def __init__(self, data):
super(BiLSTM, self).__init__()
print "build batched bilstm..."
self.gpu = data.HP_gpu
self.use_char = data.HP_use_char
self.batch_size = data.HP_batch_size
self.char_hidden_dim = 0
self.average_batch = data.HP_average_batch_loss
if self.use_char:
self.char_hidden_dim = data.HP_char_hidden_dim
self.char_embedding_dim = data.char_emb_dim
if data.char_features == "CNN":
self.char_feature = CharCNN(data.char_alphabet.size(),
self.char_embedding_dim,
self.char_hidden_dim,
data.HP_dropout, self.gpu)
elif data.char_features == "LSTM":
self.char_feature = CharBiLSTM(data.char_alphabet.size(),
self.char_embedding_dim,
self.char_hidden_dim,
data.HP_dropout, self.gpu)
else:
print "Error char feature selection, please check parameter data.char_features (either CNN or LSTM)."
exit(0)
self.embedding_dim = data.word_emb_dim
self.hidden_dim = data.HP_hidden_dim
self.drop = nn.Dropout(data.HP_dropout)
self.droplstm = nn.Dropout(data.HP_dropout)
self.word_embeddings = nn.Embedding(data.word_alphabet.size(),
self.embedding_dim)
self.bilstm_flag = data.HP_bilstm
self.lstm_layer = data.HP_lstm_layer
if data.pretrain_word_embedding is not None:
self.word_embeddings.weight.data.copy_(
torch.from_numpy(data.pretrain_word_embedding))
else:
self.word_embeddings.weight.data.copy_(
torch.from_numpy(
self.random_embedding(data.word_alphabet.size(),
self.embedding_dim)))
# The LSTM takes word embeddings as inputs, and outputs hidden states
# with dimensionality hidden_dim.
if self.bilstm_flag:
lstm_hidden = data.HP_hidden_dim // 2
else:
lstm_hidden = data.HP_hidden_dim
self.lstm = nn.LSTM(self.embedding_dim + self.char_hidden_dim,
lstm_hidden,
num_layers=self.lstm_layer,
batch_first=True,
bidirectional=self.bilstm_flag)
# The linear layer that maps from hidden state space to tag space
self.hidden2tag = nn.Linear(self.hidden_dim, data.label_alphabet_size)
if self.gpu:
self.drop = self.drop.cuda()
self.droplstm = self.droplstm.cuda()
self.word_embeddings = self.word_embeddings.cuda()
self.lstm = self.lstm.cuda()
self.hidden2tag = self.hidden2tag.cuda()
def __init__(self,
vocab_size,
word_embed_dim,
word_hidden_dim,
alphabet_size,
char_embedding_dim,
char_hidden_dim,
feature_extractor,
tag_num,
dropout,
pretrain_embed=None,
use_char=False,
use_crf=False,
use_gpu=False):
super(NamedEntityRecog, self).__init__()
self.use_crf = use_crf
self.use_char = use_char
self.drop = nn.Dropout(dropout)
self.input_dim = word_embed_dim
self.feature_extractor = feature_extractor
self.embeds = nn.Embedding(vocab_size, word_embed_dim, padding_idx=0)
if pretrain_embed is not None:
self.embeds.weight.data.copy_(torch.from_numpy(pretrain_embed))
else:
self.embeds.weight.data.copy_(
torch.from_numpy(
self.random_embedding(vocab_size, word_embed_dim)))
if self.use_char:
self.input_dim += char_hidden_dim
self.char_feature = CharCNN(alphabet_size, char_embedding_dim,
char_hidden_dim, dropout)
if feature_extractor == 'lstm':
self.lstm = nn.LSTM(self.input_dim,
word_hidden_dim,
batch_first=True,
bidirectional=True)
else:
self.word2cnn = nn.Linear(self.input_dim, word_hidden_dim * 2)
self.cnn_list = list()
for _ in range(4):
self.cnn_list.append(
nn.Conv1d(word_hidden_dim * 2,
word_hidden_dim * 2,
kernel_size=3,
padding=1))
self.cnn_list.append(nn.ReLU())
self.cnn_list.append(nn.Dropout(dropout))
self.cnn_list.append(nn.BatchNorm1d(word_hidden_dim * 2))
self.cnn = nn.Sequential(*self.cnn_list)
if self.use_crf:
self.hidden2tag = nn.Linear(word_hidden_dim * 2, tag_num + 2)
self.crf = CRF(tag_num, use_gpu)
else:
self.hidden2tag = nn.Linear(word_hidden_dim * 2, tag_num)
def __init__(self, args):
super().__init__()
L, R, lb = 7, 10, 0.0
self.wtoi, self.wvec = args.wtoi, args.wvec
qws = list(set(args.qws))
self.w2l, self.w2r = get_lr(self.wtoi.keys(), qws, L, R, lb)
self.cdim = 100
self.charcnn = CharCNN(self.cdim)
self.lW = nn.Parameter(
torch.zeros(self.cdim, self.cdim).uniform_(-0.1, 0.1))
self.rW = nn.Parameter(
torch.zeros(self.cdim, self.cdim).uniform_(-0.1, 0.1))
self.gL = nn.Linear(L + R, 2)
def __init__(self, args, model_params):
super(WordEmbedding, self).__init__()
self.gpu = args.ifgpu
self.use_char = args.useChar
self.char_hidden_dim = args.char_hidden_dim
self.char_embedding_dim = args.char_embedding_dim
self.embedding_dim = model_params.embedding_dim
self.drop = nn.Dropout(args.dropout)
#char Embedding
if self.use_char:
if args.charExtractor == "CNN":
self.char_feature = CharCNN(
model_params.char_alphabet.size(),
model_params.pretrain_char_embedding,
self.char_embedding_dim, self.char_hidden_dim,
args.dropout, self.gpu)
elif args.charExtractor == "LSTM":
self.char_feature = CharBiLSTM(
model_params.char_alphabet.size(),
model_params.pretrain_char_embedding,
self.char_embedding_dim, self.char_hidden_dim,
args.dropout, self.gpu)
else:
print(
"Error char feature selection, please check parameter data.char_feature_extractor (CNN/LSTM)."
