def __init__(self,
vocab: Vocabulary,
sentence_embedder: TextFieldEmbedder,
action_embedding_dim: int,
encoder: Seq2SeqEncoder,
attention: Attention,
decoder_beam_search: BeamSearch,
max_decoding_steps: int,
dropout: float = 0.0) -> None:
super(NlvrDirectSemanticParser,
self).__init__(vocab=vocab,
sentence_embedder=sentence_embedder,
action_embedding_dim=action_embedding_dim,
encoder=encoder,
dropout=dropout)
self._decoder_trainer = MaximumMarginalLikelihood()
self._decoder_step = BasicTransitionFunction(
encoder_output_dim=self._encoder.get_output_dim(),
action_embedding_dim=action_embedding_dim,
input_attention=attention,
activation=Activation.by_name('tanh')(),
add_action_bias=False,
dropout=dropout)
self._decoder_beam_search = decoder_beam_search
self._max_decoding_steps = max_decoding_steps
self._action_padding_index = -1
def __init__(self,
vocab: Vocabulary,
utterance_embedder: TextFieldEmbedder,
action_embedding_dim: int,
encoder: Seq2SeqEncoder,
decoder_beam_search: BeamSearch,
max_decoding_steps: int,
input_attention: Attention,
add_action_bias: bool = True,
training_beam_size: int = None,
decoder_num_layers: int = 1,
dropout: float = 0.0,
rule_namespace: str = 'rule_labels',
database_file='/atis/atis.db') -> None:
# Atis semantic parser init
super().__init__(vocab)
self._utterance_embedder = utterance_embedder
self._encoder = encoder
self._max_decoding_steps = max_decoding_steps
self._add_action_bias = add_action_bias
if dropout > 0:
self._dropout = torch.nn.Dropout(p=dropout)
else:
self._dropout = lambda x: x
self._rule_namespace = rule_namespace
self._exact_match = Average()
self._valid_sql_query = Average()
self._action_similarity = Average()
self._denotation_accuracy = Average()
self._executor = SqlExecutor(database_file)
self._action_padding_index = -1 # the padding value used by IndexField
num_actions = vocab.get_vocab_size(self._rule_namespace)
if self._add_action_bias:
input_action_dim = action_embedding_dim + 1
else:
input_action_dim = action_embedding_dim
self._action_embedder = Embedding(num_embeddings=num_actions, embedding_dim=input_action_dim)
self._output_action_embedder = Embedding(num_embeddings=num_actions, embedding_dim=action_embedding_dim)
# This is what we pass as input in the first step of decoding, when we don't have a
# previous action, or a previous utterance attention.
self._first_action_embedding = torch.nn.Parameter(torch.FloatTensor(action_embedding_dim))
self._first_attended_utterance = torch.nn.Parameter(torch.FloatTensor(encoder.get_output_dim()))
torch.nn.init.normal_(self._first_action_embedding)
torch.nn.init.normal_(self._first_attended_utterance)
self._num_entity_types = 2 # TODO(kevin): get this in a more principled way somehow?
self._entity_type_decoder_embedding = Embedding(self._num_entity_types, action_embedding_dim)
self._decoder_num_layers = decoder_num_layers
self._beam_search = decoder_beam_search
self._decoder_trainer = MaximumMarginalLikelihood(training_beam_size)
self._transition_function = LinkingTransitionFunction(encoder_output_dim=self._encoder.get_output_dim(),
action_embedding_dim=action_embedding_dim,
input_attention=input_attention,
add_action_bias=self._add_action_bias,
dropout=dropout,
num_layers=self._decoder_num_layers)
def __init__(self,
vocab: Vocabulary,
mydatabase: str,
schema_path: str,
utterance_embedder: TextFieldEmbedder,
action_embedding_dim: int,
encoder: Seq2SeqEncoder,
decoder_beam_search: BeamSearch,
max_decoding_steps: int,
input_attention: Attention,
add_action_bias: bool = True,
dropout: float = 0.0,
initializer: InitializerApplicator = InitializerApplicator(),
regularizer: Optional[RegularizerApplicator] = None) -> None:
super().__init__(vocab, regularizer)
self._utterance_embedder = utterance_embedder
self._encoder = encoder
self._max_decoding_steps = max_decoding_steps
self._add_action_bias = add_action_bias
self._dropout = torch.nn.Dropout(p=dropout)
self._exact_match = Average()
self._action_similarity = Average()
self._valid_sql_query = SqlValidity(mydatabase=mydatabase)
self._token_match = TokenSequenceAccuracy()
self._kb_match = KnowledgeBaseConstsAccuracy(schema_path=schema_path)
self._schema_free_match = GlobalTemplAccuracy(schema_path=schema_path)
self._coverage_loss = CoverageAttentionLossMetric()
# the padding value used by IndexField
self._action_padding_index = -1
num_actions = vocab.get_vocab_size("rule_labels")
input_action_dim = action_embedding_dim
if self._add_action_bias:
input_action_dim += 1
self._action_embedder = Embedding(num_embeddings=num_actions,
embedding_dim=input_action_dim)
self._output_action_embedder = Embedding(
num_embeddings=num_actions, embedding_dim=action_embedding_dim)
