprocessing helps the annotations align to the gold_preproc (bool): Whether to use gold preprocessing. Gold preprocessing helps the annotations align to the
@ -79,10 +81,10 @@ def find_threshold(
threshold_key: str, threshold_key: str,
scores_key: str, scores_key: str,
*, *,
n_trials: int = _DEFAULTS["n_trials"], n_trials: int = _DEFAULTS["n_trials"], # type: ignore
beta: float = _DEFAULTS["beta"], beta: float = _DEFAULTS["beta"], # type: ignore
use_gpu: int = _DEFAULTS["use_gpu"], use_gpu: int = _DEFAULTS["use_gpu"], # type: ignore
gold_preproc: bool = _DEFAULTS["gold_preproc"], gold_preproc: bool = _DEFAULTS["gold_preproc"], # type: ignore
silent: bool = True, silent: bool = True,
) -> Tuple[float, float]: ) -> Tuple[float, float]:
""" """
@ -93,7 +95,7 @@ def find_threshold(
threshold_key (str): Key of threshold attribute in component's configuration. threshold_key (str): Key of threshold attribute in component's configuration.
scores_key (str): Name of score to metric to optimize. scores_key (str): Name of score to metric to optimize.
n_trials (int): Number of trials to determine optimal thresholds. n_trials (int): Number of trials to determine optimal thresholds.
beta (float): Beta for F1 calculation. beta (float): Beta for F-score calculation.
use_gpu (int): GPU ID or -1 for CPU. use_gpu (int): GPU ID or -1 for CPU.
gold_preproc (bool): Whether to use gold preprocessing. Gold preprocessing helps the annotations align to the gold_preproc (bool): Whether to use gold preprocessing. Gold preprocessing helps the annotations align to the
tokenization, and may result in sequences of more consistent length. However, it may reduce runtime accuracy due tokenization, and may result in sequences of more consistent length. However, it may reduce runtime accuracy due
@ -108,10 +110,13 @@ def find_threshold(
wasabi.msg.fail("Evaluation data not found", data_path, exits=1) wasabi.msg.fail("Evaluation data not found", data_path, exits=1)
nlp = util.load_model(model) nlp = util.load_model(model)
pipe: Optional[Pipe] = None
try: try:
pipe = nlp.get_pipe(pipe_name) pipe = nlp.get_pipe(pipe_name)
except KeyError as err: except KeyError as err:
wasabi.msg.fail(title=str(err), exits=1) wasabi.msg.fail(title=str(err), exits=1)
if not isinstance(pipe, TrainablePipe):
raise TypeError(Errors.E1044)
if not silent: if not silent:
wasabi.msg.info( wasabi.msg.info(
@ -140,7 +145,9 @@ def find_threshold(
scores: Dict[float, float] = {} scores: Dict[float, float] = {}
for threshold in numpy.linspace(0, 1, n_trials): for threshold in numpy.linspace(0, 1, n_trials):
pipe.cfg = set_nested_item(pipe.cfg, config_keys, threshold) pipe.cfg = set_nested_item(pipe.cfg, config_keys, threshold)
scores[threshold] = nlp.evaluate(dev_dataset)[scores_key] scores[threshold] = nlp.evaluate(dev_dataset, scorer_cfg={"beta": beta})[
scores_key
]
if not ( if not (
isinstance(scores[threshold], float) or isinstance(scores[threshold], int) isinstance(scores[threshold], float) or isinstance(scores[threshold], int)
): ):

1
spacy/errors.py
View File

@ -939,6 +939,7 @@ class Errors(metaclass=ErrorsWithCodes):
"`{arg2}`={arg2_values} but these arguments are conflicting.") "`{arg2}`={arg2_values} but these arguments are conflicting.")
E1043 = ("Expected None or a value in range [{range_start}, {range_end}] for entity linker threshold, but got " E1043 = ("Expected None or a value in range [{range_start}, {range_end}] for entity linker threshold, but got "
"{value}.") "{value}.")
E1044 = ("Only components of type `TrainablePipe` are supported by `find_threshold()`.")
