Python nn.RegularizerApplicator方法代码示例

# 需要导入模块: from allennlp import nn [as 别名]
# 或者: from allennlp.nn import RegularizerApplicator [as 别名]
def __init__(self, vocab: Vocabulary,
                 text_field_embedder: TextFieldEmbedder,
                 verbose_metrics: bool = False,
                 dropout: float = 0.2,
                 initializer: InitializerApplicator = InitializerApplicator(),
                 regularizer: Optional[RegularizerApplicator] = None,
                 ) -> None:
        super(TextClassifier, self).__init__(vocab, regularizer)

        self.text_field_embedder = text_field_embedder
        self.dropout = torch.nn.Dropout(dropout)
        self.num_classes = self.vocab.get_vocab_size("labels")
        self.classifier_feedforward = torch.nn.Linear(self.text_field_embedder.get_output_dim()  , self.num_classes)

        self.label_accuracy = CategoricalAccuracy()
        self.label_f1_metrics = {}

        self.verbose_metrics = verbose_metrics

        for i in range(self.num_classes):
            self.label_f1_metrics[vocab.get_token_from_index(index=i, namespace="labels")] = F1Measure(positive_label=i)
        self.loss = torch.nn.CrossEntropyLoss()

        initializer(self) 

# 需要导入模块: from allennlp import nn [as 别名]
# 或者: from allennlp.nn import RegularizerApplicator [as 别名]
def __init__(self, vocab: Vocabulary,
                 input_dim: int,
                 num_classes: int,
                 label_namespace: str = "labels",
                 feedforward: Optional[FeedForward] = None,
                 dropout: Optional[float] = None,
                 verbose_metrics: bool = False,
                 initializer: InitializerApplicator = InitializerApplicator(),
                 regularizer: Optional[RegularizerApplicator] = None) -> None:
        super().__init__(vocab, regularizer)
        self.label_namespace = label_namespace
        self.input_dim = input_dim
        self.num_classes = num_classes 
        self._verbose_metrics = verbose_metrics
        if dropout:
            self.dropout = torch.nn.Dropout(dropout)
        else:
            self.dropout = None
        self._feedforward = feedforward

        if self._feedforward is not None: 
            self.projection_layer = Linear(feedforward.get_output_dim(), self.num_classes)
        else:
            self.projection_layer = Linear(self.input_dim, self.num_classes)

        self.metrics = {
                "accuracy": CategoricalAccuracy(),
                "accuracy3": CategoricalAccuracy(top_k=3),
                "accuracy5": CategoricalAccuracy(top_k=5)
        }
        self._loss = torch.nn.CrossEntropyLoss()

        initializer(self) 

# 需要导入模块: from allennlp import nn [as 别名]
# 或者: from allennlp.nn import RegularizerApplicator [as 别名]
def __init__(self, vocab: Vocabulary,
                 text_field_embedder: TextFieldEmbedder,
                 text_encoder: Seq2SeqEncoder,
                 classifier_feedforward: FeedForward,
                 verbose_metrics: False,
                 initializer: InitializerApplicator = InitializerApplicator(),
                 regularizer: Optional[RegularizerApplicator] = None,
                 ) -> None:
        super(TextClassifier, self).__init__(vocab, regularizer)

        self.text_field_embedder = text_field_embedder
        self.num_classes = self.vocab.get_vocab_size("labels")
        self.text_encoder = text_encoder
        self.classifier_feedforward = classifier_feedforward
        self.prediction_layer = torch.nn.Linear(self.classifier_feedforward.get_output_dim()  , self.num_classes)

        self.label_accuracy = CategoricalAccuracy()
        self.label_f1_metrics = {}

        self.verbose_metrics = verbose_metrics

        for i in range(self.num_classes):
            self.label_f1_metrics[vocab.get_token_from_index(index=i, namespace="labels")] = F1Measure(positive_label=i)
        self.loss = torch.nn.CrossEntropyLoss()

        self.pool = lambda text, mask: util.get_final_encoder_states(text, mask, bidirectional=True)

        initializer(self) 

# 需要导入模块: from allennlp import nn [as 别名]
# 或者: from allennlp.nn import RegularizerApplicator [as 别名]
def __init__(self, vocab: Vocabulary,
                 text_field_embedder: TextFieldEmbedder,
                 encoder: Seq2SeqEncoder,
                 include_start_end_transitions: bool = True,
                 dropout: Optional[float] = None,
                 initializer: InitializerApplicator = InitializerApplicator(),
                 regularizer: Optional[RegularizerApplicator] = None) -> None:
        super().__init__(vocab, regularizer)

        self.label_namespace = 'labels'
        self.num_tags = self.vocab.get_vocab_size(self.label_namespace)