# This is what we pass as input in the first step of decoding, when we don't have a
# previous action, or a previous utterance attention.
self._first_action_embedding = torch.nn.Parameter(
torch.FloatTensor(action_embedding_dim))
self._first_attended_utterance = torch.nn.Parameter(
torch.FloatTensor(encoder.get_output_dim()))
torch.nn.init.normal_(self._first_action_embedding)
torch.nn.init.normal_(self._first_attended_utterance)
self._beam_search = decoder_beam_search
self._decoder_trainer = MaximumMarginalLikelihood(beam_size=1)
self._transition_function = BasicTransitionFunction(
encoder_output_dim=self._encoder.get_output_dim(),
action_embedding_dim=action_embedding_dim,
input_attention=input_attention,
add_action_bias=self._add_action_bias,
dropout=dropout)
initializer(self)
def __init__(self,
vocab: Vocabulary,
utterance_embedder: TextFieldEmbedder,
action_embedding_dim: int,
encoder: Seq2SeqEncoder,
decoder_beam_search: BeamSearch,
max_decoding_steps: int,
input_attention: Attention,
database_file: str,
add_action_bias: bool = True,
dropout: float = 0.0,
initializer: InitializerApplicator = InitializerApplicator(),
regularizer: Optional[RegularizerApplicator] = None) -> None:
super().__init__(vocab, regularizer)
self._utterance_embedder = utterance_embedder
self._encoder = encoder
self._max_decoding_steps = max_decoding_steps
self._add_action_bias = add_action_bias
self._dropout = torch.nn.Dropout(p=dropout)
self._exact_match = Average()
self._valid_sql_query = Average()
self._action_similarity = Average()
self._denotation_accuracy = Average()
self._executor = SqlExecutor(database_file)
# the padding value used by IndexField
self._action_padding_index = -1
num_actions = vocab.get_vocab_size("rule_labels")
input_action_dim = action_embedding_dim
if self._add_action_bias:
input_action_dim += 1
self._action_embedder = Embedding(num_embeddings=num_actions,
embedding_dim=input_action_dim)
self._output_action_embedder = Embedding(
num_embeddings=num_actions, embedding_dim=action_embedding_dim)
# This is what we pass as input in the first step of decoding, when we don't have a
# previous action, or a previous utterance attention.
self._first_action_embedding = torch.nn.Parameter(
torch.FloatTensor(action_embedding_dim))
self._first_attended_utterance = torch.nn.Parameter(
torch.FloatTensor(encoder.get_output_dim()))
torch.nn.init.normal_(self._first_action_embedding)
torch.nn.init.normal_(self._first_attended_utterance)
self._beam_search = decoder_beam_search
self._decoder_trainer = MaximumMarginalLikelihood(beam_size=1)
self._transition_function = BasicTransitionFunction(
encoder_output_dim=self._encoder.get_output_dim(),
action_embedding_dim=action_embedding_dim,
input_attention=input_attention,
predict_start_type_separately=False,
add_action_bias=self._add_action_bias,
dropout=dropout)
initializer(self)
def setUp(self):
super().setUp()
self.initial_state = SimpleState(
[0, 1], [[], []],
[torch.Tensor([0.0]), torch.Tensor([0.0])], [0, 1])
self.decoder_step = SimpleTransitionFunction()
self.targets = torch.Tensor([[[2, 3, 4], [1, 3, 4], [1, 2, 4]],
[[3, 4, 0], [2, 3, 4], [0, 0, 0]]])
self.target_mask = torch.Tensor([[[1, 1, 1], [1, 1, 1], [1, 1, 1]],
[[1, 1, 0], [1, 1, 1], [0, 0, 0]]])
self.supervision = (self.targets, self.target_mask)