)
exit(0)
#word Embedding
self.word_embedding = nn.Embedding(model_params.word_alphabet.size(),
self.embedding_dim)
if model_params.pretrain_word_embedding is not None:
self.word_embedding.weight.data.copy_(
torch.from_numpy(model_params.pretrain_word_embedding))
else:
self.word_embedding.weight.data.copy_(
torch.from_numpy(
self.random_embedding(model_params.word_alphabet.size(),
self.embedding_dim)))
def __init__(self, data, use_position, use_cap, use_postag, use_char):
super(WordRep, self).__init__()
print "build word representation..."
self.gpu = data.HP_gpu
self.use_char = use_char
self.batch_size = data.HP_batch_size
self.char_hidden_dim = 0
self.char_all_feature = False
if self.use_char:
self.char_hidden_dim = data.HP_char_hidden_dim
self.char_embedding_dim = data.char_emb_dim
if data.char_feature_extractor == "CNN":
self.char_feature = CharCNN(data.char_alphabet.size(),
data.pretrain_char_embedding,
self.char_embedding_dim,
self.char_hidden_dim,
data.HP_dropout, self.gpu)
elif data.char_feature_extractor == "LSTM":
self.char_feature = CharBiLSTM(data.char_alphabet.size(),
data.pretrain_char_embedding,
self.char_embedding_dim,
self.char_hidden_dim,
data.HP_dropout, self.gpu)
elif data.char_feature_extractor == "GRU":
self.char_feature = CharBiGRU(data.char_alphabet.size(),
data.pretrain_char_embedding,
self.char_embedding_dim,
self.char_hidden_dim,
data.HP_dropout, self.gpu)
elif data.char_feature_extractor == "ALL":
self.char_all_feature = True
self.char_feature = CharCNN(data.char_alphabet.size(),
data.pretrain_char_embedding,
self.char_embedding_dim,
self.char_hidden_dim,
data.HP_dropout, self.gpu)
self.char_feature_extra = CharBiLSTM(
data.char_alphabet.size(), data.pretrain_char_embedding,
self.char_embedding_dim, self.char_hidden_dim,
data.HP_dropout, self.gpu)
else:
print "Error char feature selection, please check parameter data.char_feature_extractor (CNN/LSTM/GRU/ALL)."
exit(0)
self.embedding_dim = data.word_emb_dim
self.drop = nn.Dropout(data.HP_dropout)
self.word_embedding = nn.Embedding(data.word_alphabet.size(),
self.embedding_dim)
if data.pretrain_word_embedding is not None:
self.word_embedding.weight.data.copy_(
torch.from_numpy(data.pretrain_word_embedding))
else:
self.word_embedding.weight.data.copy_(
torch.from_numpy(
self.random_embedding(data.word_alphabet.size(),
self.embedding_dim)))
self.feature_num = 0
self.feature_embedding_dims = data.feature_emb_dims
self.feature_embeddings = nn.ModuleList()
if use_cap:
self.feature_num += 1
alphabet_id = data.feature_name2id['[Cap]']
emb = nn.Embedding(data.feature_alphabets[alphabet_id].size(),
self.feature_embedding_dims[alphabet_id])
emb.weight.data.copy_(
torch.from_numpy(
self.random_embedding(
data.feature_alphabets[alphabet_id].size(),
self.feature_embedding_dims[alphabet_id])))
self.feature_embeddings.append(emb)
if use_postag:
self.feature_num += 1
alphabet_id = data.feature_name2id['[POS]']
emb = nn.Embedding(data.feature_alphabets[alphabet_id].size(),
self.feature_embedding_dims[alphabet_id])
emb.weight.data.copy_(
torch.from_numpy(
self.random_embedding(
data.feature_alphabets[alphabet_id].size(),
self.feature_embedding_dims[alphabet_id])))
self.feature_embeddings.append(emb)
self.use_position = use_position
if self.use_position:
position_alphabet_id = data.re_feature_name2id['[POSITION]']
self.position_embedding_dim = data.re_feature_emb_dims[
position_alphabet_id]
self.position1_emb = nn.Embedding(
data.re_feature_alphabet_sizes[position_alphabet_id],
self.position_embedding_dim, data.pad_idx)
self.position1_emb.weight.data.copy_(
torch.from_numpy(
self.random_embedding(
data.re_feature_alphabet_sizes[position_alphabet_id],
self.position_embedding_dim)))
self.position2_emb = nn.Embedding(
data.re_feature_alphabet_sizes[position_alphabet_id],
self.position_embedding_dim, data.pad_idx)
self.position2_emb.weight.data.copy_(
torch.from_numpy(
self.random_embedding(
data.re_feature_alphabet_sizes[position_alphabet_id],
self.position_embedding_dim)))
if torch.cuda.is_available():
self.drop = self.drop.cuda(self.gpu)
self.word_embedding = self.word_embedding.cuda(self.gpu)
for idx in range(self.feature_num):
self.feature_embeddings[idx] = self.feature_embeddings[
idx].cuda(self.gpu)
if self.use_position:
self.position1_emb = self.position1_emb.cuda(self.gpu)
self.position2_emb = self.position2_emb.cuda(self.gpu)
class WordRep(nn.Module):
def __init__(self, data, use_position, use_cap, use_postag, use_char):
super(WordRep, self).__init__()
print "build word representation..."