        # encode text
        self.text_field_embedder = text_field_embedder
        self.encoder = encoder
        self.dropout = torch.nn.Dropout(dropout) if dropout else None

        # crf
        output_dim = self.encoder.get_output_dim()
        self.tag_projection_layer = TimeDistributed(Linear(output_dim, self.num_tags))
        self.crf = ConditionalRandomField(self.num_tags, constraints=None, include_start_end_transitions=include_start_end_transitions)

        self.metrics = {
            "accuracy": CategoricalAccuracy(),
            "accuracy3": CategoricalAccuracy(top_k=3)
        }
        for index, label in self.vocab.get_index_to_token_vocabulary(self.label_namespace).items():
            self.metrics['F1_' + label] = F1Measure(positive_label=index)

        initializer(self) 

# 需要导入模块: from allennlp import nn [as 别名]
# 或者: from allennlp.nn import RegularizerApplicator [as 别名]
def __init__(self, vocab: Vocabulary,
                 text_field_embedder: TextFieldEmbedder,
                 predictor_dropout=0.0,
                 labels_namespace: str = "labels",
                 detect_namespace: str = "d_tags",
                 verbose_metrics: bool = False,
                 label_smoothing: float = 0.0,
                 confidence: float = 0.0,
                 initializer: InitializerApplicator = InitializerApplicator(),
                 regularizer: Optional[RegularizerApplicator] = None) -> None:
        super(Seq2Labels, self).__init__(vocab, regularizer)

        self.label_namespaces = [labels_namespace,
                                 detect_namespace]
        self.text_field_embedder = text_field_embedder
        self.num_labels_classes = self.vocab.get_vocab_size(labels_namespace)
        self.num_detect_classes = self.vocab.get_vocab_size(detect_namespace)
        self.label_smoothing = label_smoothing
        self.confidence = confidence
        self.incorr_index = self.vocab.get_token_index("INCORRECT",
                                                       namespace=detect_namespace)

        self._verbose_metrics = verbose_metrics
        self.predictor_dropout = TimeDistributed(torch.nn.Dropout(predictor_dropout))

        self.tag_labels_projection_layer = TimeDistributed(
            Linear(text_field_embedder._token_embedders['bert'].get_output_dim(), self.num_labels_classes))

        self.tag_detect_projection_layer = TimeDistributed(
            Linear(text_field_embedder._token_embedders['bert'].get_output_dim(), self.num_detect_classes))

        self.metrics = {"accuracy": CategoricalAccuracy()}

        initializer(self) 

# 需要导入模块: from allennlp import nn [as 别名]
# 或者: from allennlp.nn import RegularizerApplicator [as 别名]
def __init__(self, vocab: Vocabulary,
                 text_field_embedder: TextFieldEmbedder,
                 encoder: Seq2SeqEncoder,
                 projection_feedforward: FeedForward,
                 inference_encoder: Seq2SeqEncoder,
                 output_feedforward: FeedForward,
                 output_logit: FeedForward,
                 final_feedforward: FeedForward,
                 coverage_loss: CoverageLoss,
                 similarity_function: SimilarityFunction = DotProductSimilarity(),
                 dropout: float = 0.5,
                 contextualize_pair_comparators: bool = False,
                 pair_context_encoder: Seq2SeqEncoder = None,
                 pair_feedforward: FeedForward = None,
                 initializer: InitializerApplicator = InitializerApplicator(),
                 regularizer: Optional[RegularizerApplicator] = None) -> None:
    	# Need to send it verbatim because otherwise FromParams doesn't work appropriately.
        super().__init__(vocab=vocab,
                         text_field_embedder=text_field_embedder,
                         encoder=encoder,
                         similarity_function=similarity_function,
                         projection_feedforward=projection_feedforward,
                         inference_encoder=inference_encoder,
                         output_feedforward=output_feedforward,
                         output_logit=output_logit,
                         final_feedforward=final_feedforward,
                         contextualize_pair_comparators=contextualize_pair_comparators,
                         coverage_loss=coverage_loss,
                         pair_context_encoder=pair_context_encoder,
                         pair_feedforward=pair_feedforward,
                         dropout=dropout,
                         initializer=initializer,
                         regularizer=regularizer)
        self._answer_loss = torch.nn.CrossEntropyLoss()

        self._accuracy = CategoricalAccuracy() 

