def __init__(
self,
is_binary: bool = True,
label_weights: Optional[Dict[str, float]] = None,
loss=None,
):
super().__init__()
if is_binary:
self.loss = loss or BinaryCrossEntropyLoss(BinaryCrossEntropyLoss.Config())
else:
self.loss = loss or CrossEntropyLoss(CrossEntropyLoss.Config())
class Config(OutputLayerBase.Config):
loss: Union[CrossEntropyLoss.Config, BinaryCrossEntropyLoss.Config,
AUCPRHingeLoss.Config, KLDivergenceBCELoss.Config,
KLDivergenceCELoss.Config, LabelSmoothedCrossEntropyLoss.
Config, ] = CrossEntropyLoss.Config()
label_weights: Dict[str, float] = {}
ignore_pad_in_loss: Optional[bool] = True
def test_doc_classification_output_layer(self):
tensorizer = LabelTensorizer()
tensorizer.vocab = Vocabulary([SpecialTokens.PAD, "foo", "bar"])
layer = ClassificationOutputLayer.from_config(
config=ClassificationOutputLayer.Config(loss=CrossEntropyLoss.Config()),
labels=tensorizer.vocab,
)
self.assertEqual(layer.loss_fn.ignore_index, 0)
# use default pad
tensorizer.vocab = Vocabulary(["foo", "bar"])
layer = ClassificationOutputLayer.from_config(
config=ClassificationOutputLayer.Config(loss=CrossEntropyLoss.Config()),
labels=tensorizer.vocab,
)
self.assertEqual(layer.loss_fn.ignore_index, -1)
class Config(OutputLayerBase.Config):
loss: Union[CrossEntropyLoss.Config, BinaryCrossEntropyLoss.Config,
MultiLabelSoftMarginLoss.Config, AUCPRHingeLoss.Config,
KLDivergenceBCELoss.Config, KLDivergenceCELoss.Config,
LabelSmoothedCrossEntropyLoss.
Config, ] = CrossEntropyLoss.Config()
label_weights: Optional[Dict[str, float]] = None
class Config(OutputLayerBase.Config):
loss: CrossEntropyLoss.Config = CrossEntropyLoss.Config()
ignore_impossible: bool = True
pos_loss_weight: float = 0.5
has_answer_loss_weight: float = 0.5
false_label: str = "False"
max_answer_len: int = 30
class Config(OutputLayerBase.Config):
loss: Union[
CrossEntropyLoss.Config,
BinaryCrossEntropyLoss.Config,
AUCPRHingeLoss.Config,
KLDivergenceBCELoss.Config,
KLDivergenceCELoss.Config,
SoftHardBCELoss.Config,
] = CrossEntropyLoss.Config()
class Config(OutputLayerBase.Config):
loss: Union[CrossEntropyLoss.Config, BinaryCrossEntropyLoss.Config,
BinaryCrossEntropyWithLogitsLoss.Config,
MultiLabelSoftMarginLoss.Config, AUCPRHingeLoss.Config,
HingeLoss.Config, KLDivergenceBCELoss.Config,
KLDivergenceCELoss.Config, LabelSmoothedCrossEntropyLoss.
Config, ] = CrossEntropyLoss.Config()
label_weights: Optional[Dict[str, float]] = None
automatic_label_weighting_method: Optional[WeightingMethod] = None
class Config(OutputLayerBase.Config):
loss: Union[
CrossEntropyLoss.Config,
BinaryCrossEntropyLoss.Config,
AUCPRHingeLoss.Config,
KLDivergenceBCELoss.Config,
KLDivergenceCELoss.Config,
SoftHardBCELoss.Config,
] = CrossEntropyLoss.Config()
label_weights: Optional[Dict[str, float]] = None
class Config(OutputLayerBase.Config):
loss: Union[CrossEntropyLoss.Config,
KLDivergenceCELoss.Config] = CrossEntropyLoss.Config()
ignore_impossible: bool = True
pos_loss_weight: float = 0.5
has_answer_loss_weight: float = 0.5
false_label: str = "False"
max_answer_len: int = 30
# For knowledge distillation we have soft and hard labels. This specifies
# the weight on loss against hard labels.
hard_weight: float = 0.0
def from_config(cls, config, tensorizers):
embedding = create_module(config.embedding,
tensorizer=tensorizers["tokens"])
representation = create_module(config.representation,
embed_dim=embedding.embedding_dim)
slots = tensorizers["slots"].vocab
decoder = create_module(config.decoder,
in_dim=representation.representation_dim,
out_dim=len(slots))
output_layer = MyTaggingOutputLayer(slots, CrossEntropyLoss(None))
return cls(embedding, representation, decoder, output_layer)
def from_config(cls, config: Config, tensorizers: Dict[str, Tensorizer]):
labels = tensorizers["labels"].labels
embedding = cls.create_embedding(config, tensorizers)
representation = create_module(config.representation,
embed_dim=embedding.embedding_dim)
decoder = create_module(
config.decoder,
in_dim=representation.representation_dim,
out_dim=len(labels),
)
output_layer = ClassificationOutputLayer(labels,
CrossEntropyLoss(None))
return cls(embedding, representation, decoder, output_layer)
def from_config(cls, config, tensorizers):
labels = tensorizers["labels"].vocab
embedding = cls.create_embedding(config, tensorizers)
representation = create_module(config.representation,
embed_dim=embedding.embedding_dim)
decoder = create_module(
config.decoder,
in_dim=representation.representation_dim,
out_dim=len(labels),
)
# TODO after migration: create_module(config.output_layer, tensorizers=tensorizers)
output_layer = WordTaggingOutputLayer(labels, CrossEntropyLoss(None))
return cls(embedding, representation, decoder, output_layer)
def from_config(cls, config: Config, tensorizers: Dict[str, Tensorizer]):
vocab = tensorizers["tokens"].vocab
labels = tensorizers["labels"].labels
embedding = WordEmbedding(len(vocab), config.embedding.embed_dim, None,
None, vocab.idx[UNK], [])
representation = create_module(config.representation,
embed_dim=embedding.embedding_dim)
decoder = create_module(
config.decoder,
in_dim=representation.representation_dim,
out_dim=len(labels),
)
output_layer = ClassificationOutputLayer(labels,
CrossEntropyLoss(None))
return cls(embedding, representation, decoder, output_layer)
def __init__(self, loss: Loss = None):
super().__init__()
self.loss = loss or CrossEntropyLoss(CrossEntropyLoss.Config())
class Config(OutputLayerBase.Config):
loss: CrossEntropyLoss.Config = CrossEntropyLoss.Config()
class Config(OutputLayerBase.Config):
loss: Union[CrossEntropyLoss.Config, BinaryCrossEntropyLoss.Config,
AUCPRHingeLoss.Config, ] = CrossEntropyLoss.Config()
class Config(OutputLayerBase.Config):
loss: CrossEntropyLoss.Config = CrossEntropyLoss.Config()
label_weights: Dict[str, float] = {}
class Config(ConfigBase):
loss: Union[CrossEntropyLoss.Config,
LabelSmoothedCrossEntropyLoss.Config,
NLLLoss.Config, ] = CrossEntropyLoss.Config()