self.gpu = data.HP_gpu
self.use_char = use_char
self.batch_size = data.HP_batch_size
self.char_hidden_dim = 0
self.char_all_feature = False
if self.use_char:
self.char_hidden_dim = data.HP_char_hidden_dim
self.char_embedding_dim = data.char_emb_dim
if data.char_feature_extractor == "CNN":
self.char_feature = CharCNN(data.char_alphabet.size(),
data.pretrain_char_embedding,
self.char_embedding_dim,
self.char_hidden_dim,
data.HP_dropout, self.gpu)
elif data.char_feature_extractor == "LSTM":
self.char_feature = CharBiLSTM(data.char_alphabet.size(),
data.pretrain_char_embedding,
self.char_embedding_dim,
self.char_hidden_dim,
data.HP_dropout, self.gpu)
elif data.char_feature_extractor == "GRU":
self.char_feature = CharBiGRU(data.char_alphabet.size(),
data.pretrain_char_embedding,
self.char_embedding_dim,
self.char_hidden_dim,
data.HP_dropout, self.gpu)
elif data.char_feature_extractor == "ALL":
self.char_all_feature = True
self.char_feature = CharCNN(data.char_alphabet.size(),
data.pretrain_char_embedding,
self.char_embedding_dim,
self.char_hidden_dim,
data.HP_dropout, self.gpu)
self.char_feature_extra = CharBiLSTM(
data.char_alphabet.size(), data.pretrain_char_embedding,
self.char_embedding_dim, self.char_hidden_dim,
data.HP_dropout, self.gpu)
else:
print "Error char feature selection, please check parameter data.char_feature_extractor (CNN/LSTM/GRU/ALL)."
exit(0)
self.embedding_dim = data.word_emb_dim
self.drop = nn.Dropout(data.HP_dropout)
self.word_embedding = nn.Embedding(data.word_alphabet.size(),
self.embedding_dim)
if data.pretrain_word_embedding is not None:
self.word_embedding.weight.data.copy_(
torch.from_numpy(data.pretrain_word_embedding))
else:
self.word_embedding.weight.data.copy_(
torch.from_numpy(
self.random_embedding(data.word_alphabet.size(),
self.embedding_dim)))
self.feature_num = 0
self.feature_embedding_dims = data.feature_emb_dims
self.feature_embeddings = nn.ModuleList()
if use_cap:
self.feature_num += 1
alphabet_id = data.feature_name2id['[Cap]']
emb = nn.Embedding(data.feature_alphabets[alphabet_id].size(),
self.feature_embedding_dims[alphabet_id])
emb.weight.data.copy_(
torch.from_numpy(
self.random_embedding(
data.feature_alphabets[alphabet_id].size(),
self.feature_embedding_dims[alphabet_id])))
self.feature_embeddings.append(emb)
if use_postag:
self.feature_num += 1
alphabet_id = data.feature_name2id['[POS]']
emb = nn.Embedding(data.feature_alphabets[alphabet_id].size(),
self.feature_embedding_dims[alphabet_id])
emb.weight.data.copy_(
torch.from_numpy(
self.random_embedding(
data.feature_alphabets[alphabet_id].size(),
self.feature_embedding_dims[alphabet_id])))
self.feature_embeddings.append(emb)
self.use_position = use_position
if self.use_position:
position_alphabet_id = data.re_feature_name2id['[POSITION]']
self.position_embedding_dim = data.re_feature_emb_dims[
position_alphabet_id]
self.position1_emb = nn.Embedding(
data.re_feature_alphabet_sizes[position_alphabet_id],
self.position_embedding_dim, data.pad_idx)
self.position1_emb.weight.data.copy_(
torch.from_numpy(
self.random_embedding(
data.re_feature_alphabet_sizes[position_alphabet_id],
self.position_embedding_dim)))
self.position2_emb = nn.Embedding(
data.re_feature_alphabet_sizes[position_alphabet_id],
self.position_embedding_dim, data.pad_idx)
self.position2_emb.weight.data.copy_(
torch.from_numpy(
self.random_embedding(
data.re_feature_alphabet_sizes[position_alphabet_id],
self.position_embedding_dim)))
if torch.cuda.is_available():
self.drop = self.drop.cuda(self.gpu)
self.word_embedding = self.word_embedding.cuda(self.gpu)
for idx in range(self.feature_num):
self.feature_embeddings[idx] = self.feature_embeddings[
idx].cuda(self.gpu)
if self.use_position:
self.position1_emb = self.position1_emb.cuda(self.gpu)
self.position2_emb = self.position2_emb.cuda(self.gpu)
def random_embedding(self, vocab_size, embedding_dim):
pretrain_emb = np.empty([vocab_size, embedding_dim])
scale = np.sqrt(3.0 / embedding_dim)
for index in range(vocab_size):
pretrain_emb[index, :] = np.random.uniform(-scale, scale,
[1, embedding_dim])
return pretrain_emb
def forward(self, word_inputs, feature_inputs, word_seq_lengths,
char_inputs, char_seq_lengths, char_seq_recover,
position1_inputs, position2_inputs):
"""
input:
word_inputs: (batch_size, sent_len)
features: list [(batch_size, sent_len), (batch_len, sent_len),...]