# High beam size ensures exhaustive search.
self.trainer = MaximumMarginalLikelihood()
def __init__(self,
vocab: Vocabulary,
question_embedder: TextFieldEmbedder,
action_embedding_dim: int,
encoder: Seq2SeqEncoder,
entity_encoder: Seq2VecEncoder,
decoder_beam_search: BeamSearch,
max_decoding_steps: int,
attention: Attention,
mixture_feedforward: FeedForward = None,
add_action_bias: bool = True,
training_beam_size: int = None,
use_neighbor_similarity_for_linking: bool = False,
dropout: float = 0.0,
num_linking_features: int = 10,
rule_namespace: str = 'rule_labels',
tables_directory: str = '/wikitables/') -> None:
use_similarity = use_neighbor_similarity_for_linking
super().__init__(vocab=vocab,
question_embedder=question_embedder,
action_embedding_dim=action_embedding_dim,
encoder=encoder,
entity_encoder=entity_encoder,
max_decoding_steps=max_decoding_steps,
add_action_bias=add_action_bias,
use_neighbor_similarity_for_linking=use_similarity,
dropout=dropout,
num_linking_features=num_linking_features,
rule_namespace=rule_namespace,
tables_directory=tables_directory)
self._beam_search = decoder_beam_search
self._decoder_trainer = MaximumMarginalLikelihood(training_beam_size)
self._decoder_step = LinkingTransitionFunction(
encoder_output_dim=self._encoder.get_output_dim(),
action_embedding_dim=action_embedding_dim,
input_attention=attention,
num_start_types=self._num_start_types,
predict_start_type_separately=True,
add_action_bias=self._add_action_bias,
mixture_feedforward=mixture_feedforward,
dropout=dropout)
def __init__(self,
vocab: Vocabulary,
encoder: Seq2SeqEncoder,
entity_encoder: Seq2VecEncoder,
decoder_beam_search: BeamSearch,
question_embedder: TextFieldEmbedder,
input_attention: Attention,
past_attention: Attention,
max_decoding_steps: int,
action_embedding_dim: int,
gnn: bool = True,
decoder_use_graph_entities: bool = True,
decoder_self_attend: bool = True,
gnn_timesteps: int = 2,
parse_sql_on_decoding: bool = True,
add_action_bias: bool = True,
use_neighbor_similarity_for_linking: bool = True,
dataset_path: str = 'dataset',
training_beam_size: int = None,
decoder_num_layers: int = 1,
dropout: float = 0.0,
rule_namespace: str = 'rule_labels',
scoring_dev_params: dict = None,
debug_parsing: bool = False) -> None:
super().__init__(vocab)
self.vocab = vocab
self._encoder = encoder
self._max_decoding_steps = max_decoding_steps
if dropout > 0:
self._dropout = torch.nn.Dropout(p=dropout)
else:
self._dropout = lambda x: x
self._rule_namespace = rule_namespace
self._question_embedder = question_embedder
self._add_action_bias = add_action_bias
self._scoring_dev_params = scoring_dev_params or {}
self.parse_sql_on_decoding = parse_sql_on_decoding
self._entity_encoder = TimeDistributed(entity_encoder)
self._use_neighbor_similarity_for_linking = use_neighbor_similarity_for_linking
self._self_attend = decoder_self_attend
self._decoder_use_graph_entities = decoder_use_graph_entities
self._action_padding_index = -1 # the padding value used by IndexField
self._exact_match = Average()
self._sql_evaluator_match = Average()
self._action_similarity = Average()
self._acc_single = Average()
self._acc_multi = Average()
self._beam_hit = Average()
self._action_embedding_dim = action_embedding_dim
num_actions = vocab.get_vocab_size(self._rule_namespace)
if self._add_action_bias:
input_action_dim = action_embedding_dim + 1
else:
input_action_dim = action_embedding_dim
self._action_embedder = Embedding(num_embeddings=num_actions,
embedding_dim=input_action_dim)
self._output_action_embedder = Embedding(
num_embeddings=num_actions, embedding_dim=action_embedding_dim)
encoder_output_dim = encoder.get_output_dim()
