def __init__(
self,
experiment_name: str,
vocab: Vocabulary,
question_embedder: TextFieldEmbedder,
schema_encoder: Seq2SeqEncoder,
beam_encoder: Seq2SeqEncoder,
tree_rep_transformer: Seq2SeqEncoder,
utterance_augmenter: Seq2SeqEncoder,
beam_summarizer: Seq2SeqEncoder,
decoder_timesteps=9,
beam_size=30,
misc_params=None,
dropout: float = 0.1,
) -> None:
super().__init__(vocab)
self._experiment_name = experiment_name
self._misc_params = misc_params
self.set_flags()
self._utterance_augmenter = utterance_augmenter
self._action_dim = beam_encoder.get_output_dim()
self._beam_size = beam_size
self._n_schema_leafs = 15
self._num_values = 10
self.tokenizer = TokenIndexer.by_name("pretrained_transformer")(
model_name="Salesforce/grappa_large_jnt"
)._allennlp_tokenizer.tokenizer
if not self.cntx_reranker:
self._noreranker_cntx_linear = torch.nn.Linear(
in_features=self._action_dim,
out_features=2 * self._action_dim)
if not self.utt_aug:
self._nobeam_cntx_linear = torch.nn.Linear(
in_features=self._action_dim,
out_features=2 * self._action_dim)
self.activation_func = torch.nn.ReLU
if self.lin_after_cntx:
self.cntx_linear = torch.nn.Sequential(
torch.nn.Linear(2 * self._action_dim, 4 * self._action_dim),
torch.nn.Dropout(p=dropout),
torch.nn.LayerNorm(4 * self._action_dim),
self.activation_func(),
torch.nn.Linear(4 * self._action_dim, 2 * self._action_dim),
)
if self.cntx_rep:
self._cntx_rep_linear = torch.nn.Linear(
in_features=self._action_dim,
out_features=2 * self._action_dim)
self._create_action_dicts()
self.op_count = self.binary_op_count + self.unary_op_count
self.xent = torch.nn.CrossEntropyLoss()
self.type_embedding = torch.nn.Embedding(self.op_count,
self._action_dim)
self.summrize_vec = torch.nn.Embedding(num_embeddings=1,
embedding_dim=self._action_dim)
self.d_frontier = 2 * self._action_dim
self.left_emb = torch.nn.Linear(in_features=self.d_frontier,
out_features=self.d_frontier)
self.right_emb = torch.nn.Linear(in_features=self.d_frontier,
out_features=self.d_frontier)
self.after_add = torch.nn.Sequential(
torch.nn.Linear(self.d_frontier, self.d_frontier),
torch.nn.Dropout(p=dropout),
torch.nn.LayerNorm(self.d_frontier),
self.activation_func(),
torch.nn.Linear(self.d_frontier, self.d_frontier),
)
self._unary_frontier_embedder = torch.nn.Sequential(
torch.nn.Linear(self.d_frontier, self.d_frontier),
torch.nn.Dropout(p=dropout),
torch.nn.LayerNorm(self.d_frontier),
self.activation_func(),
torch.nn.Linear(self.d_frontier, self.d_frontier),
)
self.op_linear = torch.nn.Linear(in_features=self.d_frontier,
out_features=self.op_count)
self.pre_op_linear = torch.nn.Sequential(
torch.nn.Linear(self.d_frontier, self.d_frontier),
torch.nn.Dropout(p=dropout),
torch.nn.LayerNorm(self.d_frontier),
self.activation_func(),
)
assert (self._action_dim % 2) == 0
self.vocab = vocab
self._question_embedder = question_embedder
self._schema_encoder = schema_encoder
self._beam_encoder = beam_encoder
self._beam_summarizer = beam_summarizer
self._tree_rep_transformer = tree_rep_transformer
self._decoder_timesteps = decoder_timesteps
self._beam_size = beam_size
self.q_emb_dim = question_embedder.get_output_dim()
self.dropout_prob = dropout
self._action_dim = beam_encoder.get_output_dim()
self._span_score_func = torch.nn.Linear(self._action_dim, 2)
self._pooler = BagOfEmbeddingsEncoder(embedding_dim=self._action_dim)
self._rank_schema = torch.nn.Sequential(
torch.nn.Linear(self._action_dim, self._action_dim),
torch.nn.Dropout(p=dropout),
torch.nn.LayerNorm(self._action_dim),
torch.nn.Tanh(),
torch.nn.Linear(self._action_dim, 1),
)
self._rank_beam = torch.nn.Sequential(
torch.nn.Linear(2 * self._action_dim, 2 * self._action_dim),
torch.nn.Dropout(p=dropout),
torch.nn.LayerNorm(2 * self._action_dim),
torch.nn.Tanh(),
torch.nn.Linear(2 * self._action_dim, 1),
)
self._emb_to_action_dim = torch.nn.Linear(
in_features=self.q_emb_dim,
out_features=self._action_dim,
)
self._create_type_tensor()
self._bce_loss = torch.nn.BCEWithLogitsLoss(reduction="none")
self._softmax = torch.nn.Softmax(dim=1)
self._final_beam_acc = Average()
self._reranker_acc = Average()
self._spider_acc = Average()
self._leafs_acc = Average()
self._batch_size = -1
self._device = None
self._evaluate_func = partial(
evaluate_single,
db_dir=os.path.join("dataset", "database"),
table_file=os.path.join("dataset", "tables.json"),
)
def test_read_from_file(self, lazy):
reader = LevenshteinReader(
token_indexers={"tokens": TokenIndexer.by_name("single_id")()},
lazy=False)
self._check_outputs(reader)