# 需要导入模块: from allennlp import nn [as 别名]
# 或者: from allennlp.nn import RegularizerApplicator [as 别名]
def __init__(self, vocab: Vocabulary,
                 text_field_embedder: TextFieldEmbedder,
                 encoder: Seq2SeqEncoder,
                 projection_feedforward: FeedForward,
                 inference_encoder: Seq2SeqEncoder,
                 output_feedforward: FeedForward,
                 output_logit: FeedForward,
                 final_feedforward: FeedForward,
                 coverage_loss: CoverageLoss,
                 similarity_function: SimilarityFunction = DotProductSimilarity(),
                 dropout: float = 0.5,
                 contextualize_pair_comparators: bool = False,
                 pair_context_encoder: Seq2SeqEncoder = None,
                 pair_feedforward: FeedForward = None,
                 initializer: InitializerApplicator = InitializerApplicator(),
                 regularizer: Optional[RegularizerApplicator] = None) -> None:
        super().__init__(vocab=vocab,
                         text_field_embedder=text_field_embedder,
                         encoder=encoder,
                         similarity_function=similarity_function,
                         projection_feedforward=projection_feedforward,
                         inference_encoder=inference_encoder,
                         output_feedforward=output_feedforward,
                         output_logit=output_logit,
                         final_feedforward=final_feedforward,
                         coverage_loss=coverage_loss,
                         contextualize_pair_comparators=contextualize_pair_comparators,
                         pair_context_encoder=pair_context_encoder,
                         pair_feedforward=pair_feedforward,
                         dropout=dropout,
                         initializer=initializer,
                         regularizer=regularizer)
        self._ignore_index = -1
        self._answer_loss = torch.nn.CrossEntropyLoss(ignore_index=self._ignore_index)
        self._coverage_loss = coverage_loss

        self._accuracy = CategoricalAccuracy()
        self._entailment_f1 = F1Measure(self._label2idx["entailment"]) 

# 需要导入模块: from allennlp import nn [as 别名]
# 或者: from allennlp.nn import RegularizerApplicator [as 别名]
def __init__(self, vocab: Vocabulary,
                 text_field_embedder: TextFieldEmbedder,
                 phrase_layer: Seq2SeqEncoder,
                 projected_layer: Seq2SeqEncoder,
                 flow_layer: Seq2SeqEncoder,
                 contextual_passage: Seq2SeqEncoder,
                 contextual_question: Seq2SeqEncoder,
                 dropout: float = 0.2,
                 regularizer: Optional[RegularizerApplicator] = None,
                 initializer: InitializerApplicator = InitializerApplicator(),
                 ):

        super(MultiGranularityHierarchicalAttentionFusionNetworks, self).__init__(vocab, regularizer)
        self._text_field_embedder = text_field_embedder
        self._phrase_layer = phrase_layer
        self._encoding_dim = self._phrase_layer.get_output_dim()
        self.projected_layer = torch.nn.Linear(self._encoding_dim + 1024, self._encoding_dim)
        self.fuse = FusionLayer(self._encoding_dim)
        self.projected_lstm = projected_layer
        self.flow = flow_layer
        self.contextual_layer_p = contextual_passage
        self.contextual_layer_q = contextual_question
        self.linear_self_align = torch.nn.Linear(self._encoding_dim, 1)
        self.bilinear_layer_s = BilinearSeqAtt(self._encoding_dim, self._encoding_dim)
        self.bilinear_layer_e = BilinearSeqAtt(self._encoding_dim, self._encoding_dim)
        self.yesno_predictor = torch.nn.Linear(self._encoding_dim, 3)
        self.relu = torch.nn.ReLU()

        self._max_span_length = 30

        self._span_start_accuracy = CategoricalAccuracy()
        self._span_end_accuracy = CategoricalAccuracy()
        self._span_accuracy = BooleanAccuracy()
        self._squad_metrics = SquadEmAndF1()
        self._span_yesno_accuracy = CategoricalAccuracy()
        self._official_f1 = Average()
        self._variational_dropout = InputVariationalDropout(dropout)

        self._loss = torch.nn.CrossEntropyLoss()
        initializer(self) 