word_seq_lengths: list of batch_size, (batch_size,1)
char_inputs: (batch_size*sent_len, word_length)
char_seq_lengths: list of whole batch_size for char, (batch_size*sent_len, 1)
char_seq_recover: variable which records the char order information, used to recover char order
output:
Variable(batch_size, sent_len, hidden_dim)
"""
batch_size = word_inputs.size(0)
sent_len = word_inputs.size(1)
word_embs = self.word_embedding(word_inputs)
word_list = [word_embs]
for idx in range(self.feature_num):
word_list.append(self.feature_embeddings[idx](feature_inputs[idx]))
if self.use_char:
## calculate char lstm last hidden
char_features = self.char_feature.get_last_hiddens(
char_inputs,
char_seq_lengths.cpu().numpy())
char_features = char_features[char_seq_recover]
char_features = char_features.view(batch_size, sent_len, -1)
## concat word and char together
word_list.append(char_features)
word_embs = torch.cat([word_embs, char_features], 2)
if self.char_all_feature:
char_features_extra = self.char_feature_extra.get_last_hiddens(
char_inputs,
char_seq_lengths.cpu().numpy())
char_features_extra = char_features_extra[char_seq_recover]
char_features_extra = char_features_extra.view(
batch_size, sent_len, -1)
## concat word and char together
word_list.append(char_features_extra)
if self.use_position:
position1_feature = self.position1_emb(position1_inputs)
position2_feature = self.position2_emb(position2_inputs)
word_list.append(position1_feature)
word_list.append(position2_feature)
word_embs = torch.cat(word_list, 2)
word_represent = self.drop(word_embs)
return word_represent
def __init__(self, data):
super(WordRep, self).__init__()
print "build word representation..."
self.gpu = data.HP_gpu
self.use_char = data.use_char
self.batch_size = data.HP_batch_size
self.char_hidden_dim = 0
self.char_all_feature = False
if self.use_char:
self.char_hidden_dim = data.HP_char_hidden_dim
self.char_embedding_dim = data.char_emb_dim
if data.char_seq_feature == "CNN":
self.char_feature = CharCNN(data.char_alphabet.size(),
self.char_embedding_dim,
self.char_hidden_dim,
data.HP_dropout, self.gpu)
elif data.char_seq_feature == "LSTM":
self.char_feature = CharBiLSTM(data.char_alphabet.size(),
self.char_embedding_dim,
self.char_hidden_dim,
data.HP_dropout, self.gpu)
elif data.char_seq_feature == "GRU":
self.char_feature = CharBiGRU(data.char_alphabet.size(),
self.char_embedding_dim,
self.char_hidden_dim,
data.HP_dropout, self.gpu)
elif data.char_seq_feature == "ALL":
self.char_all_feature = True
self.char_feature = CharCNN(data.char_alphabet.size(),
self.char_embedding_dim,
self.char_hidden_dim,
data.HP_dropout, self.gpu)
self.char_feature_extra = CharBiLSTM(data.char_alphabet.size(),
self.char_embedding_dim,
self.char_hidden_dim,
data.HP_dropout, self.gpu)
else:
print "Error char feature selection, please check parameter data.char_seq_feature (CNN/LSTM/GRU/ALL)."
exit(0)
self.embedding_dim = data.word_emb_dim
self.drop = nn.Dropout(data.HP_dropout)
self.word_embedding = nn.Embedding(data.word_alphabet.size(),
self.embedding_dim)
if data.pretrain_word_embedding is not None:
self.word_embedding.weight.data.copy_(
torch.from_numpy(data.pretrain_word_embedding))
else:
self.word_embedding.weight.data.copy_(
torch.from_numpy(
self.random_embedding(data.word_alphabet.size(),
self.embedding_dim)))
self.feature_num = data.feature_num
self.feature_embedding_dims = data.feature_emb_dims
self.feature_embeddings = nn.ModuleList()
for idx in range(self.feature_num):
self.feature_embeddings.append(
nn.Embedding(data.feature_alphabets[idx].size(),
self.feature_embedding_dims[idx]))
for idx in range(self.feature_num):
if data.pretrain_feature_embeddings[idx] is not None:
self.feature_embeddings[idx].weight.data.copy_(
torch.from_numpy(data.pretrain_feature_embeddings[idx]))
else:
self.feature_embeddings[idx].weight.data.copy_(
torch.from_numpy(
self.random_embedding(
data.feature_alphabets[idx].size(),
self.feature_embedding_dims[idx])))
if self.gpu:
self.drop = self.drop.cuda()
self.word_embedding = self.word_embedding.cuda()
for idx in range(self.feature_num):
self.feature_embeddings[idx] = self.feature_embeddings[
idx].cuda()
def __init__(self, data, opt):
super(WordRep, self).__init__()
self.gpu = opt.gpu
self.batch_size = opt.batch_size
self.use_elmo = False
if opt.elmo:
logging.info("use elmo, loading ...")
