def __init__(
self,
dictionary,
word_dropout_params=None,
embed_dim=512,
freeze_embed=False,
hidden_dim=512,
num_layers=1,
cell_type="lstm",
dropout_in=0.1,
dropout_out=0.1,
residual_level=None,
bidirectional=False,
add_encoder_output_as_decoder_input=False,
char_rnn_params=None,
):
super().__init__(dictionary)
self.dictionary = dictionary
self.dropout_in = dropout_in
self.dropout_out = dropout_out
self.residual_level = residual_level
self.hidden_dim = hidden_dim
self.bidirectional = bidirectional
self.add_encoder_output_as_decoder_input = add_encoder_output_as_decoder_input
num_embeddings = len(dictionary)
self.padding_idx = dictionary.pad()
self.char_rnn_params = char_rnn_params
if self.char_rnn_params is not None:
self.char_rnn_encoder = char_rnn_encoder.CharRNN(
dictionary=self.dictionary,
embed_dim=embed_dim,
hidden_dim=char_rnn_params["char_rnn_units"],
num_layers=char_rnn_params["char_rnn_layers"],
bidirectional=True,
word_delimiter=char_rnn_params["word_delimiter"],
)
self.word_dim = char_rnn_params["char_rnn_units"]
else:
self.embed_tokens = Embedding(
num_embeddings=num_embeddings,
embedding_dim=embed_dim,
padding_idx=self.padding_idx,
freeze_embed=freeze_embed,
)
self.word_dim = embed_dim
self.cell_type = cell_type
self.layers = nn.ModuleList([])
for layer in range(num_layers):
self.layers.append(
RNNLayer(
self.word_dim if layer == 0 else hidden_dim,
hidden_dim,
self.cell_type,
True if bidirectional and layer == 0 else False,
))
self.num_layers = len(self.layers)
self.word_dropout_module = None
if (word_dropout_params and
word_dropout_params["word_dropout_freq_threshold"] is not None
and word_dropout_params["word_dropout_freq_threshold"] > 0):
self.word_dropout_module = word_dropout.WordDropout(
dictionary, word_dropout_params)
def __init__(
self,
dictionary,
embed_dim=512,
freeze_embed=False,
cell_type="lstm",
hidden_dim=512,
num_layers=1,
dropout_in=0.1,
dropout_out=0.1,
residual_level=None,
bidirectional=False,
word_dropout_params=None,
padding_value=0,
char_rnn_params=None,
):
assert cell_type == "lstm", 'sequence-lstm requires cell_type="lstm"'
super().__init__(dictionary)
self.dictionary = dictionary
self.dropout_in = dropout_in
self.dropout_out = dropout_out
self.residual_level = residual_level
self.hidden_dim = hidden_dim
self.bidirectional = bidirectional
num_embeddings = len(dictionary)
self.padding_idx = dictionary.pad()
self.padding_value = padding_value
self.char_rnn_params = char_rnn_params
if self.char_rnn_params is not None:
self.char_rnn_encoder = char_rnn_encoder.CharRNN(
dictionary=self.dictionary,
embed_dim=embed_dim,
hidden_dim=char_rnn_params["char_rnn_units"],
num_layers=char_rnn_params["char_rnn_layers"],
bidirectional=True,
word_delimiter=char_rnn_params["word_delimiter"],
)
self.word_dim = char_rnn_params["char_rnn_units"]
else:
self.embed_tokens = Embedding(
num_embeddings=num_embeddings,
embedding_dim=embed_dim,
padding_idx=self.padding_idx,
freeze_embed=freeze_embed,
)
self.word_dim = embed_dim
self.layers = nn.ModuleList([])
for layer in range(num_layers):
is_layer_bidirectional = self.bidirectional and layer == 0
self.layers.append(
LSTMSequenceEncoder.LSTM(
self.word_dim if layer == 0 else hidden_dim,
hidden_dim // 2 if is_layer_bidirectional else hidden_dim,
num_layers=1,
dropout=self.dropout_out,
bidirectional=is_layer_bidirectional,
))
self.num_layers = len(self.layers)
self.word_dropout_module = None
if (word_dropout_params and
word_dropout_params["word_dropout_freq_threshold"] is not None
and word_dropout_params["word_dropout_freq_threshold"] > 0):
self.word_dropout_module = word_dropout.WordDropout(
dictionary, word_dropout_params)