# 需要导入模块: from allennlp import nn [as 别名]
# 或者: from allennlp.nn import RegularizerApplicator [as 别名]
def __init__(self,
                 vocab: Vocabulary,
                 source_text_embedder: TextFieldEmbedder,
                 encoder: Seq2SeqEncoder,
                 decoder: SeqDecoder,
                 tied_source_embedder_key: Optional[str] = None,
                 initializer: InitializerApplicator = InitializerApplicator(),
                 regularizer: Optional[RegularizerApplicator] = None) -> None:

        super(CustomComposedSeq2Seq, self).__init__(vocab, regularizer)

        self._source_text_embedder = source_text_embedder
        self._encoder = encoder
        self._decoder = decoder

        if self._encoder.get_output_dim() != self._decoder.get_output_dim():
            raise ConfigurationError(f"Encoder output dimension {self._encoder.get_output_dim()} should be"
                                     f" equal to decoder dimension {self._decoder.get_output_dim()}.")
        if tied_source_embedder_key:
            if not isinstance(self._source_text_embedder, BasicTextFieldEmbedder):
                raise ConfigurationError("Unable to tie embeddings,"
                                         "Source text embedder is not an instance of `BasicTextFieldEmbedder`.")
            source_embedder = self._source_text_embedder._token_embedders[tied_source_embedder_key]
            if not isinstance(source_embedder, Embedding):
                raise ConfigurationError("Unable to tie embeddings,"
                                         "Selected source embedder is not an instance of `Embedding`.")
            if source_embedder.get_output_dim() != self._decoder.target_embedder.get_output_dim():
                raise ConfigurationError(f"Output Dimensions mismatch between"
                                         f"source embedder and target embedder.")
            self._source_text_embedder._token_embedders[tied_source_embedder_key] = self._decoder.target_embedder
        initializer(self) 

# 需要导入模块: from allennlp import nn [as 别名]
# 或者: from allennlp.nn import RegularizerApplicator [as 别名]
def __init__(self,
                 vocab: Vocabulary,
                 text_field_embedder: TextFieldEmbedder,
                 question_encoder: Seq2SeqEncoder,
                 passage_encoder: Seq2SeqEncoder,
                 pair_encoder: AttentionEncoder,
                 self_encoder: AttentionEncoder,
                 output_layer: QAOutputLayer,
                 initializer: InitializerApplicator = InitializerApplicator(),
                 regularizer: Optional[RegularizerApplicator] = None,
                 share_encoder: bool = False):

        super().__init__(vocab, regularizer)
        self.text_field_embedder = text_field_embedder
        self.question_encoder = question_encoder
        self.passage_encoder = passage_encoder
        self.pair_encoder = pair_encoder
        self.self_encoder = self_encoder
        self.output_layer = output_layer

        self._span_start_accuracy = CategoricalAccuracy()
        self._span_end_accuracy = CategoricalAccuracy()
        self._span_accuracy = BooleanAccuracy()
        self._squad_metrics = SquadEmAndF1()
        self.share_encoder = share_encoder
        self.loss = torch.nn.CrossEntropyLoss()
        initializer(self) 

# 需要导入模块: from allennlp import nn [as 别名]
# 或者: from allennlp.nn import RegularizerApplicator [as 别名]
def __init__(self, vocab: Vocabulary,
                 text_field_embedder: TextFieldEmbedder,
                 encoder: Seq2SeqEncoder,
                 calculate_span_f1: bool = None,
                 label_encoding: Optional[str] = None,
                 label_namespace: str = "labels",
                 verbose_metrics: bool = False,
                 initializer: InitializerApplicator = InitializerApplicator(),
                 regularizer: Optional[RegularizerApplicator] = None) -> None:
        super(SimpleTagger, self).__init__(vocab, regularizer)

        self.label_namespace = label_namespace
        self.text_field_embedder = text_field_embedder
        self.num_classes = self.vocab.get_vocab_size(label_namespace)
        self.encoder = encoder
        self._verbose_metrics = verbose_metrics
        self.tag_projection_layer = TimeDistributed(Linear(self.encoder.get_output_dim(),
                                                           self.num_classes))

        check_dimensions_match(text_field_embedder.get_output_dim(), encoder.get_input_dim(),
                               "text field embedding dim", "encoder input dim")

        # We keep calculate_span_f1 as a constructor argument for API consistency with
        # the CrfTagger, even it is redundant in this class
        # (label_encoding serves the same purpose).
        if calculate_span_f1 and not label_encoding:
            raise ConfigurationError("calculate_span_f1 is True, but "
                                     "no label_encoding was specified.")
        self.metrics = {
            "accuracy": CategoricalAccuracy(),
            "accuracy3": CategoricalAccuracy(top_k=3)
        }

        if calculate_span_f1 or label_encoding:
            self._f1_metric = SpanBasedF1Measure(vocab,
                                                 tag_namespace=label_namespace,
                                                 label_encoding=label_encoding)
        else:
            self._f1_metric = None

        initializer(self) 