self.use_elmo = True
self.elmo = Embedder(data.config['elmo_path'])
# we project the elmo representation to the same dim of char embedding
self.elmo_projection = nn.Linear(
self.elmo.config['encoder']['projection_dim'] * 2,
opt.char_hidden_dim, False)
self.elmo_drop = nn.Dropout(opt.dropout)
else:
self.char_hidden_dim = opt.char_hidden_dim
self.char_embedding_dim = opt.char_emb_dim
self.char_feature = CharCNN(data.char_alphabet.size(), None,
self.char_embedding_dim,
self.char_hidden_dim, opt.dropout,
self.gpu)
self.embedding_dim = data.word_emb_dim
self.drop = nn.Dropout(opt.dropout)
self.word_embedding = nn.Embedding(data.word_alphabet.size(),
self.embedding_dim)
if data.pretrain_word_embedding is not None:
self.word_embedding.weight.data.copy_(
torch.from_numpy(data.pretrain_word_embedding))
else:
self.word_embedding.weight.data.copy_(
torch.from_numpy(
self.random_embedding(data.word_alphabet.size(),
self.embedding_dim)))
if data.feat_config is not None:
self.feature_num = len(data.feature_alphabets)
self.feature_embedding_dims = data.feature_emb_dims
self.feature_embeddings = nn.ModuleList()
for idx in range(self.feature_num):
emb = nn.Embedding(data.feature_alphabets[idx].size(),
self.feature_embedding_dims[idx])
emb.weight.data.copy_(
torch.from_numpy(
self.random_embedding(
data.feature_alphabets[idx].size(),
self.feature_embedding_dims[idx])))
self.feature_embeddings.append(emb)
else:
self.feature_num = 0
if opt.gpu >= 0 and torch.cuda.is_available():
self.drop = self.drop.cuda(self.gpu)
self.word_embedding = self.word_embedding.cuda(self.gpu)
if data.feat_config is not None:
for idx in range(self.feature_num):
self.feature_embeddings[idx] = self.feature_embeddings[
idx].cuda(self.gpu)
if opt.elmo:
self.elmo_projection = self.elmo_projection.cuda(self.gpu)
self.elmo_drop = self.elmo_drop.cuda(self.gpu)
class WordRep(nn.Module):
def __init__(self, data, opt):
super(WordRep, self).__init__()
self.gpu = opt.gpu
self.batch_size = opt.batch_size
self.use_elmo = False
if opt.elmo:
logging.info("use elmo, loading ...")
self.use_elmo = True
self.elmo = Embedder(data.config['elmo_path'])
# we project the elmo representation to the same dim of char embedding
self.elmo_projection = nn.Linear(
self.elmo.config['encoder']['projection_dim'] * 2,
opt.char_hidden_dim, False)
self.elmo_drop = nn.Dropout(opt.dropout)
else:
self.char_hidden_dim = opt.char_hidden_dim
self.char_embedding_dim = opt.char_emb_dim
self.char_feature = CharCNN(data.char_alphabet.size(), None,
self.char_embedding_dim,
self.char_hidden_dim, opt.dropout,
self.gpu)
self.embedding_dim = data.word_emb_dim
self.drop = nn.Dropout(opt.dropout)
self.word_embedding = nn.Embedding(data.word_alphabet.size(),
self.embedding_dim)
if data.pretrain_word_embedding is not None:
self.word_embedding.weight.data.copy_(
torch.from_numpy(data.pretrain_word_embedding))
else:
self.word_embedding.weight.data.copy_(
torch.from_numpy(
self.random_embedding(data.word_alphabet.size(),
self.embedding_dim)))
if data.feat_config is not None:
self.feature_num = len(data.feature_alphabets)
self.feature_embedding_dims = data.feature_emb_dims
self.feature_embeddings = nn.ModuleList()
for idx in range(self.feature_num):
emb = nn.Embedding(data.feature_alphabets[idx].size(),
self.feature_embedding_dims[idx])
emb.weight.data.copy_(
torch.from_numpy(
self.random_embedding(
data.feature_alphabets[idx].size(),
self.feature_embedding_dims[idx])))
self.feature_embeddings.append(emb)
else:
self.feature_num = 0
if opt.gpu >= 0 and torch.cuda.is_available():
self.drop = self.drop.cuda(self.gpu)
self.word_embedding = self.word_embedding.cuda(self.gpu)
if data.feat_config is not None:
for idx in range(self.feature_num):
self.feature_embeddings[idx] = self.feature_embeddings[
idx].cuda(self.gpu)
if opt.elmo:
self.elmo_projection = self.elmo_projection.cuda(self.gpu)
self.elmo_drop = self.elmo_drop.cuda(self.gpu)
def random_embedding(self, vocab_size, embedding_dim):
pretrain_emb = np.zeros([vocab_size, embedding_dim])
scale = np.sqrt(3.0 / embedding_dim)
for index in range(vocab_size):
pretrain_emb[index, :] = np.random.uniform(-scale, scale,
[1, embedding_dim])
return pretrain_emb
def forward(self, word_inputs, word_seq_lengths, char_inputs,
char_seq_lengths, char_seq_recover, feature_inputs,
text_inputs):
"""
input:
word_inputs: (batch_size, sent_len)
features: list [(batch_size, sent_len), (batch_len, sent_len),...]
word_seq_lengths: list of batch_size, (batch_size,1)
char_inputs: (batch_size*sent_len, word_length)
char_seq_lengths: list of whole batch_size for char, (batch_size*sent_len, 1)
char_seq_recover: variable which records the char order information, used to recover char order
output:
Variable(batch_size, sent_len, hidden_dim)
"""
batch_size = word_inputs.size(0)
sent_len = word_inputs.size(1)
word_embs = self.word_embedding(word_inputs)
word_list = [word_embs]
for idx in range(self.feature_num):
word_list.append(self.feature_embeddings[idx](feature_inputs[idx]))
if self.use_elmo:
with torch.no_grad():
elmo_rep = torch.from_numpy(
np.array(self.elmo.sents2elmo(
text_inputs))) # batch, seq_len, 1024
if self.gpu >= 0 and torch.cuda.is_available():
elmo_rep = elmo_rep.cuda(self.gpu)
char_features = self.elmo_drop(self.elmo_projection(elmo_rep))
# char_features = elmo_rep
else:
char_features = self.char_feature.get_last_hiddens(
char_inputs,
char_seq_lengths.cpu().numpy())
char_features = char_features[char_seq_recover]
char_features = char_features.view(batch_size, sent_len, -1)
word_list.append(char_features)
word_embs = torch.cat(word_list, 2)
word_represent = self.drop(word_embs)
return word_represent
def __init__(self, data):
super(WordRep, self).__init__()
print "Build word representation..."