if gnn:
encoder_output_dim += action_embedding_dim
self._first_action_embedding = torch.nn.Parameter(
torch.FloatTensor(action_embedding_dim))
self._first_attended_utterance = torch.nn.Parameter(
torch.FloatTensor(encoder_output_dim))
self._first_attended_output = torch.nn.Parameter(
torch.FloatTensor(action_embedding_dim))
torch.nn.init.normal_(self._first_action_embedding)
torch.nn.init.normal_(self._first_attended_utterance)
torch.nn.init.normal_(self._first_attended_output)
self._num_entity_types = 9
self._embedding_dim = question_embedder.get_output_dim()
self._entity_type_encoder_embedding = Embedding(
self._num_entity_types, self._embedding_dim)
self._entity_type_decoder_embedding = Embedding(
self._num_entity_types, action_embedding_dim)
self._linking_params = torch.nn.Linear(16, 1)
torch.nn.init.uniform_(self._linking_params.weight, 0, 1)
num_edge_types = 3
self._gnn = GatedGraphConv(self._embedding_dim,
gnn_timesteps,
num_edge_types=num_edge_types,
dropout=dropout)
self._decoder_num_layers = decoder_num_layers
self._beam_search = decoder_beam_search
self._decoder_trainer = MaximumMarginalLikelihood(training_beam_size)
if decoder_self_attend:
self._transition_function = AttendPastSchemaItemsTransitionFunction(
encoder_output_dim=encoder_output_dim,
action_embedding_dim=action_embedding_dim,
input_attention=input_attention,
past_attention=past_attention,
predict_start_type_separately=False,
add_action_bias=self._add_action_bias,
dropout=dropout,
num_layers=self._decoder_num_layers)
else:
self._transition_function = LinkingTransitionFunction(
encoder_output_dim=encoder_output_dim,
action_embedding_dim=action_embedding_dim,
input_attention=input_attention,
predict_start_type_separately=False,
add_action_bias=self._add_action_bias,
dropout=dropout,
num_layers=self._decoder_num_layers)
self._ent2ent_ff = FeedForward(action_embedding_dim, 1,
action_embedding_dim,
Activation.by_name('relu')())
self._neighbor_params = torch.nn.Linear(self._embedding_dim,
self._embedding_dim)
# TODO: Remove hard-coded dirs
self._evaluate_func = partial(
evaluate,
db_dir=os.path.join(dataset_path, 'database'),
table=os.path.join(dataset_path, 'tables.json'),
check_valid=False)
self.debug_parsing = debug_parsing
def __init__(
self,
vocab: Vocabulary,
question_embedder: TextFieldEmbedder,
action_embedding_dim: int,
encoder: Seq2SeqEncoder,
decoder_beam_search: BeamSearch,
max_decoding_steps: int,
attention: Attention,
mixture_feedforward: FeedForward = None,
add_action_bias: bool = True,
dropout: float = 0.0,
num_linking_features: int = 0,
num_entity_bits: int = 0,
entity_bits_output: bool = True,
use_entities: bool = False,
denotation_only: bool = False,
# Deprecated parameter to load older models
entity_encoder: Seq2VecEncoder = None, # pylint: disable=unused-argument
entity_similarity_mode: str = "dot_product",
rule_namespace: str = 'rule_labels') -> None:
super(QuarelSemanticParser, self).__init__(vocab)
self._question_embedder = question_embedder
self._encoder = encoder
self._beam_search = decoder_beam_search
self._max_decoding_steps = max_decoding_steps
if dropout > 0:
self._dropout = torch.nn.Dropout(p=dropout)
else:
self._dropout = lambda x: x
self._rule_namespace = rule_namespace
self._denotation_accuracy = Average()
self._action_sequence_accuracy = Average()
self._has_logical_form = Average()
self._embedding_dim = question_embedder.get_output_dim()
self._use_entities = use_entities
# Note: there's only one non-trivial entity type in QuaRel for now, so most of the
# entity_type stuff is irrelevant
self._num_entity_types = 4 # TODO(mattg): get this in a more principled way somehow?