# 需要导入模块: from allennlp import nn [as 别名]
# 或者: from allennlp.nn import RegularizerApplicator [as 别名]
def __init__(self,
                 vocab: Vocabulary,
                 task: str,
                 encoder: Seq2SeqEncoder,
                 label_smoothing: float = 0.0,
                 dropout: float = 0.0,
                 adaptive: bool = False,
                 features: List[str] = None,
                 initializer: InitializerApplicator = InitializerApplicator(),
                 regularizer: Optional[RegularizerApplicator] = None) -> None:
        super(TagDecoder, self).__init__(vocab, regularizer)

        self.task = task
        self.encoder = encoder
        self.output_dim = encoder.get_output_dim()
        self.label_smoothing = label_smoothing
        self.num_classes = self.vocab.get_vocab_size(task)
        self.adaptive = adaptive
        self.features = features if features else []

        self.metrics = {
            "acc": CategoricalAccuracy(),
            # "acc3": CategoricalAccuracy(top_k=3)
        }

        if self.adaptive:
            # TODO
            adaptive_cutoffs = [round(self.num_classes / 15), 3 * round(self.num_classes / 15)]
            self.task_output = AdaptiveLogSoftmaxWithLoss(self.output_dim,
                                                          self.num_classes,
                                                          cutoffs=adaptive_cutoffs,
                                                          div_value=4.0)
        else:
            self.task_output = TimeDistributed(Linear(self.output_dim, self.num_classes))

        self.feature_outputs = torch.nn.ModuleDict()
        self.features_metrics = {}
        for feature in self.features:
            self.feature_outputs[feature] = TimeDistributed(Linear(self.output_dim,
                                                                   vocab.get_vocab_size(feature)))
            self.features_metrics[feature] = {
                "acc": CategoricalAccuracy(),
            }

        initializer(self) 

# 需要导入模块: from allennlp import nn [as 别名]
# 或者: from allennlp.nn import RegularizerApplicator [as 别名]
def __init__(self,
                 vocab: Vocabulary,
                 tasks: List[str],
                 text_field_embedder: TextFieldEmbedder,
                 encoder: Seq2SeqEncoder,
                 decoders: Dict[str, Model],
                 post_encoder_embedder: TextFieldEmbedder = None,
                 dropout: float = 0.0,
                 word_dropout: float = 0.0,
                 mix_embedding: int = None,
                 layer_dropout: int = 0.0,
                 initializer: InitializerApplicator = InitializerApplicator(),
                 regularizer: Optional[RegularizerApplicator] = None) -> None:
        super(UdifyModel, self).__init__(vocab, regularizer)

        self.tasks = sorted(tasks)
        self.vocab = vocab
        self.bert_vocab = BertTokenizer.from_pretrained("config/archive/bert-base-multilingual-cased/vocab.txt").vocab
        self.text_field_embedder = text_field_embedder
        self.post_encoder_embedder = post_encoder_embedder
        self.shared_encoder = encoder
        self.word_dropout = word_dropout
        self.dropout = torch.nn.Dropout(p=dropout)
        self.decoders = torch.nn.ModuleDict(decoders)

        if mix_embedding:
            self.scalar_mix = torch.nn.ModuleDict({
                task: ScalarMixWithDropout(mix_embedding,
                                           do_layer_norm=False,
                                           dropout=layer_dropout)
                for task in self.decoders
            })
        else:
            self.scalar_mix = None

        self.metrics = {}

        for task in self.tasks:
            if task not in self.decoders:
                raise ConfigurationError(f"Task {task} has no corresponding decoder. Make sure their names match.")

        check_dimensions_match(text_field_embedder.get_output_dim(), encoder.get_input_dim(),
                               "text field embedding dim", "encoder input dim")

        initializer(self)
        self._count_params() 