self.gpu = data.HP_gpu
self.use_char = data.use_char
self.use_trans = data.use_trans
self.batch_size = data.HP_batch_size
self.char_hidden_dim = 0
self.char_all_feature = False
self.use_mapping = data.use_mapping
self.mapping_func = data.mapping_func
# character-level
if self.use_trans:
if self.use_mapping:
# linear mapping
self.w = nn.Linear(data.word_emb_dim, data.HP_trans_hidden_dim)
# non-linear mapping:w + tanh or w + sigmoid
if self.mapping_func:
if self.mapping_func == "tanh":
self.non_linear = nn.Tanh()
elif self.mapping_func == "sigmoid":
self.non_linear = nn.Sigmoid()
self.trans_hidden_dim = data.HP_trans_hidden_dim
self.trans_embedding_dim = data.trans_emb_dim
self.trans_feature = TransBiLSTM(data.translation_alphabet.size(), self.trans_embedding_dim,
self.trans_hidden_dim,
data.HP_dropout, data.pretrain_trans_embedding, self.gpu)
# word-level
if self.use_char:
self.char_hidden_dim = data.HP_char_hidden_dim
self.char_embedding_dim = data.char_emb_dim
if data.char_seq_feature == "CNN":
self.char_feature = CharCNN(data.char_alphabet.size(), self.char_embedding_dim, self.char_hidden_dim,
data.HP_dropout, data.pretrain_char_embedding, self.gpu)
elif data.char_seq_feature == "LSTM":
self.char_feature = CharBiLSTM(data.char_alphabet.size(), self.char_embedding_dim, self.char_hidden_dim,
data.HP_dropout, data.pretrain_char_embedding, self.gpu)
elif data.char_seq_feature == "GRU":
self.char_feature = CharBiGRU(data.char_alphabet.size(), self.char_embedding_dim, self.char_hidden_dim,
data.HP_dropout, self.gpu)
elif data.char_seq_feature == "ALL":
self.char_all_feature = True
self.char_feature = CharCNN(data.char_alphabet.size(), self.char_embedding_dim, self.char_hidden_dim,
data.HP_dropout, data.pretrain_char_embedding, self.gpu)
self.char_feature_extra = CharBiLSTM(data.char_alphabet.size(), self.char_embedding_dim,
self.char_hidden_dim, data.HP_dropout, self.gpu)
else:
print "Error char feature selection, please check parameter data.char_seq_feature (CNN/LSTM/GRU/ALL)."
exit(0)
# Word embedding
self.embedding_dim = data.word_emb_dim
self.drop = nn.Dropout(data.HP_dropout)
self.word_embedding = nn.Embedding(data.word_alphabet.size(), self.embedding_dim)
if data.pretrain_word_embedding is not None:
self.word_embedding.weight.data.copy_(torch.from_numpy(data.pretrain_word_embedding))
else:
self.word_embedding.weight.data.copy_(
torch.from_numpy(self.random_embedding(data.word_alphabet.size(), self.embedding_dim)))
# not use
self.feature_num = data.feature_num
self.feature_embedding_dims = data.feature_emb_dims
self.feature_embeddings = nn.ModuleList()
for idx in range(self.feature_num):
self.feature_embeddings.append(
nn.Embedding(data.feature_alphabets[idx].size(), self.feature_embedding_dims[idx]))
for idx in range(self.feature_num):
if data.pretrain_feature_embeddings[idx] is not None:
self.feature_embeddings[idx].weight.data.copy_(torch.from_numpy(data.pretrain_feature_embeddings[idx]))
else:
self.feature_embeddings[idx].weight.data.copy_(torch.from_numpy(
self.random_embedding(data.feature_alphabets[idx].size(), self.feature_embedding_dims[idx])))
if self.gpu:
self.drop = self.drop.cuda()
self.word_embedding = self.word_embedding.cuda()
for idx in range(self.feature_num):
self.feature_embeddings[idx] = self.feature_embeddings[idx].cuda()
class WordRep(nn.Module):
def __init__(self, data):
super(WordRep, self).__init__()
print "Build word representation..."
self.gpu = data.HP_gpu
self.use_char = data.use_char
self.use_trans = data.use_trans
self.batch_size = data.HP_batch_size
self.char_hidden_dim = 0
self.char_all_feature = False
self.use_mapping = data.use_mapping
self.mapping_func = data.mapping_func
# character-level
if self.use_trans:
if self.use_mapping:
# linear mapping
self.w = nn.Linear(data.word_emb_dim, data.HP_trans_hidden_dim)
# non-linear mapping:w + tanh or w + sigmoid
if self.mapping_func:
if self.mapping_func == "tanh":
self.non_linear = nn.Tanh()
elif self.mapping_func == "sigmoid":
self.non_linear = nn.Sigmoid()
self.trans_hidden_dim = data.HP_trans_hidden_dim
self.trans_embedding_dim = data.trans_emb_dim
self.trans_feature = TransBiLSTM(data.translation_alphabet.size(), self.trans_embedding_dim,
self.trans_hidden_dim,
data.HP_dropout, data.pretrain_trans_embedding, self.gpu)
# word-level
if self.use_char:
self.char_hidden_dim = data.HP_char_hidden_dim
self.char_embedding_dim = data.char_emb_dim
if data.char_seq_feature == "CNN":
self.char_feature = CharCNN(data.char_alphabet.size(), self.char_embedding_dim, self.char_hidden_dim,
data.HP_dropout, data.pretrain_char_embedding, self.gpu)
elif data.char_seq_feature == "LSTM":
self.char_feature = CharBiLSTM(data.char_alphabet.size(), self.char_embedding_dim, self.char_hidden_dim,
data.HP_dropout, data.pretrain_char_embedding, self.gpu)
elif data.char_seq_feature == "GRU":
self.char_feature = CharBiGRU(data.char_alphabet.size(), self.char_embedding_dim, self.char_hidden_dim,
data.HP_dropout, self.gpu)
elif data.char_seq_feature == "ALL":
self.char_all_feature = True
self.char_feature = CharCNN(data.char_alphabet.size(), self.char_embedding_dim, self.char_hidden_dim,
data.HP_dropout, data.pretrain_char_embedding, self.gpu)
self.char_feature_extra = CharBiLSTM(data.char_alphabet.size(), self.char_embedding_dim,
self.char_hidden_dim, data.HP_dropout, self.gpu)
else:
print "Error char feature selection, please check parameter data.char_seq_feature (CNN/LSTM/GRU/ALL)."