self._num_start_types = 1 # Hardcoded until we feed lf syntax into the model
self._entity_type_encoder_embedding = Embedding(
self._num_entity_types, self._embedding_dim)
self._entity_type_decoder_embedding = Embedding(
self._num_entity_types, action_embedding_dim)
self._entity_similarity_layer = None
self._entity_similarity_mode = entity_similarity_mode
if self._entity_similarity_mode == "weighted_dot_product":
self._entity_similarity_layer = \
TimeDistributed(torch.nn.Linear(self._embedding_dim, 1, bias=False))
# Center initial values around unweighted dot product
self._entity_similarity_layer._module.weight.data += 1 # pylint: disable=protected-access
elif self._entity_similarity_mode == "dot_product":
pass
else:
raise ValueError("Invalid entity_similarity_mode: {}".format(
self._entity_similarity_mode))
if num_linking_features > 0:
self._linking_params = torch.nn.Linear(num_linking_features, 1)
else:
self._linking_params = None
self._decoder_trainer = MaximumMarginalLikelihood()
self._encoder_output_dim = self._encoder.get_output_dim()
if entity_bits_output:
self._encoder_output_dim += num_entity_bits
self._entity_bits_output = entity_bits_output
self._debug_count = 10
self._num_denotation_cats = 2 # Hardcoded for simplicity
self._denotation_only = denotation_only
if self._denotation_only:
self._denotation_accuracy_cat = CategoricalAccuracy()
self._denotation_classifier = torch.nn.Linear(
self._encoder_output_dim, self._num_denotation_cats)
# Rest of init not needed for denotation only where no decoding to actions needed
return
self._action_padding_index = -1 # the padding value used by IndexField
num_actions = vocab.get_vocab_size(self._rule_namespace)
self._num_actions = num_actions
self._action_embedder = Embedding(num_embeddings=num_actions,
embedding_dim=action_embedding_dim)
# We are tying the action embeddings used for input and output
# self._output_action_embedder = Embedding(num_embeddings=num_actions, embedding_dim=action_embedding_dim)
self._output_action_embedder = self._action_embedder # tied weights
self._add_action_bias = add_action_bias
if self._add_action_bias:
self._action_biases = Embedding(num_embeddings=num_actions,
embedding_dim=1)
# This is what we pass as input in the first step of decoding, when we don't have a
# previous action, or a previous question attention.
self._first_action_embedding = torch.nn.Parameter(
torch.FloatTensor(action_embedding_dim))
self._first_attended_question = torch.nn.Parameter(
torch.FloatTensor(self._encoder_output_dim))
torch.nn.init.normal_(self._first_action_embedding)
torch.nn.init.normal_(self._first_attended_question)
self._decoder_step = LinkingTransitionFunction(
encoder_output_dim=self._encoder_output_dim,
action_embedding_dim=action_embedding_dim,
input_attention=attention,
num_start_types=self._num_start_types,
predict_start_type_separately=False,
add_action_bias=self._add_action_bias,
mixture_feedforward=mixture_feedforward,
dropout=dropout)
def __init__(
self,
vocab: Vocabulary,
encoder: Seq2SeqEncoder,
dropout: float = 0.0,
object_loss_multiplier: float = 0.0,
denotation_loss_multiplier: float = 1.0,
tokens_namespace: str = "tokens",
rule_namespace: str = "rule_labels",
denotation_namespace: str = "labels",
num_parse_only_batches: int = 0,
use_gold_program_for_eval: bool = False,
nmn_settings: Dict = None,
) -> None:
# Atis semantic parser init
super().__init__(vocab)
self._encoder = encoder
self._dropout = torch.nn.Dropout(p=dropout)
self._obj_loss_multiplier = object_loss_multiplier
self._denotation_loss_multiplier = denotation_loss_multiplier
self._tokens_namespace = tokens_namespace
self._rule_namespace = rule_namespace
self._denotation_namespace = denotation_namespace
self._num_parse_only_batches = num_parse_only_batches
self._use_gold_program_for_eval = use_gold_program_for_eval
self._nmn_settings = nmn_settings
self._training_batches_so_far = 0
self._denotation_accuracy = CategoricalAccuracy()
self._proposal_accuracy = CategoricalAccuracy()
# TODO(mattg): use FullSequenceMatch instead of this.
self._program_accuracy = Average()
self.loss = torch.nn.BCELoss()
self._action_padding_index = -1 # the padding value used by IndexField
num_actions = vocab.get_vocab_size(self._rule_namespace)
action_embedding_dim = 100
self._add_action_bias = True
if self._add_action_bias:
input_action_dim = action_embedding_dim + 1
else:
input_action_dim = action_embedding_dim
self._action_embedder = Embedding(num_embeddings=num_actions,
embedding_dim=input_action_dim)
self._output_action_embedder = Embedding(
num_embeddings=num_actions, embedding_dim=action_embedding_dim)
self._language_parameters = VisualReasoningGqaParameters(
hidden_dim=self._encoder.get_output_dim(),
initializer=self._encoder.encoder.model.init_bert_weights,
)
