def glove_linear_cnn(
glove_path: Union[Path, TextIO],
glove_dim: int,
num_classes: int,
vocab: Vocabulary,
filters: List[int],
out_channels: int,
freeze: bool = True,
saved_glove_file: Optional[Path] = None,
dropout: float = 0,
) -> LinearCnn:
"""Return LinearCnn with embedding layer initialised with GloVe."""
glove = load_glove(
input_file=glove_path,
glove_dim=glove_dim,
vocab=vocab,
saved_glove_file=saved_glove_file,
)
embedding = nn.Embedding.from_pretrained(glove, freeze=freeze)
return LinearCnn(
embedding=embedding,
num_classes=num_classes,
filters=filters,
out_channels=out_channels,
dropout=dropout,
)
def test_synthetic_glove() -> None:
"""Test if GloVe loader works with synthetic data."""
vocab = data.Vocabulary(["test", "new"])
input_file = StringIO(glove_str)
tensor = glove.load_glove(input_file, vocab, 3)
assert tensor.shape == (vocab.vocab_size(), 3)
def glove_bc_lstm(
glove_path: Union[Path, TextIO],
glove_dim: int,
num_classes: int,
vocab: Vocabulary,
filters: List[int],
out_channels: int,
freeze: bool = True,
saved_glove_file: Optional[Path] = None,
rnn_hidden_size: int = 100,
rnn_num_layers: int = 1,
bidirectional: bool = False,
rnn_dropout: float = 0,
) -> BcLstm:
"""
Return bcLSTM with embedding layer initialised with GloVe.
Both RNNs used (utterance and dialogue) are LSTMs.
"""
glove = load_glove(
input_file=glove_path,
glove_dim=glove_dim,
vocab=vocab,
saved_glove_file=saved_glove_file,
)
embedding = nn.Embedding.from_pretrained(glove, freeze=freeze)
utterance_lstm = nn.LSTM(
input_size=out_channels * len(filters),
hidden_size=rnn_hidden_size,
num_layers=rnn_num_layers,
bidirectional=bidirectional,
dropout=rnn_dropout,
batch_first=True,
)
dialogue_input = (1 + bidirectional) * utterance_lstm.hidden_size
dialogue_lstm = nn.LSTM(
input_size=dialogue_input,
hidden_size=rnn_hidden_size,
num_layers=rnn_num_layers,
bidirectional=False, # this doesn't make sense to be bidirectional [1]
dropout=rnn_dropout,
batch_first=True,
)
return BcLstm(
embedding=embedding,
num_classes=num_classes,
dialogue_rnn=dialogue_lstm,
utterance_rnn=utterance_lstm,
filters=filters,
out_channels=out_channels,
)
def test_real_glove() -> None:
"""Test if GloVe loader works with real data."""
glove_dim = 50
data_path = Path('./data/dev/metadata.csv')
dataset = MeldLinearTextDataset(data_path)
input_file = Path(f'./data/glove/glove.6B.{glove_dim}d.txt')
tensor = glove.load_glove(input_file, dataset.vocab, glove_dim)
assert tensor.shape == (dataset.vocab.vocab_size(), glove_dim)
def glove_contextual_lstm(
glove_path: Union[Path, TextIO],
glove_dim: int,
num_classes: int,
vocab: Vocabulary,
freeze: bool = True,
saved_glove_file: Optional[Path] = None,
rnn_hidden_size: int = 100,
rnn_num_layers: int = 1,
bidirectional: bool = False,
rnn_dropout: float = 0,
) -> ContextualRnn:
"""
Return ContextualRnn with embedding layer initialised with GloVe.
Both RNNs used (utterance and dialogue) are LSTMs.
"""
glove = load_glove(
input_file=glove_path,
glove_dim=glove_dim,
vocab=vocab,
saved_glove_file=saved_glove_file,
)
embedding = nn.Embedding.from_pretrained(glove, freeze=freeze)
utterance_lstm = nn.LSTM(
input_size=embedding.embedding_dim,
hidden_size=rnn_hidden_size,
num_layers=rnn_num_layers,
bidirectional=bidirectional,
dropout=rnn_dropout,
batch_first=True,
)
dialogue_lstm = nn.LSTM(
input_size=utterance_lstm.hidden_size,
hidden_size=rnn_hidden_size,
num_layers=rnn_num_layers,
bidirectional=False, # this doesn't make sense to be bidirectional [1]
dropout=rnn_dropout,
batch_first=True,
)
return ContextualRnn(
embedding=embedding,
num_classes=num_classes,
dialogue_rnn=dialogue_lstm,
utterance_rnn=utterance_lstm,
)
def glove_linear_cnn_lstm(
glove_path: Union[Path, TextIO],
glove_dim: int,
num_classes: int,
vocab: Vocabulary,
filters: List[int],
out_channels: int,
rnn_hidden_size: int = 100,
rnn_num_layers: int = 1,
bidirectional: bool = False,
rnn_dropout: float = 0,
freeze: bool = True,
saved_glove_file: Optional[Path] = None,
) -> LinearCnnRnn:
"""Return LinearCnnRnn with embedding layer initialised with GloVe."""
glove = load_glove(
input_file=glove_path,
glove_dim=glove_dim,
vocab=vocab,
saved_glove_file=saved_glove_file,
)
embedding = nn.Embedding.from_pretrained(glove, freeze=freeze)
lstm = nn.LSTM(
input_size=out_channels * len(filters),
hidden_size=rnn_hidden_size,
num_layers=rnn_num_layers,
bidirectional=bidirectional,
dropout=rnn_dropout,
batch_first=True,
)
return LinearCnnRnn(
embedding=embedding,
num_classes=num_classes,
rnn=lstm,
filters=filters,
out_channels=out_channels,
)
def glove_simple(
glove_path: Union[Path, TextIO],
glove_dim: int,
num_classes: int,
vocab: Vocabulary,
freeze: bool = True,
saved_glove_file: Optional[Path] = None,
) -> Simple:
"""Return Simple with embedding layer initialised with GloVe."""
glove = load_glove(
input_file=glove_path,
glove_dim=glove_dim,
vocab=vocab,
saved_glove_file=saved_glove_file,
)
embedding = nn.Embedding.from_pretrained(glove, freeze=freeze)
return Simple(
embedding=embedding,
num_classes=num_classes,
)
def glove_linear_lstm(
glove_path: Union[Path, TextIO],
glove_dim: int,
num_classes: int,
vocab: Vocabulary,
freeze: bool = True,
saved_glove_file: Optional[Path] = None,
rnn_hidden_size: int = 100,
rnn_num_layers: int = 1,
bidirectional: bool = False,
rnn_dropout: float = 0,
) -> LinearRnn:
"""
Return RnnClassifier with embedding layer initialised with GloVe.
RNN used is an LSTM.
"""
glove = load_glove(
input_file=glove_path,
glove_dim=glove_dim,
vocab=vocab,
saved_glove_file=saved_glove_file,
)
embedding = nn.Embedding.from_pretrained(glove, freeze=freeze)
lstm = nn.LSTM(
input_size=embedding.embedding_dim,
hidden_size=rnn_hidden_size,
num_layers=rnn_num_layers,
bidirectional=bidirectional,
dropout=rnn_dropout,
batch_first=True,
)
return LinearRnn(
embedding=embedding,
num_classes=num_classes,
rnn=lstm,
)