def __init__(self, vocab: Vocabulary,
text_field_embedder: TextFieldEmbedder,
gate_sent_encoder: Seq2SeqEncoder,
gate_self_attention_layer: Seq2SeqEncoder,
bert_projection: FeedForward,
span_gate: Seq2SeqEncoder,
dropout: float = 0.2,
output_att_scores: bool = True,
regularizer: Optional[RegularizerApplicator] = None) -> None:
super(PTNChainBidirectionalAttentionFlow, self).__init__(vocab, regularizer)
self._text_field_embedder = text_field_embedder
self._dropout = torch.nn.Dropout(p=dropout)
self._output_att_scores = output_att_scores
self._span_gate = span_gate
self._bert_projection = bert_projection
#self._gate_sent_encoder = gate_sent_encoder
self._gate_self_attention_layer = gate_self_attention_layer
self._gate_sent_encoder = None
self._gate_self_attention_layer = None
self._f1_metrics = AttF1Measure(0.5, top_k=False)
self._loss_trackers = {'loss': Average(),
'rl_loss': Average()}
self.evd_sup_acc_metric = ChainAccuracy()
self.evd_ans_metric = Average()
self.evd_beam_ans_metric = Average()
def __init__(self,
vocab: Vocabulary,
text_field_embedder: TextFieldEmbedder,
dropout: float = 0.2,
regularizer: Optional[RegularizerApplicator] = None) -> None:
super(BidirectionalAttentionFlow, self).__init__(vocab, regularizer)
self._text_field_embedder = text_field_embedder
encoding_dim = text_field_embedder.get_output_dim()
self._dropout = torch.nn.Dropout(p=dropout)
self._squad_metrics = SquadEmAndF1()
self._categorical_acc = CategoricalAccuracy()
self._coref_f1_metric = AttF1Measure(0.1)
self.linear_start = nn.Linear(encoding_dim, 1)
self.linear_end = nn.Linear(encoding_dim, 1)
self.linear_type = nn.Linear(encoding_dim, 3)
self._loss_trackers = {'loss': Average()}
def calc_evd_f1(pred_labels, gold_labels):
evd_metric = AttF1Measure(0.5) # We just use 0.5 as the TH since pred_labels should only contain 0 and 1
T_P, N_P, N_T = evd_metric(torch.tensor(pred_labels).float(), torch.tensor(gold_labels).float())
precision = float(T_P) / float(N_P + 1e-13)
recall = float(T_P) / float(N_T + 1e-13)
f1 = 2. * ((precision * recall) / (precision + recall + 1e-13))
return precision, recall, f1
def __init__(self,
vocab: Vocabulary,
text_field_embedder: TextFieldEmbedder,
phrase_layer: Seq2SeqEncoder,
modeling_layer: Seq2SeqEncoder,
span_start_encoder: Seq2SeqEncoder,
span_end_encoder: Seq2SeqEncoder,
dropout: float = 0.2,
regularizer: Optional[RegularizerApplicator] = None) -> None:
super(BidirectionalAttentionFlow, self).__init__(vocab, regularizer)
self._text_field_embedder = text_field_embedder
self._phrase_layer = phrase_layer
encoding_dim = phrase_layer.get_output_dim()
self._modeling_layer = modeling_layer
self._dropout = torch.nn.Dropout(p=dropout)
self._squad_metrics = SquadEmAndF1()
self._f1_metrics = F1Measure(1)
self.linear_1 = nn.Sequential(
nn.Linear(encoding_dim * 4, encoding_dim), nn.ReLU())
self.linear_2 = nn.Sequential(
nn.Linear(encoding_dim * 4, encoding_dim), nn.ReLU())
self.qc_att = BiAttention(encoding_dim, dropout)
self._coref_f1_metric = AttF1Measure(0.1)
self._span_start_encoder = span_start_encoder
self._span_end_encoder = span_end_encoder
self.linear_start = nn.Linear(encoding_dim, 1)
self.linear_end = nn.Linear(encoding_dim, 1)
self.linear_type = nn.Linear(encoding_dim * 3, 3)
self._loss_trackers = {
'loss': Average(),
'start_loss': Average(),
'end_loss': Average(),
'type_loss': Average()
}
def __init__(self, vocab: Vocabulary,
text_field_embedder: TextFieldEmbedder,
phrase_layer: Seq2SeqEncoder,
phrase_layer_sp: Seq2SeqEncoder,
span_start_encoder: Seq2SeqEncoder,
span_end_encoder: Seq2SeqEncoder,
self_attention_layer: Seq2SeqEncoder,
gate_sent_encoder: Seq2SeqEncoder,
gate_self_attention_layer: Seq2SeqEncoder,
type_encoder: Seq2SeqEncoder,
modeling_layer: Seq2SeqEncoder,