# This is what we pass as input in the first step of decoding, when we don't have a
# previous action, or a previous utterance attention.
self._first_action_embedding = torch.nn.Parameter(
torch.FloatTensor(action_embedding_dim))
# encoder_output_dim = self._lxrt_encoder.get_output_dim()
encoder_output_dim = self._encoder.get_output_dim()
self._first_attended_utterance = torch.nn.Parameter(
torch.FloatTensor(encoder_output_dim))
torch.nn.init.normal_(self._first_action_embedding)
torch.nn.init.normal_(self._first_attended_utterance)
self._decoder_num_layers = 1
self._beam_search = BeamSearch(beam_size=10)
self._decoder_trainer = MaximumMarginalLikelihood()
self._transition_function = BasicTransitionFunction(
encoder_output_dim=encoder_output_dim,
action_embedding_dim=action_embedding_dim,
input_attention=AdditiveAttention(vector_dim=encoder_output_dim,
matrix_dim=encoder_output_dim),
add_action_bias=self._add_action_bias,
dropout=dropout,
num_layers=self._decoder_num_layers,
)
self._language_parameters.apply(
self._encoder.encoder.model.init_bert_weights)
# attention.apply(self._lxrt_encoder.encoder.model.init_bert_weights)
# self._transition_function.apply(self._lxrt_encoder.encoder.model.init_bert_weights)
# Our language is constant across instances, so we just create one up front that we can
# re-use to construct the `GrammarStatelet`.
self._world = VisualReasoningGqaLanguage(None, None, None, None, None)
def __init__(self,
vocab: Vocabulary,
encoder: Seq2SeqEncoder,
entity_encoder: Seq2VecEncoder,
decoder_beam_search: BeamSearch,
question_embedder: TextFieldEmbedder,
input_attention: Attention,
past_attention: Attention,
graph_attention: Attention,
max_decoding_steps: int,
action_embedding_dim: int,
enable_gating: bool = False,
ablation_mode: str = None,
gnn: bool = True,
graph_loss_lambda: float = 0.5,
decoder_use_graph_entities: bool = True,
decoder_self_attend: bool = True,
gnn_timesteps: int = 2,
pruning_gnn_timesteps: int = 2,
parse_sql_on_decoding: bool = True,
add_action_bias: bool = False,
use_neighbor_similarity_for_linking: bool = True,
dataset_path: str = 'dataset',
log_path: str = '',
training_beam_size: int = None,
decoder_num_layers: int = 1,
dropout: float = 0.0,
rule_namespace: str = 'rule_labels') -> None:
super().__init__(vocab, encoder, entity_encoder, question_embedder, gnn_timesteps, dropout, rule_namespace)
self.enable_gating = enable_gating
self.ablation_mode = ablation_mode
self._log_path = log_path
self._max_decoding_steps = max_decoding_steps
self._add_action_bias = add_action_bias
self._parse_sql_on_decoding = parse_sql_on_decoding
self._self_attend = decoder_self_attend
self._decoder_use_graph_entities = decoder_use_graph_entities
self._use_neighbor_similarity_for_linking = use_neighbor_similarity_for_linking
self._action_padding_index = -1 # the padding value used by IndexField
self._exact_match = Average()
self._sql_evaluator_match = Average()
self._action_similarity = Average()
self._beam_hit = Average()
self._action_embedding_dim = action_embedding_dim
self._graph_loss_lambda = graph_loss_lambda
num_actions = vocab.get_vocab_size(self._rule_namespace)
if self._add_action_bias:
input_action_dim = action_embedding_dim + 1
else:
input_action_dim = action_embedding_dim
self._action_embedder = Embedding(num_embeddings=num_actions, embedding_dim=input_action_dim)
self._output_action_embedder = Embedding(num_embeddings=num_actions, embedding_dim=action_embedding_dim)
self._embedding_projector = torch.nn.Linear(question_embedder.get_output_dim(), self._embedding_dim, bias=False)
self._bert_embedding_dim = question_embedder.get_output_dim()
encoder_output_dim = self._encoder.get_output_dim() + self._embedding_dim
self._neighbor_encoder = TimeDistributed(BagOfEmbeddingsEncoder(self._embedding_dim, averaged=True))
self._first_action_embedding = torch.nn.Parameter(torch.FloatTensor(action_embedding_dim))
self._first_attended_utterance = torch.nn.Parameter(torch.FloatTensor(encoder_output_dim))
