explosion/spaCy
1
1
Fork 0
mirror of https://github.com/explosion/spaCy.git synced 2025-04-24 19:11:58 +03:00
Code Issues Packages Projects Releases Wiki Activity

[ci skip] Small updates

Browse Source
This commit is contained in:
Lj Miranda 2022-08-25 16:26:03 +08:00
parent b728eaae18
commit 43bf05275f
1 changed files with 4 additions and 6 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

10
spacy/pipeline/spancat_exclusive.py
View File

@ -18,9 +18,9 @@ from .trainable_pipe import TrainablePipe
@registry.layers("spacy.Softmax.v1")
def build_linear_logistic(nO=None, nI=None) -> Model[Floats2d, Floats2d]:
"""An output layer for multi-label classification. It uses a linear layer
followed by a logistic activation.
def build_softmax(nO=None, nI=None) -> Model[Floats2d, Floats2d]:
"""
An output layer for softmax classification.
"""
return Softmax_v2(nI=nI, nO=nO)
@ -376,12 +376,10 @@ class SpanCategorizerExclusive(TrainablePipe):
offset += spans.lengths[i]
target = self.model.ops.asarray(target, dtype="f") # type: ignore
negative_samples = numpy.nonzero(negative_spans)[0]
breakpoint()
target[negative_samples, self._negative_label] = 1.0
d_scores = scores - target
neg_weight = self.cfg["negative_weight"]
if neg_weight != 1.0:
d_scores[negative_samples] *= neg_weight
d_scores[negative_samples] *= neg_weight
loss = float((d_scores**2).sum())
return loss, d_scores