exit(0)
# Word embedding
self.embedding_dim = data.word_emb_dim
self.drop = nn.Dropout(data.HP_dropout)
self.word_embedding = nn.Embedding(data.word_alphabet.size(), self.embedding_dim)
if data.pretrain_word_embedding is not None:
self.word_embedding.weight.data.copy_(torch.from_numpy(data.pretrain_word_embedding))
else:
self.word_embedding.weight.data.copy_(
torch.from_numpy(self.random_embedding(data.word_alphabet.size(), self.embedding_dim)))
# not use
self.feature_num = data.feature_num
self.feature_embedding_dims = data.feature_emb_dims
self.feature_embeddings = nn.ModuleList()
for idx in range(self.feature_num):
self.feature_embeddings.append(
nn.Embedding(data.feature_alphabets[idx].size(), self.feature_embedding_dims[idx]))
for idx in range(self.feature_num):
if data.pretrain_feature_embeddings[idx] is not None:
self.feature_embeddings[idx].weight.data.copy_(torch.from_numpy(data.pretrain_feature_embeddings[idx]))
else:
self.feature_embeddings[idx].weight.data.copy_(torch.from_numpy(
self.random_embedding(data.feature_alphabets[idx].size(), self.feature_embedding_dims[idx])))
if self.gpu:
self.drop = self.drop.cuda()
self.word_embedding = self.word_embedding.cuda()
for idx in range(self.feature_num):
self.feature_embeddings[idx] = self.feature_embeddings[idx].cuda()
def random_embedding(self, vocab_size, embedding_dim):
pretrain_emb = np.empty([vocab_size, embedding_dim])
scale = np.sqrt(3.0 / embedding_dim)
for index in range(vocab_size):
pretrain_emb[index, :] = np.random.uniform(-scale, scale, [1, embedding_dim])
return pretrain_emb
def forward(self, word_inputs, feature_inputs, word_seq_lengths, char_inputs, char_seq_lengths, char_seq_recover,
trans_inputs, trans_seq_length, trans_seq_recover):
"""
input:
word_inputs: (batch_size, sent_len)
features: list [(batch_size, sent_len), (batch_len, sent_len),...]
word_seq_lengths: list of batch_size, (batch_size,1)
char_inputs: (batch_size*sent_len, word_length)
char_seq_lengths: list of whole batch_size for char, (batch_size*sent_len, 1)
char_seq_recover: variable which records the char order information, used to recover char order
output:
Variable(batch_size, sent_len, hidden_dim)
"""
batch_size = word_inputs.size(0)
sent_len = word_inputs.size(1)
word_embs = self.word_embedding(word_inputs)
word_list = [word_embs]
for idx in range(self.feature_num):
word_list.append(self.feature_embeddings[idx](feature_inputs[idx]))
if self.use_char:
# calculate char lstm last hidden
char_features, _ = self.char_feature.get_last_hiddens(char_inputs, char_seq_lengths.cpu().numpy())
char_features = char_features[char_seq_recover]
char_features = char_features.view(batch_size, sent_len, -1)
word_list.append(char_features)
if self.char_all_feature:
char_features_extra, _ = self.char_feature_extra.get_last_hiddens(char_inputs,
char_seq_lengths.cpu().numpy())
char_features_extra = char_features_extra[char_seq_recover]
char_features_extra = char_features_extra.view(batch_size, sent_len, -1)
# concat word and char together
word_list.append(char_features_extra)
trans_features_wc_temp = None
word_embed_mapping = None
word_embs_temp = word_embs.view(batch_size * sent_len, -1)
if self.use_trans:
trans_features, trans_rnn_length = self.trans_feature.get_last_hiddens(trans_inputs,
trans_seq_length.cpu().numpy())
if self.use_mapping:
trans_features_wc = trans_features
if self.gpu:
trans_features_wc.cuda()
trans_features_wc = trans_features_wc[trans_seq_recover]
trans_inputs = trans_inputs[trans_seq_recover]
for index, line in enumerate(trans_inputs):
if line[0].data.cpu().numpy()[0] == 0:
if self.mapping_func:
trans_features_wc[index] = self.non_linear(self.w(word_embs_temp[index]))
else:
trans_features_wc[index] = self.w(word_embs_temp[index])
trans_features_wc_temp = trans_features_wc
trans_features_wc = trans_features_wc.view(batch_size, sent_len, -1)
word_list.append(trans_features_wc)
if self.mapping_func:
word_embed_mapping = self.non_linear(self.w(word_embs_temp))
else:
word_embed_mapping = self.w(word_embs_temp)
else:
trans_features = trans_features[trans_seq_recover]
trans_features = trans_features.view(batch_size, sent_len, -1)
word_list.append(trans_features)
word_embs = torch.cat(word_list, 2)
word_represent = self.drop(word_embs)
return word_represent, word_embed_mapping, trans_features_wc_temp
def __init__(self, data, use_position, use_cap, use_postag, use_char):
super(WordRep, self).__init__()
self.gpu = data.HP_gpu
self.use_char = use_char
self.batch_size = data.HP_batch_size
self.char_hidden_dim = 0
if self.use_char:
self.char_hidden_dim = data.HP_char_hidden_dim
self.char_embedding_dim = data.char_emb_dim
self.char_feature = CharCNN(data.char_alphabet.size(), None,
