spaCy/spacy/scorer.py

650 lines
23 KiB
Python
Raw Blame History

This file contains ambiguous Unicode characters

This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.

import numpy as np
from .gold import tags_to_entities, GoldParse, DocAnnotation
from .errors import Errors
class PRFScore(object):
"""
A precision / recall / F score
"""
def __init__(self):
self.tp = 0
self.fp = 0
self.fn = 0
def score_set(self, cand, gold):
self.tp += len(cand.intersection(gold))
self.fp += len(cand - gold)
self.fn += len(gold - cand)
@property
def precision(self):
return self.tp / (self.tp + self.fp + 1e-100)
@property
def recall(self):
return self.tp / (self.tp + self.fn + 1e-100)
@property
def fscore(self):
p = self.precision
r = self.recall
return 2 * ((p * r) / (p + r + 1e-100))
class ROCAUCScore(object):
"""
An AUC ROC score.
"""
def __init__(self):
self.golds = []
self.cands = []
self.saved_score = 0.0
self.saved_score_at_len = 0
def score_set(self, cand, gold):
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.
except ValueError:
self.saved_score = -float("inf")
self.saved_score_at_len = len(self.golds)
return self.saved_score
class Scorer(object):
"""Compute evaluation scores."""
def __init__(self, eval_punct=False, pipeline=None):
"""Initialize the Scorer.
eval_punct (bool): Evaluate the dependency attachments to and from
punctuation.
RETURNS (Scorer): The newly created object.
DOCS: https://spacy.io/api/scorer#init
"""
self.tokens = PRFScore()
self.sbd = PRFScore()
self.unlabelled = PRFScore()
self.labelled = PRFScore()
self.labelled_per_dep = dict()
self.tags = PRFScore()
self.sent_starts = PRFScore()
self.ner = PRFScore()
self.ner_per_ents = dict()
self.eval_punct = eval_punct
self.textcat = None
self.textcat_per_cat = dict()
self.textcat_positive_label = None
self.textcat_multilabel = False
if pipeline:
for name, model in pipeline:
if name == "textcat":
self.textcat_positive_label = model.cfg.get("positive_label", None)
if self.textcat_positive_label:
self.textcat = PRFScore()
if not model.cfg.get("exclusive_classes", False):
self.textcat_multilabel = True
for label in model.cfg.get("labels", []):
self.textcat_per_cat[label] = ROCAUCScore()
else:
for label in model.cfg.get("labels", []):
self.textcat_per_cat[label] = PRFScore()
@property
def tags_acc(self):
"""RETURNS (float): Part-of-speech tag accuracy (fine grained tags,
i.e. `Token.tag`).
"""
return self.tags.fscore * 100
@property
def sent_p(self):
"""RETURNS (float): F-score for identification of sentence starts.
i.e. `Token.is_sent_start`).
"""
return self.sent_starts.precision * 100
@property
def sent_r(self):
"""RETURNS (float): F-score for identification of sentence starts.
i.e. `Token.is_sent_start`).
"""
return self.sent_starts.recall * 100
@property
def sent_f(self):
"""RETURNS (float): F-score for identification of sentence starts.
i.e. `Token.is_sent_start`).
"""
return self.sent_starts.fscore * 100
@property
def token_acc(self):
"""RETURNS (float): Tokenization accuracy."""
return self.tokens.precision * 100
@property
def uas(self):
"""RETURNS (float): Unlabelled dependency score."""
return self.unlabelled.fscore * 100
@property
def las(self):
"""RETURNS (float): Labelled dependency score."""
return self.labelled.fscore * 100
@property
def las_per_type(self):
"""RETURNS (dict): Scores per dependency label.
"""
return {
k: {"p": v.precision * 100, "r": v.recall * 100, "f": v.fscore * 100}
for k, v in self.labelled_per_dep.items()
}
@property
def ents_p(self):
"""RETURNS (float): Named entity accuracy (precision)."""
return self.ner.precision * 100
@property
def ents_r(self):
"""RETURNS (float): Named entity accuracy (recall)."""
return self.ner.recall * 100
@property
def ents_f(self):
"""RETURNS (float): Named entity accuracy (F-score)."""
return self.ner.fscore * 100
@property
def ents_per_type(self):
"""RETURNS (dict): Scores per entity label.
