def test_can_init_scaled_dot(self):
legacy_attention = Attention.from_params(
Params({
"type": "scaled_dot_product",
"scaling_factor": 9
}))
isinstance(legacy_attention, ScaledDotProductAttention)
def test_can_init_linear(self):
legacy_attention = Attention.from_params(
Params({
"type": "linear",
"tensor_1_dim": 3,
"tensor_2_dim": 3
}))
isinstance(legacy_attention, LinearAttention)
def __init__(
self,
vocab: Vocabulary,
text_field_embedder: TextFieldEmbedder,
elmo_text_field_embedder: TextFieldEmbedder,
quote_response_encoder: Seq2SeqEncoder,
quote_response_encoder_aux: Seq2VecEncoder,
classifier_feedforward: FeedForward,
classifier_feedforward_2: FeedForward,
initializer: InitializerApplicator = InitializerApplicator(),
regularizer: Optional[RegularizerApplicator] = None,
report_auxiliary_metrics: bool = False,
# predict_mode: bool = False,
) -> None:
super(SarcasmClassifier, self).__init__(vocab, regularizer)
self.text_field_embedder = text_field_embedder
self.elmo_text_field_embedder = elmo_text_field_embedder
self.num_classes = self.vocab.get_vocab_size("labels")
self.num_classes_emotions = self.vocab.get_vocab_size("emotion_labels")
self.quote_response_encoder = quote_response_encoder
self.quote_response_encoder_aux = quote_response_encoder_aux
self.classifier_feedforward = classifier_feedforward
self.classifier_feedforward_2 = classifier_feedforward_2
self.attention_seq2seq = Attention(
quote_response_encoder.get_output_dim())
self.label_acc_metrics = {"accuracy": CategoricalAccuracy()}
self.label_f1_metrics = {}
self.label_f1_metrics_emotions = {}
# for i in range(self.num_classes):
# self.label_f1_metrics[vocab.get_token_from_index(index=i, namespace="label")] =\
# F1Measure(positive_label=i)
for i in range(self.num_classes):
self.label_f1_metrics[vocab.get_token_from_index(index=i, namespace="labels")] =\
F1Measure(positive_label=i)
for i in range(self.num_classes_emotions):
self.label_f1_metrics_emotions[vocab.get_token_from_index(index=i, namespace="emotion_labels")] =\
F1Measure(positive_label=i)
self.loss = torch.nn.CrossEntropyLoss()
# self.attention_seq2seq = Attention(quote_response_encoder.get_output_dim())
self.report_auxiliary_metrics = report_auxiliary_metrics
# self.predict_mode = predict_mode
initializer(self)
from __future__ import absolute_import
import torch
#overrides
from allennlp.modules.attention.attention import Attention
from allennlp.modules.similarity_functions import DotProductSimilarity, SimilarityFunction
class LegacyAttention(Attention):
u"""
Computes attention between a vector and a matrix using a similarity function.
This should be considered deprecated, as it consumes more memory than the specialized attention modules.
"""
def __init__(self, similarity_function=None, normalize=True):
super(LegacyAttention, self).__init__(normalize)
self._similarity_function = similarity_function or DotProductSimilarity(
)
#overrides
def _forward_internal(self, vector, matrix):
tiled_vector = vector.unsqueeze(1).expand(vector.size()[0],
matrix.size()[1],
vector.size()[1])
return self._similarity_function(tiled_vector, matrix)
LegacyAttention = Attention.register(u"legacy")(LegacyAttention)
activation : ``Activation``, optional (default=linear (i.e. no activation))
An activation function applied after the ``x^T W y + b`` calculation. Default is no
activation.
normalize : ``bool``, optional (default: ``True``)
If true, we normalize the computed similarities with a softmax, to return a probability
distribution for your attention. If false, this is just computing a similarity score.
"""
def __init__(self,
vector_dim,
matrix_dim,
activation=None,
normalize=True):
super(BilinearAttention, self).__init__(normalize)
self._weight_matrix = Parameter(torch.Tensor(vector_dim, matrix_dim))
self._bias = Parameter(torch.Tensor(1))
self._activation = activation or Activation.by_name(u'linear')()
self.reset_parameters()
def reset_parameters(self):
torch.nn.init.xavier_uniform_(self._weight_matrix)
self._bias.data.fill_(0)
#overrides
def _forward_internal(self, vector, matrix):
intermediate = vector.mm(self._weight_matrix).unsqueeze(1)
return self._activation(
intermediate.bmm(matrix.transpose(1, 2)).squeeze(1) + self._bias)
BilinearAttention = Attention.register(u"bilinear")(BilinearAttention)
def test_can_init_cosine(self):
legacy_attention = Attention.from_params(Params({"type": "cosine"}))
isinstance(legacy_attention, CosineAttention)
def test_can_init_legacy(self):
legacy_attention = Attention.from_params(Params({"type": "legacy"}))
isinstance(legacy_attention, LegacyAttention)
def test_can_init_dot(self):
legacy_attention = Attention.from_params(
Params({"type": "dot_product"}))
isinstance(legacy_attention, DotProductAttention)
def test_can_init_dot(self):
legacy_attention = Attention.from_params(Params({"type": "dot_product"}))
isinstance(legacy_attention, DotProductAttention)
def test_can_init_cosine(self):
legacy_attention = Attention.from_params(Params({"type": "cosine"}))
isinstance(legacy_attention, CosineAttention)
def test_can_init_linear(self):
legacy_attention = Attention.from_params(Params({"type": "linear",
"tensor_1_dim": 3,
"tensor_2_dim": 3}))
isinstance(legacy_attention, LinearAttention)