self.char_embedding_dim,
self.char_hidden_dim, data.HP_dropout,
self.gpu)
self.embedding_dim = data.word_emb_dim
self.drop = nn.Dropout(data.HP_dropout)
self.word_embedding = nn.Embedding(data.word_alphabet.size(),
self.embedding_dim)
if data.pretrain_word_embedding is not None:
self.word_embedding.weight.data.copy_(
torch.from_numpy(data.pretrain_word_embedding))
else:
self.word_embedding.weight.data.copy_(
torch.from_numpy(
self.random_embedding(data.word_alphabet.size(),
self.embedding_dim)))
self.feature_num = 0
self.feature_embedding_dims = data.feature_emb_dims
self.feature_embeddings = nn.ModuleList()
if use_cap:
self.feature_num += 1
alphabet_id = data.feature_name2id['[Cap]']
emb = nn.Embedding(data.feature_alphabets[alphabet_id].size(),
self.feature_embedding_dims[alphabet_id])
emb.weight.data.copy_(
torch.from_numpy(
self.random_embedding(
data.feature_alphabets[alphabet_id].size(),
self.feature_embedding_dims[alphabet_id])))
self.feature_embeddings.append(emb)
if use_postag:
self.feature_num += 1
alphabet_id = data.feature_name2id['[POS]']
emb = nn.Embedding(data.feature_alphabets[alphabet_id].size(),
self.feature_embedding_dims[alphabet_id])
emb.weight.data.copy_(
torch.from_numpy(
self.random_embedding(
data.feature_alphabets[alphabet_id].size(),
self.feature_embedding_dims[alphabet_id])))
self.feature_embeddings.append(emb)
self.use_position = use_position
if self.use_position:
position_alphabet_id = data.re_feature_name2id['[POSITION]']
self.position_embedding_dim = data.re_feature_emb_dims[
position_alphabet_id]
self.position1_emb = nn.Embedding(
data.re_feature_alphabet_sizes[position_alphabet_id],
self.position_embedding_dim, data.pad_idx)
self.position1_emb.weight.data.copy_(
torch.from_numpy(
self.random_embedding(
data.re_feature_alphabet_sizes[position_alphabet_id],
self.position_embedding_dim)))
self.position2_emb = nn.Embedding(
data.re_feature_alphabet_sizes[position_alphabet_id],
self.position_embedding_dim, data.pad_idx)
self.position2_emb.weight.data.copy_(
torch.from_numpy(
self.random_embedding(
data.re_feature_alphabet_sizes[position_alphabet_id],
self.position_embedding_dim)))
if torch.cuda.is_available():
self.drop = self.drop.cuda(self.gpu)
self.word_embedding = self.word_embedding.cuda(self.gpu)
for idx in range(self.feature_num):
self.feature_embeddings[idx] = self.feature_embeddings[
idx].cuda(self.gpu)
if self.use_position:
self.position1_emb = self.position1_emb.cuda(self.gpu)
self.position2_emb = self.position2_emb.cuda(self.gpu)
def __init__(self, name, embedding_size):
self.name = name
self.embedding_size = embedding_size
self.is_cuda_available = torch.cuda.is_available()
self.cnn = CharCNN(16, embedding_size, 3)
self.log("Done")
def __init__(self,
word_sents,
name,
embedding_size,
filler,
force_train=False,
ce_enabled=True,
tf_enabled=True):
# The following parameters help to distinguish models
self.name = name
self.embedding_size = embedding_size
# Filler is a functor which provides embedding vector
# when a word or its substring cannot be found in a dictionary
self.filler = filler
# Character embedding dramatically increases the precision of taggers
# but it's not always necessary
self.ce_enabled = ce_enabled
self.tf_enabled = tf_enabled
# Grammer is used to transform a word into a list of its
# character n-grams
self.grammer = CharGram()
# Used models
self.word_model = None
self.char_model = None
# CUDA flag
self.is_cuda_available = torch.cuda.is_available()
# CharCNN
self.cnn = CharCNN(16, embedding_size, 5)
# Model misses.
# Can be used for debugging and as an aggregate estimation
# of vocabulary completeness
self.num_word_misses = 0
self.num_char_misses = 0
self.num_words = 0
self.num_grams = 0
# Trained models should be stored
word_model_path = "internal/{}_{}_word_model.w2v".format(
name, embedding_size)
# Very
# Important
# Decision
train_word_model = False
if force_train:
train_word_model = True
else:
if not os.path.exists(word_model_path):
train_word_model = True
# Word model
if train_word_model:
# Train model
self.log("Training word model...")
iterator = LowercaseSentences(word_sents)
# iterator = word_sents
self.word_model = gensim.models.Word2Vec(iterator,
size=embedding_size,
sg=1,
workers=4,
min_count=3)
self.word_model.save(word_model_path)
else:
# Load
self.log("Loading existing word model...")
self.word_model = gensim.models.Word2Vec.load(word_model_path)
self.log("Done")
def __init__(self, char_size, output_size, kernel_width, vocab_size):
super(CharCNNTrainer, self).__init__()
self.cnn = CharCNN(char_size, output_size, kernel_width)
self.linear = nn.Linear(output_size, vocab_size)
if torch.cuda.is_available():
self.linear = self.linear.cuda()