modeling_layer_sp: Seq2SeqEncoder,
span_gate: Seq2SeqEncoder,
dropout: float = 0.2,
output_att_scores: bool = True,
weights_file: str = None,
ft_reward: str = "ans",
regularizer: Optional[RegularizerApplicator] = None) -> None:
super(FineTuneRLBidirectionalAttentionFlow, self).__init__(vocab, regularizer)
self._text_field_embedder = text_field_embedder
self._phrase_layer = phrase_layer
self._phrase_layer_sp = phrase_layer_sp
self._dropout = torch.nn.Dropout(p=dropout)
self._modeling_layer = modeling_layer
self._modeling_layer_sp = modeling_layer_sp
self._span_start_encoder = span_start_encoder
self._span_end_encoder = span_end_encoder
self._type_encoder = type_encoder
self._self_attention_layer = self_attention_layer
self._output_att_scores = output_att_scores
self._ft_reward = ft_reward.split('+')
encoding_dim = span_start_encoder.get_output_dim()
self._span_gate = span_gate
self.qc_att = BiAttention(encoding_dim, 0.0)
self.qc_att_sp = BiAttention(encoding_dim, dropout)
self._gate_sent_encoder = gate_sent_encoder
self._gate_self_attention_layer = gate_self_attention_layer
self.linear_start = nn.Linear(encoding_dim, 1)
self.linear_end = nn.Linear(encoding_dim, 1)
self.linear_type = nn.Linear(encoding_dim * 3, 3)
self._squad_metrics = SquadEmAndF1_RT()
self._f1_metrics = AttF1Measure(0.5, top_k=False)
self._loss_trackers = {'loss': Average(),
'rl_loss': Average()}
if self._span_gate.evd_decoder._train_helper_type == 'teacher_forcing':
raise ValueError("train_helper should not be teacher forcing during fine tune!")
self.evd_sup_acc_metric = ChainAccuracy()
self.evd_ans_metric = Average()
self.evd_beam_ans_metric = Average()
self.evd_beam2_ans_metric = Average()
# load the weights
if weights_file:
model_state = torch.load(weights_file, map_location=util.device_mapping(0))
self.load_state_dict(model_state)
for p in self._text_field_embedder.parameters():
p.requires_grad = False
def __init__(self,
vocab: Vocabulary,
text_field_embedder: TextFieldEmbedder,
phrase_layer: Seq2SeqEncoder,
phrase_layer_sp: Seq2SeqEncoder,
span_start_encoder: Seq2SeqEncoder,
span_end_encoder: Seq2SeqEncoder,
self_attention_layer: Seq2SeqEncoder,
gate_sent_encoder: Seq2SeqEncoder,
gate_self_attention_layer: Seq2SeqEncoder,
type_encoder: Seq2SeqEncoder,
modeling_layer: Seq2SeqEncoder,
modeling_layer_sp: Seq2SeqEncoder,
span_gate: Seq2SeqEncoder,
dropout: float = 0.2,
output_att_scores: bool = True,
regularizer: Optional[RegularizerApplicator] = None) -> None:
super(RLBidirectionalAttentionFlow, self).__init__(vocab, regularizer)
self._text_field_embedder = text_field_embedder
self._phrase_layer = phrase_layer
self._phrase_layer_sp = phrase_layer_sp
self._dropout = torch.nn.Dropout(p=dropout)
self._modeling_layer = modeling_layer
self._modeling_layer_sp = modeling_layer_sp
self._span_start_encoder = span_start_encoder
self._span_end_encoder = span_end_encoder
self._type_encoder = type_encoder
self._self_attention_layer = self_attention_layer
self._output_att_scores = output_att_scores
encoding_dim = span_start_encoder.get_output_dim()
self._span_gate = span_gate
self.qc_att = BiAttention(encoding_dim, dropout)
self.qc_att_sp = BiAttention(encoding_dim, dropout)
self._gate_sent_encoder = gate_sent_encoder
self._gate_self_attention_layer = gate_self_attention_layer
self.linear_start = nn.Linear(encoding_dim, 1)
self.linear_end = nn.Linear(encoding_dim, 1)
self.linear_type = nn.Linear(encoding_dim * 3, 3)
self._squad_metrics = SquadEmAndF1_RT()
self._f1_metrics = AttF1Measure(0.5, top_k=False)
self._loss_trackers = {
'loss': Average(),
'start_loss': Average(),
'end_loss': Average(),
'type_loss': Average(),
'rl_loss': Average()
}
self.evd_sup_acc_metric = ChainAccuracy()
self.evd_ans_metric = Average()
self.evd_beam_ans_metric = Average()