def __init__(self, emb_matrix, config, num_tags):
super(NER_SOFTMAX_CHAR, self).__init__()
embd_vector = torch.from_numpy(emb_matrix['word']).float()
self.word_embeds = nn.Embedding.from_pretrained(embd_vector, freeze=False)
embd_vector = torch.from_numpy(emb_matrix['char']).float()
self.char_embeds = nn.Embedding.from_pretrained(embd_vector, freeze=False)
self.lstm_char = nn.LSTM(self.char_embeds.embedding_dim,
config.char_lstm_dim,
num_layers=1, bidirectional=True, batch_first=True)
input_size = self.word_embeds.embedding_dim + config.char_lstm_dim * 2
self.lstm = nn.LSTM(input_size,
config.word_lstm_dim,
num_layers=1, bidirectional=True, batch_first=True)
self.dropout = nn.Dropout(config.dropout_rate)
self.hidden_layer = nn.Linear(config.word_lstm_dim * 2, config.word_lstm_dim)
self.tanh_layer = torch.nn.Tanh()
self.hidden2tag = nn.Linear(config.word_lstm_dim, num_tags)
self.config = config
model_utils.init_lstm_wt(self.lstm_char)
model_utils.init_lstm_wt(self.lstm)
model_utils.init_linear_wt(self.hidden_layer)
model_utils.init_linear_wt(self.hidden2tag)
def __init__(self, vocab, config):
super(NER_SOFTMAX_CHAR, self).__init__()
word_emb_matrix = get_word_embd(vocab, config)
embd_vector = torch.from_numpy(word_emb_matrix).float()
self.word_embeds = nn.Embedding.from_pretrained(embd_vector,
freeze=False)
self.char_embeds = nn.Embedding(len(vocab.char_to_id),
config.char_embd_dim,
padding_idx=Constants.PAD_ID)
if config.is_caps:
self.caps_embeds = nn.Embedding(vocab.get_caps_cardinality(),
config.caps_embd_dim,
padding_idx=Constants.PAD_ID)
self.lstm_char = nn.LSTM(self.char_embeds.embedding_dim,
config.char_lstm_dim,
num_layers=1,
bidirectional=True,
batch_first=True)
if config.is_caps:
self.lstm = nn.LSTM(self.word_embeds.embedding_dim +
config.char_embd_dim * 2 +
config.caps_embd_dim,
config.word_lstm_dim,
num_layers=1,
bidirectional=True,
batch_first=True)
else:
self.lstm = nn.LSTM(self.word_embeds.embedding_dim +
config.char_embd_dim * 2,
config.word_lstm_dim,
num_layers=1,
bidirectional=True,
batch_first=True)
self.dropout = nn.Dropout(config.dropout_rate)
self.hidden_layer = nn.Linear(config.word_lstm_dim * 2,
config.word_lstm_dim)
self.tanh_layer = torch.nn.Tanh()
self.hidden2tag = nn.Linear(config.word_lstm_dim, len(vocab.id_to_tag))
self.config = config
init_lstm_wt(self.lstm_char)
init_lstm_wt(self.lstm)
init_linear_wt(self.hidden_layer)
init_linear_wt(self.hidden2tag)
self.char_embeds.weight.data.uniform_(-1., 1.)
if config.is_caps:
self.caps_embeds.weight.data.uniform_(-1., 1.)