def __init__(self,
input_dim: int,
hidden_dim: int,
projection_dim: int,
feedforward_hidden_dim: int,
num_layers: int,
num_attention_heads: int,
dropout_prob: float = 0.1,
residual_dropout_prob: float = 0.2,
attention_dropout_prob: float = 0.1) -> None:
super().__init__()
self._attention_layers: List[MultiHeadSelfAttention] = []
self._feedfoward_layers: List[FeedForward] = []
self._layer_norm_layers: List[LayerNorm] = []
self._feed_forward_layer_norm_layers: List[LayerNorm] = []
self._reset_gate_layers: List[FeedForward] = []
feedfoward_input_dim = input_dim
for i in range(num_layers):
feedfoward = FeedForward(
feedfoward_input_dim,
activations=[
Activation.by_name('relu')(),
Activation.by_name('linear')()
],
hidden_dims=[feedforward_hidden_dim, hidden_dim],
num_layers=2,
dropout=dropout_prob)
# Note: Please use `ModuleList` in new code. It provides better
# support for running on multiple GPUs. We've kept `add_module` here
# solely for backwards compatibility with existing serialized models.
self.add_module(f"feedforward_{i}", feedfoward)
self._feedfoward_layers.append(feedfoward)
feedforward_layer_norm = LayerNorm(feedfoward.get_output_dim())
self.add_module(f"feedforward_layer_norm_{i}",
feedforward_layer_norm)
self._feed_forward_layer_norm_layers.append(feedforward_layer_norm)
self_attention = MultiHeadSelfAttention(
num_heads=num_attention_heads,
input_dim=hidden_dim,
attention_dim=projection_dim,
values_dim=projection_dim,
attention_dropout_prob=attention_dropout_prob)
self.add_module(f"self_attention_{i}", self_attention)
self._attention_layers.append(self_attention)
reset_gate = FeedForward(
feedforward_hidden_dim,
activations=Activation.by_name('sigmoid')(),
hidden_dims=hidden_dim,
num_layers=1,
dropout=dropout_prob)
self.add_module(f"reset_gate_{i}", reset_gate)
self._reset_gate_layers.append(reset_gate)
layer_norm = LayerNorm(self_attention.get_output_dim())
self.add_module(f"layer_norm_{i}", layer_norm)
self._layer_norm_layers.append(layer_norm)
feedfoward_input_dim = hidden_dim
self.dropout = Dropout(residual_dropout_prob)
self._input_dim = input_dim
self._output_dim = self._attention_layers[-1].get_output_dim()
def __init__(self,
vocab: Vocabulary,
text_field_embedder: TextFieldEmbedder,
num_highway_layers: int,
phrase_layer: Seq2SeqEncoder,
similarity_function: SimilarityFunction,
multi_headed_attention_layer: MultiHeadSelfAttention,
modeling_layer: Seq2SeqEncoder,
span_end_encoder: Seq2SeqEncoder,
dropout: float = 0.2,
mask_lstms: bool = True,
initializer: InitializerApplicator = InitializerApplicator(),
regularizer: Optional[RegularizerApplicator] = None) -> None:
super(BidirectionalAttentionFlow, self).__init__(vocab, regularizer)
self._text_field_embedder = text_field_embedder
self._highway_layer = TimeDistributed(
Highway(text_field_embedder.get_output_dim(), num_highway_layers))
self._phrase_layer = phrase_layer
self._matrix_attention = LegacyMatrixAttention(similarity_function)
self._modeling_layer = modeling_layer
self._span_end_encoder = span_end_encoder
#New Self Attention layer
self._self_attention_layer = multi_headed_attention_layer
self._sa_matrix_attention = LegacyMatrixAttention(similarity_function)
selfattent_dim = multi_headed_attention_layer.get_output_dim()
#print("Self Attention Output Dim:",selfattent_dim,"\n")
encoding_dim = phrase_layer.get_output_dim()
modeling_dim = modeling_layer.get_output_dim()
#span_start_input_dim = encoding_dim * 4 + modeling_dim
span_start_input_dim = encoding_dim * 4 + modeling_dim + 2 * selfattent_dim
self._span_start_predictor = TimeDistributed(
torch.nn.Linear(span_start_input_dim, 1))
span_end_encoding_dim = span_end_encoder.get_output_dim()
#span_end_input_dim = encoding_dim * 4 + span_end_encoding_dim
span_end_input_dim = encoding_dim * 4 + span_end_encoding_dim + 2 * selfattent_dim
self._span_end_predictor = TimeDistributed(
torch.nn.Linear(span_end_input_dim, 1))
# Bidaf has lots of layer dimensions which need to match up - these aren't necessarily
# obvious from the configuration files, so we check here.
check_dimensions_match(modeling_layer.get_input_dim(),
4 * encoding_dim + 2 * selfattent_dim,
"modeling layer input dim", "4 * encoding dim")
check_dimensions_match(text_field_embedder.get_output_dim(),
phrase_layer.get_input_dim(),
"text field embedder output dim",
"phrase layer input dim")
check_dimensions_match(
span_end_encoder.get_input_dim(),
4 * encoding_dim + 3 * modeling_dim + 2 * selfattent_dim,
"span end encoder input dim",
"4 * encoding dim + 3 * modeling dim")
self._na_accuracy = CategoricalAccuracy()
self._span_start_accuracy = CategoricalAccuracy()
self._span_end_accuracy = CategoricalAccuracy()
self._span_accuracy = BooleanAccuracy()
self._squad_metrics = SquadEmAndF1()
self._na_dense = lambda in_dim: torch.nn.Linear(in_dim, 2).cuda()
if dropout > 0:
self._dropout = torch.nn.Dropout(p=dropout)
else:
self._dropout = lambda x: x
self._mask_lstms = mask_lstms
initializer(self)