# 需要导入模块: from allennlp import nn [as 别名]
# 或者: from allennlp.nn import RegularizerApplicator [as 别名]
def __init__(self, vocab: Vocabulary,
                 encoder_keys: List[str],
                 mask_key: str,
                 pair2vec_config_file: str,
                 pair2vec_model_file: str,
                 text_field_embedder: TextFieldEmbedder,
                 encoder: Seq2SeqEncoder,
                 similarity_function: SimilarityFunction,
                 projection_feedforward: FeedForward,
                 inference_encoder: Seq2SeqEncoder,
                 output_feedforward: FeedForward,
                 output_logit: FeedForward,
                 initializer: InitializerApplicator = InitializerApplicator(),
                 dropout: float = 0.5,
                 pair2vec_dropout: float = 0.0,
                 bidirectional_pair2vec: bool = True,
                 regularizer: Optional[RegularizerApplicator] = None) -> None:
        super().__init__(vocab, regularizer)
        self._vocab = vocab
        self.pair2vec = util.get_pair2vec(pair2vec_config_file, pair2vec_model_file)
        self._encoder_keys = encoder_keys
        self._mask_key = mask_key
        self._text_field_embedder = text_field_embedder
        self._projection_feedforward = projection_feedforward
        self._encoder = encoder
        from allennlp.modules.matrix_attention import DotProductMatrixAttention

        self._matrix_attention = DotProductMatrixAttention()


        self._inference_encoder = inference_encoder
        self._pair2vec_dropout = torch.nn.Dropout(pair2vec_dropout)
        self._bidirectional_pair2vec = bidirectional_pair2vec

        if dropout:
            self.dropout = torch.nn.Dropout(dropout)
            self.rnn_input_dropout = VariationalDropout(dropout)
        else:
            self.dropout = None
            self.rnn_input_dropout = None

        self._output_feedforward = output_feedforward
        self._output_logit = output_logit

        self._num_labels = vocab.get_vocab_size(namespace="labels")


        self._accuracy = CategoricalAccuracy()
        self._loss = torch.nn.CrossEntropyLoss()

        initializer(self) 

# 需要导入模块: from allennlp import nn [as 别名]
# 或者: from allennlp.nn import RegularizerApplicator [as 别名]
def __init__(self, vocab: Vocabulary,
                 text_field_embedder: TextFieldEmbedder,
                 encoder: Seq2SeqEncoder,
                 similarity_function: SimilarityFunction,
                 projection_feedforward: FeedForward,
                 inference_encoder: Seq2SeqEncoder,
                 output_feedforward: FeedForward,
                 output_logit: FeedForward,
                 final_feedforward: FeedForward,
                 coverage_loss: CoverageLoss = None,
                 contextualize_pair_comparators: bool = False,
                 pair_context_encoder: Seq2SeqEncoder = None,
                 pair_feedforward: FeedForward = None,
                 optimize_coverage_for: List = ["entailment", "neutral"],
                 dropout: float = 0.5,
                 initializer: InitializerApplicator = InitializerApplicator(),
                 regularizer: Optional[RegularizerApplicator] = None) -> None:
        super().__init__(vocab, regularizer)
        self._label2idx = self.vocab.get_token_to_index_vocabulary('labels')

        self._text_field_embedder = text_field_embedder

        self._entailment_comparator_layer_1 = EsimComparatorLayer1(encoder, dropout)
        self._entailment_comparator_layer_2 = EsimComparatorLayer2(similarity_function)

        self._td_entailment_comparator_layer_1 = TimeDistributed(self._entailment_comparator_layer_1)
        self._td_entailment_comparator_layer_2 = TimeDistributed(self._entailment_comparator_layer_2)

        self._entailment_comparator_layer_3plus_local = EsimComparatorLayer3Plus(projection_feedforward, inference_encoder,
                                                                                 output_feedforward, dropout)
        self._td_entailment_comparator_layer_3plus_local = EachOutputTimeDistributed(self._entailment_comparator_layer_3plus_local)

        self._entailment_comparator_layer_3plus_global = copy.deepcopy(self._entailment_comparator_layer_3plus_local)

        self._contextualize_pair_comparators = contextualize_pair_comparators

        if not self._contextualize_pair_comparators:
            self._output_logit = output_logit
            self._td_output_logit = TimeDistributed(self._output_logit)

        self._final_feedforward = final_feedforward
        self._td_final_feedforward = TimeDistributed(final_feedforward)

        linear = torch.nn.Linear(2*self._entailment_comparator_layer_3plus_local.get_output_dim(),
                                 self._final_feedforward.get_input_dim())
        self._local_global_projection = torch.nn.Sequential(linear, torch.nn.ReLU())

        if self._contextualize_pair_comparators:
            self._pair_context_encoder = pair_context_encoder
            self._td_pair_feedforward = TimeDistributed(pair_feedforward)

        self._coverage_loss = coverage_loss

        # Do not apply initializer. If you do, make sure it doesn't reinitialize transferred parameters.