self._first_attended_output = torch.nn.Parameter(torch.FloatTensor(action_embedding_dim))
torch.nn.init.normal_(self._first_action_embedding)
torch.nn.init.normal_(self._first_attended_utterance)
torch.nn.init.normal_(self._first_attended_output)
self._entity_type_decoder_embedding = Embedding(self._num_entity_types, action_embedding_dim)
self._decoder_num_layers = decoder_num_layers
self._beam_search = decoder_beam_search
self._decoder_trainer = MaximumMarginalLikelihood(training_beam_size)
self._graph_pruning = GraphPruning(3, self._embedding_dim, encoder.get_output_dim(), dropout,
timesteps=pruning_gnn_timesteps)
if decoder_self_attend:
self._transition_function = AttendPastSchemaItemsTransitionFunction(encoder_output_dim=encoder_output_dim,
action_embedding_dim=action_embedding_dim,
input_attention=input_attention,
past_attention=past_attention,
enable_gating=self.enable_gating,
ablation_mode=self.ablation_mode,
predict_start_type_separately=False,
add_action_bias=self._add_action_bias,
dropout=dropout,
num_layers=self._decoder_num_layers)
else:
self._transition_function = LinkingTransitionFunction(encoder_output_dim=encoder_output_dim,
action_embedding_dim=action_embedding_dim,
input_attention=input_attention,
predict_start_type_separately=False,
add_action_bias=self._add_action_bias,
dropout=dropout,
num_layers=self._decoder_num_layers)
if self.enable_gating:
self._graph_attention = graph_attention
else:
self._graph_attention = DotProductAttention()
self._embedding_sim_attn = CosineMatrixAttention()
# TODO: Remove hard-coded dirs
self._evaluate_func = partial(evaluate,
db_dir=os.path.join(dataset_path, 'database'),
table=os.path.join(dataset_path, 'tables.json'),
check_valid=False)
def __init__(self,
vocab: Vocabulary,
decoder_beam_search: BeamSearch,
input_attention: Attention,
past_attention: Attention,
max_decoding_steps: int,
action_embedding_dim: int,
decoder_self_attend: bool = True,
parse_sql_on_decoding: bool = True,
add_action_bias: bool = True,
dataset_path: str = 'spider',
training_beam_size: int = None,
decoder_num_layers: int = 1,
dropout: float = 0.0,
rule_namespace: str = 'rule_labels',
scoring_dev_params: dict = None,
debug_parsing: bool = False) -> None:
super().__init__(vocab)
self.vocab = vocab
self._max_decoding_steps = max_decoding_steps
if dropout > 0:
self._dropout = torch.nn.Dropout(p=dropout)
else:
self._dropout = lambda x: x
self._rule_namespace = rule_namespace
self._add_action_bias = add_action_bias
self._scoring_dev_params = scoring_dev_params or {}
self.parse_sql_on_decoding = parse_sql_on_decoding
self._self_attend = decoder_self_attend
self._action_padding_index = -1
self._exact_match = Average()
self._sql_evaluator_match = Average()
self._action_similarity = Average()
self._acc_single = Average()
self._acc_multi = Average()
self._beam_hit = Average()
num_actions = vocab.get_vocab_size(self._rule_namespace)
if self._add_action_bias:
input_action_dim = action_embedding_dim + 1
else:
input_action_dim = action_embedding_dim
self._action_embedder = Embedding(num_embeddings=num_actions,
embedding_dim=input_action_dim)
self._output_action_embedder = Embedding(
num_embeddings=num_actions, embedding_dim=action_embedding_dim)
encoder_output_dim = 256
self._first_action_embedding = torch.nn.Parameter(
torch.FloatTensor(action_embedding_dim))
self._first_attended_utterance = torch.nn.Parameter(
torch.FloatTensor(encoder_output_dim))
self._first_attended_output = torch.nn.Parameter(
torch.FloatTensor(action_embedding_dim))
torch.nn.init.normal_(self._first_action_embedding)
torch.nn.init.normal_(self._first_attended_utterance)
torch.nn.init.normal_(self._first_attended_output)
# The linear layer transforms the initial encoding to match the dimensions specified in the article.
self.input_linear = Linear(768, 256)
# A single layer of the RAT encoder.
self.rat_layer = RatEncoderLayer(d_model=256,
nhead=8,
dim_feedforward=1024)