# Deprecated model shortcuts, only used in errors and warnings # Deprecated model shortcuts, only used in errors and warnings

60
spacy/scorer.py
View File

@ -22,15 +22,12 @@ class PRFScore:
"""A precision / recall / F score.""" """A precision / recall / F score."""
def __init__( def __init__(
self, self, *, tp: int = 0, fp: int = 0, fn: int = 0, beta: float = 1
*,
tp: int = 0,
fp: int = 0,
fn: int = 0,
) -> None: ) -> None:
self.tp = tp self.tp = tp
self.fp = fp self.fp = fp
self.fn = fn self.fn = fn
self.beta = beta
def __len__(self) -> int: def __len__(self) -> int:
return self.tp + self.fp + self.fn return self.tp + self.fp + self.fn
@ -42,8 +39,12 @@ class PRFScore:
return self return self
def __add__(self, other): def __add__(self, other):
assert self.beta == other.beta
return PRFScore( return PRFScore(
tp=self.tp + other.tp, fp=self.fp + other.fp, fn=self.fn + other.fn tp=self.tp + other.tp,
fp=self.fp + other.fp,
fn=self.fn + other.fn,
beta=self.beta,
) )
def score_set(self, cand: set, gold: set) -> None: def score_set(self, cand: set, gold: set) -> None:
@ -63,7 +64,7 @@ class PRFScore:
def fscore(self) -> float: def fscore(self) -> float:
p = self.precision p = self.precision
r = self.recall r = self.recall
return 2 * ((p * r) / (p + r + 1e-100)) return (1 + self.beta**2) * ((p * r) / ((self.beta**2 * p) + r + 1e-100))
def to_dict(self) -> Dict[str, float]: def to_dict(self) -> Dict[str, float]:
return {"p": self.precision, "r": self.recall, "f": self.fscore} return {"p": self.precision, "r": self.recall, "f": self.fscore}
@ -101,6 +102,8 @@ class ROCAUCScore:
class Scorer: class Scorer:
"""Compute evaluation scores.""" """Compute evaluation scores."""
BETA = 1
def __init__( def __init__(
self, self,
nlp: Optional["Language"] = None, nlp: Optional["Language"] = None,
@ -149,8 +152,9 @@ class Scorer:
DOCS: https://spacy.io/api/scorer#score_tokenization DOCS: https://spacy.io/api/scorer#score_tokenization
""" """
acc_score = PRFScore() beta = cfg.get("beta", Scorer.BETA)
prf_score = PRFScore() acc_score = PRFScore(beta=beta)
prf_score = PRFScore(beta=beta)
for example in examples: for example in examples:
gold_doc = example.reference gold_doc = example.reference
pred_doc = example.predicted pred_doc = example.predicted
@ -210,7 +214,7 @@ class Scorer:
DOCS: https://spacy.io/api/scorer#score_token_attr DOCS: https://spacy.io/api/scorer#score_token_attr
""" """
tag_score = PRFScore() tag_score = PRFScore(beta=cfg.get("beta", Scorer.BETA))
for example in examples: for example in examples:
gold_doc = example.reference gold_doc = example.reference
pred_doc = example.predicted pred_doc = example.predicted
@ -261,7 +265,8 @@ class Scorer:
key attr_micro_p/r/f and the per-feat PRF scores under key attr_micro_p/r/f and the per-feat PRF scores under
attr_per_feat. attr_per_feat.