"""
return {
k: {"p": v.precision * 100, "r": v.recall * 100, "f": v.fscore * 100}
for k, v in self.ner_per_ents.items()
}
@property
def textcat_score(self):
"""RETURNS (float): f-score on positive label for binary exclusive,
macro-averaged f-score for 3+ exclusive,
macro-averaged AUC ROC score for multilabel (-1 if undefined)
"""
if not self.textcat_multilabel:
# binary multiclass
if self.textcat_positive_label:
return self.textcat.fscore * 100
# other multiclass
return (
sum([score.fscore for label, score in self.textcat_per_cat.items()])
/ (len(self.textcat_per_cat) + 1e-100)
* 100
)
# multilabel
return max(
sum([score.score for label, score in self.textcat_per_cat.items()])
/ (len(self.textcat_per_cat) + 1e-100),
-1,
)
@property
def textcats_per_cat(self):
"""RETURNS (dict): Scores per textcat label.
"""
if not self.textcat_multilabel:
return {
k: {"p": v.precision * 100, "r": v.recall * 100, "f": v.fscore * 100}
for k, v in self.textcat_per_cat.items()
}
return {
k: {"roc_auc_score": max(v.score, -1)}
for k, v in self.textcat_per_cat.items()
}
@property
def scores(self):
"""RETURNS (dict): All scores with keys `uas`, `las`, `ents_p`,
`ents_r`, `ents_f`, `tags_acc`, `token_acc`, and `textcat_score`.
"""
return {
"uas": self.uas,
"las": self.las,
"las_per_type": self.las_per_type,
"ents_p": self.ents_p,
"ents_r": self.ents_r,
"ents_f": self.ents_f,
"ents_per_type": self.ents_per_type,
"tags_acc": self.tags_acc,
"sent_p": self.sent_p,
"sent_r": self.sent_r,
"sent_f": self.sent_f,
"token_acc": self.token_acc,
"textcat_score": self.textcat_score,
"textcats_per_cat": self.textcats_per_cat,
}
def score(self, example, verbose=False, punct_labels=("p", "punct")):
"""Update the evaluation scores from a single Doc / GoldParse pair.
example (Example): The predicted annotations + correct annotations.
verbose (bool): Print debugging information.
punct_labels (tuple): Dependency labels for punctuation. Used to
evaluate dependency attachments to punctuation if `eval_punct` is
`True`.
DOCS: https://spacy.io/api/scorer#score
"""
if isinstance(example, tuple) and len(example) == 2:
doc, gold = example
else:
gold = example.gold
doc = example.doc
if len(doc) != len(gold):
doc_annotation = DocAnnotation(cats=gold.cats)
token_annotation = gold.orig
gold = GoldParse.from_annotation(doc, doc_annotation, token_annotation)
orig = gold.orig
gold_deps = set()
gold_deps_per_dep = {}
gold_tags = set()
gold_sent_starts = set()
gold_ents = set(tags_to_entities(orig.entities))
for id_, tag, head, dep, sent_start in zip(
orig.ids, orig.tags, orig.heads, orig.deps, orig.sent_starts
):
gold_tags.add((id_, tag))
if sent_start:
gold_sent_starts.add(id_)
if dep not in (None, "") and dep.lower() not in punct_labels:
gold_deps.add((id_, head, dep.lower()))
if dep.lower() not in self.labelled_per_dep:
self.labelled_per_dep[dep.lower()] = PRFScore()
if dep.lower() not in gold_deps_per_dep:
gold_deps_per_dep[dep.lower()] = set()
gold_deps_per_dep[dep.lower()].add((id_, head, dep.lower()))
cand_deps = set()
cand_deps_per_dep = {}
cand_tags = set()
cand_sent_starts = set()
for token in doc:
if token.orth_.isspace():
continue
gold_i = gold.cand_to_gold[token.i]
if gold_i is None:
self.tokens.fp += 1
else:
self.tokens.tp += 1
cand_tags.add((gold_i, token.tag_))
if token.is_sent_start:
cand_sent_starts.add(gold_i)
if token.dep_.lower() not in punct_labels and token.orth_.strip():
gold_head = gold.cand_to_gold[token.head.i]
# 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:
self.unlabelled.fp += 1
self.labelled.fp += 1
else:
cand_deps.add((gold_i, gold_head, token.dep_.lower()))
if token.dep_.lower() not in self.labelled_per_dep:
self.labelled_per_dep[token.dep_.lower()] = PRFScore()
if token.dep_.lower() not in cand_deps_per_dep:
cand_deps_per_dep[token.dep_.lower()] = set()
cand_deps_per_dep[token.dep_.lower()].add(
(gold_i, gold_head, token.dep_.lower())
)
if "-" not in [token[-1] for token in orig.entities]:
# Find all NER labels in gold and doc
ent_labels = set([x[0] for x in gold_ents] + [k.label_ for k in doc.ents])
# Set up all labels for per type scoring and prepare gold per type
gold_per_ents = {ent_label: set() for ent_label in ent_labels}
for ent_label in ent_labels:
if ent_label not in self.ner_per_ents:
self.ner_per_ents[ent_label] = PRFScore()
gold_per_ents[ent_label].update(
[x for x in gold_ents if x[0] == ent_label]
)
# Find all candidate labels, for all and per type
cand_ents = set()
cand_per_ents = {ent_label: set() for ent_label in ent_labels}
for ent in doc.ents:
first = gold.cand_to_gold[ent.start]
last = gold.cand_to_gold[ent.end - 1]
if first is None or last is None:
self.ner.fp += 1
self.ner_per_ents[ent.label_].fp += 1
else:
cand_ents.add((ent.label_, first, last))
cand_per_ents[ent.label_].add((ent.label_, first, last))
# Scores per ent
for k, v in self.ner_per_ents.items():
if k in cand_per_ents:
v.score_set(cand_per_ents[k], gold_per_ents[k])
# Score for all ents
self.ner.score_set(cand_ents, gold_ents)
self.tags.score_set(cand_tags, gold_tags)
self.sent_starts.score_set(cand_sent_starts, gold_sent_starts)
self.labelled.score_set(cand_deps, gold_deps)
for dep in self.labelled_per_dep:
self.labelled_per_dep[dep].score_set(
cand_deps_per_dep.get(dep, set()), gold_deps_per_dep.get(dep, set())
)
self.unlabelled.score_set(
set(item[:2] for item in cand_deps), set(item[:2] for item in gold_deps)
)
if (
len(gold.cats) > 0
and set(self.textcat_per_cat) == set(gold.cats)
and set(gold.cats) == set(doc.cats)
):
goldcat = max(gold.cats, key=gold.cats.get)
candcat = max(doc.cats, key=doc.cats.get)
if self.textcat_positive_label:
self.textcat.score_set(
set([self.textcat_positive_label]) & set([candcat]),
set([self.textcat_positive_label]) & set([goldcat]),
)
for label in self.textcat_per_cat:
if self.textcat_multilabel:
self.textcat_per_cat[label].score_set(
doc.cats[label], gold.cats[label]
)
else:
self.textcat_per_cat[label].score_set(
set([label]) & set([candcat]), set([label]) & set([goldcat])
)
elif len(self.textcat_per_cat) > 0:
model_labels = set(self.textcat_per_cat)
eval_labels = set(gold.cats)
raise ValueError(
Errors.E162.format(model_labels=model_labels, eval_labels=eval_labels)
)
if verbose:
gold_words = orig.words
for w_id, h_id, dep in cand_deps - gold_deps:
print("F", gold_words[w_id], dep, gold_words[h_id])
for w_id, h_id, dep in gold_deps - cand_deps:
print("M", gold_words[w_id], dep, gold_words[h_id])
#############################################################################
#
# 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
# permission.
#
#
# 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.
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.
"""
pos_label = 1.0
y_true = np.ravel(y_true)
y_score = np.ravel(y_score)
# make y_true a boolean vector
y_true = y_true == pos_label
# 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]
weight = 1.0
# 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)
if not np.all(
np.isclose(
out.take(-1, axis=axis), expected, rtol=rtol, atol=atol, equal_nan=True
)
):
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