# A transformer encoder module consists of 8 RAT layers.
self.encoder = RatEncoder(self.rat_layer, 8)
self._decoder_num_layers = decoder_num_layers
self._beam_search = decoder_beam_search
self._decoder_trainer = MaximumMarginalLikelihood(training_beam_size)
def __init__(self,
question_embedder: TextFieldEmbedder,
input_memory_embedder: TextFieldEmbedder,
output_memory_embedder: TextFieldEmbedder,
question_encoder: Seq2SeqEncoder,
input_memory_encoder: Seq2VecEncoder,
output_memory_encoder: Seq2VecEncoder,
decoder_beam_search: BeamSearch,
input_attention: Attention,
past_attention: Attention,
action_embedding_dim: int,
max_decoding_steps: int,
nhop: int,
decoding_nhop: int,
vocab: Vocabulary,
dataset_path: str = 'dataset',
parse_sql_on_decoding: bool = True,
training_beam_size: int = None,
add_action_bias: bool = True,
decoder_self_attend: bool = True,
decoder_num_layers: int = 1,
dropout: float = 0.0,
rule_namespace: str = 'rule_labels') -> None:
super().__init__(vocab)
self.question_embedder = question_embedder
self._input_mm_embedder = input_memory_embedder
self._output_mm_embedder = output_memory_embedder
self._question_encoder = question_encoder
self._input_mm_encoder = TimeDistributed(input_memory_encoder)
self._output_mm_encoder = TimeDistributed(output_memory_encoder)
self.parse_sql_on_decoding = parse_sql_on_decoding
self._self_attend = decoder_self_attend
self._max_decoding_steps = max_decoding_steps
self._add_action_bias = add_action_bias
self._rule_namespace = rule_namespace
num_actions = vocab.get_vocab_size(self._rule_namespace)
if self._add_action_bias:
input_action_dim = action_embedding_dim + 1
else:
input_action_dim = action_embedding_dim
self._action_embedder = Embedding(num_embeddings=num_actions,
embedding_dim=input_action_dim)
self._input_action_embedder = Embedding(
num_embeddings=num_actions, embedding_dim=action_embedding_dim)
self._output_action_embedder = Embedding(
num_embeddings=num_actions, embedding_dim=action_embedding_dim)
self._num_entity_types = 9
self._entity_type_decoder_input_embedding = Embedding(
self._num_entity_types, action_embedding_dim)
self._entity_type_decoder_output_embedding = Embedding(
self._num_entity_types, action_embedding_dim)
self._entity_type_encoder_embedding = Embedding(
self._num_entity_types,
(int)(question_encoder.get_output_dim() / 2))
self._decoder_num_layers = decoder_num_layers
self._action_embedding_dim = action_embedding_dim
self._ent2ent_ff = FeedForward(action_embedding_dim, 1,
action_embedding_dim,
Activation.by_name('relu')())
if dropout > 0:
self._dropout = torch.nn.Dropout(p=dropout)
else:
self._dropout = lambda x: x
self._first_action_embedding = torch.nn.Parameter(
torch.FloatTensor(action_embedding_dim))
self._first_attended_utterance = torch.nn.Parameter(
torch.FloatTensor(question_encoder.get_output_dim()))
torch.nn.init.normal_(self._first_action_embedding)
torch.nn.init.normal_(self._first_attended_utterance)
if self._self_attend:
self._transition_function = AttendPastSchemaItemsTransitionFunction(
encoder_output_dim=question_encoder.get_output_dim(),
action_embedding_dim=action_embedding_dim,
input_attention=input_attention,
past_attention=past_attention,
decoding_nhop=decoding_nhop,
predict_start_type_separately=False,
add_action_bias=self._add_action_bias,
dropout=dropout,
num_layers=self._decoder_num_layers)
else:
self._transition_function = LinkingTransitionFunction(
encoder_output_dim=question_encoder.get_output_dim(),
action_embedding_dim=action_embedding_dim,
input_attention=input_attention,
predict_start_type_separately=False,
add_action_bias=self._add_action_bias,
dropout=dropout,
num_layers=self._decoder_num_layers)
self._mm_attn = MemAttn(question_encoder.get_output_dim(), nhop)
self._beam_search = decoder_beam_search
self._decoder_trainer = MaximumMarginalLikelihood(training_beam_size)
self._action_padding_index = -1 # the padding value used by IndexField
self._exact_match = Average()
self._sql_evaluator_match = Average()
self._action_similarity = Average()
self._acc_single = Average()
self._acc_multi = Average()
self._beam_hit = Average()
# TODO: Remove hard-coded dirs
self._evaluate_func = partial(
evaluate,
db_dir=os.path.join(dataset_path, 'database'),
table=os.path.join(dataset_path, 'tables.json'),
check_valid=False)