""" """
micro_score = PRFScore() beta = cfg.get("beta", Scorer.BETA)
micro_score = PRFScore(beta=beta)
per_feat = {} per_feat = {}
for example in examples: for example in examples:
pred_doc = example.predicted pred_doc = example.predicted
@ -276,7 +281,7 @@ class Scorer:
for feat in morph.split(Morphology.FEATURE_SEP): for feat in morph.split(Morphology.FEATURE_SEP):
field, values = feat.split(Morphology.FIELD_SEP) field, values = feat.split(Morphology.FIELD_SEP)
if field not in per_feat: if field not in per_feat:
per_feat[field] = PRFScore() per_feat[field] = PRFScore(beta=beta)
if field not in gold_per_feat: if field not in gold_per_feat:
gold_per_feat[field] = set() gold_per_feat[field] = set()
gold_per_feat[field].add((gold_i, feat)) gold_per_feat[field].add((gold_i, feat))
@ -298,7 +303,7 @@ class Scorer:
for feat in morph.split(Morphology.FEATURE_SEP): for feat in morph.split(Morphology.FEATURE_SEP):
field, values = feat.split(Morphology.FIELD_SEP) field, values = feat.split(Morphology.FIELD_SEP)
if field not in per_feat: if field not in per_feat:
per_feat[field] = PRFScore() per_feat[field] = PRFScore(beta=beta)
if field not in pred_per_feat: if field not in pred_per_feat:
pred_per_feat[field] = set() pred_per_feat[field] = set()
pred_per_feat[field].add((gold_i, feat)) pred_per_feat[field].add((gold_i, feat))
@ -353,7 +358,8 @@ class Scorer:
DOCS: https://spacy.io/api/scorer#score_spans DOCS: https://spacy.io/api/scorer#score_spans
""" """
score = PRFScore() beta = cfg.get("beta", Scorer.BETA)
score = PRFScore(beta=beta)
score_per_type = dict() score_per_type = dict()
for example in examples: for example in examples:
pred_doc = example.predicted pred_doc = example.predicted
@ -372,7 +378,7 @@ class Scorer:
gold_per_type: Dict[str, Set] = {label: set() for label in labels} gold_per_type: Dict[str, Set] = {label: set() for label in labels}
for label in labels: for label in labels:
if label not in score_per_type: if label not in score_per_type:
score_per_type[label] = PRFScore() score_per_type[label] = PRFScore(beta=beta)
# Find all predidate labels, for all and per type # Find all predidate labels, for all and per type
gold_spans = set() gold_spans = set()
pred_spans = set() pred_spans = set()
@ -469,9 +475,10 @@ class Scorer:
DOCS: https://spacy.io/api/scorer#score_cats DOCS: https://spacy.io/api/scorer#score_cats
""" """
beta = cfg.get("beta", Scorer.BETA)
if threshold is None: if threshold is None:
threshold = 0.5 if multi_label else 0.0 threshold = 0.5 if multi_label else 0.0
f_per_type = {label: PRFScore() for label in labels} f_per_type = {label: PRFScore(beta=beta) for label in labels}
auc_per_type = {label: ROCAUCScore() for label in labels} auc_per_type = {label: ROCAUCScore() for label in labels}
labels = set(labels) labels = set(labels)
if labels: if labels:
@ -519,7 +526,7 @@ class Scorer:
pred_label, pred_score = max(pred_cats.items(), key=lambda it: it[1]) pred_label, pred_score = max(pred_cats.items(), key=lambda it: it[1])
if pred_score >= threshold: if pred_score >= threshold:
f_per_type[pred_label].fp += 1 f_per_type[pred_label].fp += 1
micro_prf = PRFScore() micro_prf = PRFScore(beta=beta)
for label_prf in f_per_type.values(): for label_prf in f_per_type.values():
micro_prf.tp += label_prf.tp micro_prf.tp += label_prf.tp
micro_prf.fn += label_prf.fn micro_prf.fn += label_prf.fn
@ -576,6 +583,7 @@ class Scorer:
DOCS: https://spacy.io/api/scorer#score_links DOCS: https://spacy.io/api/scorer#score_links
""" """
beta = cfg.get("beta", Scorer.BETA)
f_per_type = {} f_per_type = {}
for example in examples: for example in examples:
gold_ent_by_offset = {} gold_ent_by_offset = {}
@ -589,7 +597,7 @@ class Scorer:
if gold_span is not None: if gold_span is not None:
label = gold_span.label_ label = gold_span.label_
if label not in f_per_type: if label not in f_per_type:
f_per_type[label] = PRFScore() f_per_type[label] = PRFScore(beta=beta)
gold = gold_span.kb_id_ gold = gold_span.kb_id_
# only evaluating entities that overlap between gold and pred, # only evaluating entities that overlap between gold and pred,
# to disentangle the performance of the NEL from the NER # to disentangle the performance of the NEL from the NER
@ -608,7 +616,7 @@ class Scorer:
# a wrong prediction (e.g. Q42 != Q3) counts as both a FP as well as a FN # a wrong prediction (e.g. Q42 != Q3) counts as both a FP as well as a FN
f_per_type[label].fp += 1 f_per_type[label].fp += 1
f_per_type[label].fn += 1 f_per_type[label].fn += 1
micro_prf = PRFScore() micro_prf = PRFScore(beta=beta)
for label_prf in f_per_type.values(): for label_prf in f_per_type.values():
micro_prf.tp += label_prf.tp micro_prf.tp += label_prf.tp
micro_prf.fn += label_prf.fn micro_prf.fn += label_prf.fn
@ -663,8 +671,9 @@ class Scorer:
DOCS: https://spacy.io/api/scorer#score_deps DOCS: https://spacy.io/api/scorer#score_deps
""" """
unlabelled = PRFScore() beta = cfg.get("beta", Scorer.BETA)
labelled = PRFScore() unlabelled = PRFScore(beta=beta)
labelled = PRFScore(beta=beta)
labelled_per_dep = dict() labelled_per_dep = dict()
missing_indices = set() missing_indices = set()
for example in examples: for example in examples:
@ -680,7 +689,7 @@ class Scorer:
if dep not in ignore_labels: if dep not in ignore_labels:
gold_deps.add((gold_i, head.i, dep)) gold_deps.add((gold_i, head.i, dep))
if dep not in labelled_per_dep: if dep not in labelled_per_dep:
labelled_per_dep[dep] = PRFScore() labelled_per_dep[dep] = PRFScore(beta=beta)
if dep not in gold_deps_per_dep: if dep not in gold_deps_per_dep:
gold_deps_per_dep[dep] = set() gold_deps_per_dep[dep] = set()
gold_deps_per_dep[dep].add((gold_i, head.i, dep)) gold_deps_per_dep[dep].add((gold_i, head.i, dep))
@ -711,7 +720,7 @@ class Scorer:
else: else:
pred_deps.add((gold_i, gold_head, dep)) pred_deps.add((gold_i, gold_head, dep))
if dep not in labelled_per_dep: if dep not in labelled_per_dep:
labelled_per_dep[dep] = PRFScore() labelled_per_dep[dep] = PRFScore(beta=beta)
if dep not in pred_deps_per_dep: if dep not in pred_deps_per_dep:
pred_deps_per_dep[dep] = set() pred_deps_per_dep[dep] = set()
pred_deps_per_dep[dep].add((gold_i, gold_head, dep)) pred_deps_per_dep[dep].add((gold_i, gold_head, dep))
@ -742,6 +751,7 @@ class Scorer:
def get_ner_prf(examples: Iterable[Example], **kwargs) -> Dict[str, Any]: def get_ner_prf(examples: Iterable[Example], **kwargs) -> Dict[str, Any]:
"""Compute micro-PRF and per-entity PRF scores for a sequence of examples.""" """Compute micro-PRF and per-entity PRF scores for a sequence of examples."""
score_per_type = defaultdict(PRFScore) score_per_type = defaultdict(PRFScore)
beta = kwargs.get("beta", Scorer.BETA)
for eg in examples: for eg in examples:
if not eg.y.has_annotation("ENT_IOB"): if not eg.y.has_annotation("ENT_IOB"):
continue continue
@ -749,7 +759,7 @@ def get_ner_prf(examples: Iterable[Example], **kwargs) -> Dict[str, Any]:
align_x2y = eg.alignment.x2y align_x2y = eg.alignment.x2y
for pred_ent in eg.x.ents: for pred_ent in eg.x.ents:
if pred_ent.label_ not in score_per_type: if pred_ent.label_ not in score_per_type:
score_per_type[pred_ent.label_] = PRFScore() score_per_type[pred_ent.label_] = PRFScore(beta=beta)
indices = align_x2y[pred_ent.start : pred_ent.end] indices = align_x2y[pred_ent.start : pred_ent.end]
if len(indices): if len(indices):
g_span = eg.y[indices[0] : indices[-1] + 1] g_span = eg.y[indices[0] : indices[-1] + 1]
@ -765,7 +775,7 @@ def get_ner_prf(examples: Iterable[Example], **kwargs) -> Dict[str, Any]:
score_per_type[pred_ent.label_].fp += 1 score_per_type[pred_ent.label_].fp += 1
for label, start, end in golds: for label, start, end in golds:
score_per_type[label].fn += 1 score_per_type[label].fn += 1
totals = PRFScore() totals = PRFScore(beta=beta)
for prf in score_per_type.values(): for prf in score_per_type.values():
totals += prf totals += prf
if len(totals) > 0: if len(totals) > 0: