def __init__(self, config, softmatcher, encoder=None, print_info=True):
super(SoftSequence, self).__init__()
self.config = config
self.device = config.device
self.encoder = SoftEncoder(self.config)
if encoder is not None:
self.encoder = encoder
self.softmatch_encoder = softmatcher.encoder
self.softmatch_attention = softmatcher.attention
self.label_size = config.label_size
self.inferencer = LinearCRF(config, print_info=print_info)
self.hidden2tag = nn.Linear(config.hidden_dim * 2,
self.label_size).to(self.device)
self.w1 = nn.Linear(config.hidden_dim,
config.hidden_dim // 2).to(self.device)
self.w2 = nn.Linear(config.hidden_dim // 2,
config.hidden_dim // 2).to(self.device)
self.attn1 = nn.Linear(config.hidden_dim // 2, 1).to(self.device)
self.attn2 = nn.Linear(config.hidden_dim + config.hidden_dim // 2,
1).to(self.device)
self.attn3 = nn.Linear(config.hidden_dim // 2, 1).to(self.device)
self.applying = Variable(torch.randn(config.hidden_dim,
config.hidden_dim // 2),
requires_grad=True).to(self.device)
self.tanh = nn.Tanh().to(self.device)
self.perturb = nn.Dropout(config.dropout).to(self.device)
def __init__(self, config: Config, print_info: bool = True):
super(NNCRF, self).__init__()
self.device = config.device
self.encoder = BiLSTMEncoder(config, print_info=print_info)
self.inferencer = None
if config.use_crf_layer:
self.inferencer = LinearCRF(config, print_info=print_info)
def __init__(self, cfig):
super(BertCRF, self).__init__(cfig)
#self.device = cfig.device
self.num_labels = len(cfig.label2idx)
self.bert = BertModel(cfig)
self.dropout = nn.Dropout(cfig.hidden_dropout_prob)
self.classifier = nn.Linear(cfig.hidden_size, len(cfig.label2idx))
self.inferencer = LinearCRF(cfig)
self.init_weights()
def __init__(self, config, encoder=None, print_info=True):
super(SoftSequenceNaive, self).__init__()
self.config = config
self.device = config.device
self.encoder = SoftEncoder(self.config)
if encoder is not None:
self.encoder = encoder
self.label_size = config.label_size
self.inferencer = LinearCRF(config, print_info=print_info)
self.hidden2tag = nn.Linear(config.hidden_dim, self.label_size).to(self.device)
def __init__(self, config, encoder=None, print_info=True):
super(SoftSequenceNaive, self).__init__()
self.config = config
self.device = config.device
self.encoder = SoftEncoder(self.config)
self.label_size = config.label_size
self.inferencer = LinearCRF(config, print_info=print_info)
self.hidden2tag = nn.Linear(config.hidden_dim,
self.label_size).to(self.device)
self.dsc_loss = DSCLoss(gamma=2)
self.bert = AutoModel.from_pretrained(self.config.bert_path).to(
self.device)
self.tokenizer = AutoTokenizer.from_pretrained(self.config.bert_path)
class BertCRF(BertPreTrainedModel):
def __init__(self, cfig):
super(BertCRF, self).__init__(cfig)
#self.device = cfig.device
self.num_labels = len(cfig.label2idx)
self.bert = BertModel(cfig)
self.dropout = nn.Dropout(cfig.hidden_dropout_prob)
self.classifier = nn.Linear(cfig.hidden_size, len(cfig.label2idx))
self.inferencer = LinearCRF(cfig)
self.init_weights()
def forward(self, input_ids, input_seq_lens=None, annotation_mask=None, labels=None,
attention_mask=None, token_type_ids=None,
position_ids=None, head_mask=None, add_crf=False):
outputs = self.bert(input_ids, attention_mask=attention_mask,
token_type_ids=token_type_ids, position_ids=position_ids, head_mask=head_mask)
sequence_output = outputs[0]
sequence_output = self.dropout(sequence_output) # (batch_size, seq_length, hidden_size)
logits = self.classifier(sequence_output) # (batch_size, seq_length, num_labels)
if labels is not None:
batch_size = input_ids.size(0)
sent_len = input_ids.size(1) # one batch max seq length
maskTemp = torch.arange(1, sent_len + 1, dtype=torch.long).view(1, sent_len).expand(batch_size, sent_len).to(self.device)
mask = torch.le(maskTemp, input_seq_lens.view(batch_size, 1).expand(batch_size, sent_len)).to(self.device)
unlabed_score, labeled_score = self.inferencer(logits, input_seq_lens, labels, attention_mask)
return unlabed_score - labeled_score
else:
bestScores, decodeIdx = self.inferencer.decode(logits, input_seq_lens, annotation_mask)
return bestScores, decodeIdx
# obsolete
def decode(self, input_ids, input_seq_lens=None, annotation_mask=None, attention_mask=None) -> Tuple[torch.Tensor, torch.Tensor]:
"""
Decode the batch input
:param batchInput:
:return:
"""
features = self.bert(input_ids, attention_mask=attention_mask, token_type_ids=None, position_ids=None, head_mask=None)
features = self.dropout(features) # (batch_size, seq_length, hidden_size)
logits = self.classifier(features) # (batch_size, seq_length, num_labels)
bestScores, decodeIdx = self.inferencer.decode(logits, input_seq_lens, annotation_mask)
return bestScores, decodeIdx
class NNCRF(nn.Module):
def __init__(self, config, print_info: bool = True):
super(NNCRF, self).__init__()
self.device = config.device
self.encoder = BiLSTMEncoder(config, print_info=print_info)
self.inferencer = LinearCRF(config, print_info=print_info)
@overrides
def forward(self, words: torch.Tensor,
word_seq_lens: torch.Tensor,
batch_context_emb: torch.Tensor,
chars: torch.Tensor,
char_seq_lens: torch.Tensor,
annotation_mask : torch.Tensor,
marginals: torch.Tensor,
tags: torch.Tensor) -> torch.Tensor:
"""
Calculate the negative loglikelihood.
:param words: (batch_size x max_seq_len)
:param word_seq_lens: (batch_size)
:param batch_context_emb: (batch_size x max_seq_len x context_emb_size)
:param chars: (batch_size x max_seq_len x max_char_len)
:param char_seq_lens: (batch_size x max_seq_len)
:param tags: (batch_size x max_seq_len)
:return: the loss with shape (batch_size)
"""
lstm_scores = self.encoder(words, word_seq_lens, batch_context_emb, chars, char_seq_lens)
batch_size = words.size(0)
sent_len = words.size(1)
maskTemp = torch.arange(1, sent_len + 1, dtype=torch.long).view(1, sent_len).expand(batch_size, sent_len).to(self.device)
mask = torch.le(maskTemp, word_seq_lens.view(batch_size, 1).expand(batch_size, sent_len)).to(self.device)
unlabed_score, labeled_score = self.inferencer(lstm_scores, word_seq_lens, tags, mask)
return unlabed_score - labeled_score
def decode(self, batchInput: Tuple) -> Tuple[torch.Tensor, torch.Tensor]:
"""
Decode the batch input
:param batchInput:
:return:
"""
wordSeqTensor, wordSeqLengths, batch_context_emb, charSeqTensor, charSeqLengths, annotation_mask, marginals, tagSeqTensor = batchInput
features = self.encoder(wordSeqTensor, wordSeqLengths, batch_context_emb,charSeqTensor,charSeqLengths)
bestScores, decodeIdx = self.inferencer.decode(features, wordSeqLengths, annotation_mask)
return bestScores, decodeIdx
def get_marginal(self, batchInput: Tuple) -> torch.Tensor:
wordSeqTensor, wordSeqLengths, batch_context_emb, charSeqTensor, charSeqLengths, annotation_mask, marginals, tagSeqTensor = batchInput
features = self.encoder(wordSeqTensor, wordSeqLengths, batch_context_emb, charSeqTensor, charSeqLengths)
marginals = self.inferencer.compute_constrained_marginal(features, wordSeqLengths, annotation_mask)
return marginals
class NNCRF(nn.Module):
def __init__(self, config, print_info: bool = True):
super(NNCRF, self).__init__()
self.device = config.device
self.encoder = BiLSTMEncoder(config, print_info=print_info)
self.inferencer = LinearCRF(config, print_info=print_info)
@overrides
def forward(self, sent_emb_tensor: torch.Tensor,
type_id_tensor: torch.Tensor, sent_seq_lens: torch.Tensor,
batch_context_emb: torch.Tensor, chars: torch.Tensor,
char_seq_lens: torch.Tensor,
tags: torch.Tensor) -> torch.Tensor:
"""
Calculate the negative loglikelihood.
:param words: (batch_size x max_seq_len)
:param word_seq_lens: (batch_size)
:param batch_context_emb: (batch_size x max_seq_len x context_emb_size)
:param chars: (batch_size x max_seq_len x max_char_len)
:param char_seq_lens: (batch_size x max_seq_len)
:param tags: (batch_size x max_seq_len)
:return: the total negative log-likelihood loss
"""
# print("sents: ",sents)
lstm_scores = self.encoder(sent_emb_tensor, type_id_tensor,
sent_seq_lens, batch_context_emb, chars,
char_seq_lens)
# lstm_scores = self.encoder(sent_emb_tensor, sent_seq_lens, chars, char_seq_lens)
batch_size = sent_emb_tensor.size(0)
sent_len = sent_emb_tensor.size(1)
maskTemp = torch.arange(1, sent_len + 1, dtype=torch.long).view(
1, sent_len).expand(batch_size, sent_len).to(self.device)
mask = torch.le(
maskTemp,
sent_seq_lens.view(batch_size, 1).expand(batch_size,
sent_len)).to(self.device)
unlabed_score, labeled_score = self.inferencer(lstm_scores,
sent_seq_lens, tags,
mask)
return unlabed_score - labeled_score
def decode(
self, batchInput: Tuple[torch.Tensor, torch.Tensor, torch.Tensor,
torch.Tensor, torch.Tensor, torch.Tensor,
torch.Tensor]
) -> Tuple[torch.Tensor, torch.Tensor]:
"""
Decode the batch input
:param batchInput:
:return:
"""
wordSeqTensor, typeTensor, wordSeqLengths, batch_context_emb, charSeqTensor, charSeqLengths, tagSeqTensor = batchInput
features = self.encoder(wordSeqTensor, typeTensor, wordSeqLengths,
batch_context_emb, charSeqTensor,
charSeqLengths)
bestScores, decodeIdx = self.inferencer.decode(features,
wordSeqLengths)
# print(bestScores, decodeIdx)
return bestScores, decodeIdx
class SoftSequenceNaive(nn.Module):
def __init__(self, config, encoder=None, print_info=True):
super(SoftSequenceNaive, self).__init__()
self.config = config
self.device = config.device
self.encoder = SoftEncoder(self.config)
if encoder is not None:
self.encoder = encoder
self.label_size = config.label_size
self.inferencer = LinearCRF(config, print_info=print_info)
self.hidden2tag = nn.Linear(config.hidden_dim, self.label_size).to(self.device)
def forward(self, word_seq_tensor: torch.Tensor,
word_seq_lens: torch.Tensor,
batch_context_emb: torch.Tensor,
char_inputs: torch.Tensor,
char_seq_lens: torch.Tensor, tags):
batch_size = word_seq_tensor.size(0)
max_sent_len = word_seq_tensor.size(1)
output, sentence_mask, _, _ = \
self.encoder(word_seq_tensor, word_seq_lens, batch_context_emb, char_inputs, char_seq_lens, None)
lstm_scores = self.hidden2tag(output)
maskTemp = torch.arange(1, max_sent_len + 1, dtype=torch.long).view(1, max_sent_len).expand(batch_size, max_sent_len).to(self.device)
mask = torch.le(maskTemp, word_seq_lens.view(batch_size, 1).expand(batch_size, max_sent_len)).to(self.device)
if self.inferencer is not None:
unlabeled_score, labeled_score = self.inferencer(lstm_scores, word_seq_lens, tags, mask)
sequence_loss = unlabeled_score - labeled_score
else:
sequence_loss = self.compute_nll_loss(lstm_scores, tags, mask, word_seq_lens)
return sequence_loss
def decode(self, word_seq_tensor: torch.Tensor,
word_seq_lens: torch.Tensor,
batch_context_emb: torch.Tensor,
char_inputs: torch.Tensor,
char_seq_lens: torch.Tensor):
soft_output, soft_sentence_mask, _, _ = \
self.encoder(word_seq_tensor, word_seq_lens, batch_context_emb, char_inputs, char_seq_lens, None)
lstm_scores = self.hidden2tag(soft_output)
if self.inferencer is not None:
bestScores, decodeIdx = self.inferencer.decode(lstm_scores, word_seq_lens, None)
return bestScores, decodeIdx
class NNCRF(nn.Module):
def __init__(self, config, print_info: bool = True):
super(NNCRF, self).__init__()
self.device = config.device
self.encoder = BiLSTMEncoder(config, print_info=print_info)
self.inferencer = LinearCRF(config, print_info=print_info)
@overrides
def forward(self, words: torch.Tensor, word_seq_lens: torch.Tensor,
batch_context_emb: torch.Tensor, chars: torch.Tensor,
char_seq_lens: torch.Tensor,
label_mask_tensor: torch.Tensor) -> torch.Tensor:
"""
Calculate the negative loglikelihood.
:param words: (batch_size x max_seq_len)
:param word_seq_lens: (batch_size)
:param batch_context_emb: (batch_size x max_seq_len x context_emb_size)
:param chars: (batch_size x max_seq_len x max_char_len)
:param char_seq_lens: (batch_size x max_seq_len)
:param label_mask_tensor: (batch_size x max_seq_len x num_labels)
:return: the loss with shape (batch_size)
"""
lstm_scores = self.encoder(words, word_seq_lens, batch_context_emb,
chars, char_seq_lens)
unlabed_score, labeled_score = self.inferencer(lstm_scores,
word_seq_lens,
label_mask_tensor)
return unlabed_score - labeled_score
def decode(
self, batchInput: Tuple[torch.Tensor, torch.Tensor, torch.Tensor,
torch.Tensor, torch.Tensor, torch.Tensor]
) -> Tuple[torch.Tensor, torch.Tensor]:
"""
Decode the batch input
:param batchInput:
:return:
"""
wordSeqTensor, wordSeqLengths, batch_context_emb, charSeqTensor, charSeqLengths, tagSeqTensor = batchInput
features = self.encoder(wordSeqTensor, wordSeqLengths,
batch_context_emb, charSeqTensor,
charSeqLengths)
bestScores, decodeIdx = self.inferencer.decode(features,
wordSeqLengths)
return bestScores, decodeIdx
class NNCRF(nn.Module):
def __init__(self, config: Config,
print_info: bool = True):
super(NNCRF, self).__init__()
self.device = config.device
self.encoder = BiLSTMEncoder(config, print_info=print_info)
self.inferencer = None
if config.use_crf_layer:
self.inferencer = LinearCRF(config, print_info=print_info)
@overrides
def forward(self, words: torch.Tensor,
word_seq_lens: torch.Tensor,
batch_context_emb: torch.Tensor,
chars: torch.Tensor,
char_seq_lens: torch.Tensor,
tags: torch.Tensor) -> torch.Tensor:
"""
Calculate the negative loglikelihood.
:param words: (batch_size x max_seq_len)
:param word_seq_lens: (batch_size)
:param batch_context_emb: (batch_size x max_seq_len x context_emb_size)
:param chars: (batch_size x max_seq_len x max_char_len)
:param char_seq_lens: (batch_size x max_seq_len)
:param tags: (batch_size x max_seq_len)
:return: the loss with shape (batch_size)
"""
batch_size = words.size(0)
max_sent_len = words.size(1)
#Shape: (batch_size, max_seq_len, num_labels)
lstm_scores = self.encoder(words, word_seq_lens, batch_context_emb, chars, char_seq_lens)
maskTemp = torch.arange(1, max_sent_len + 1, dtype=torch.long).view(1, max_sent_len).expand(batch_size, max_sent_len).to(self.device)
mask = torch.le(maskTemp, word_seq_lens.view(batch_size, 1).expand(batch_size, max_sent_len)).to(self.device)
if self.inferencer is not None:
unlabed_score, labeled_score = self.inferencer(lstm_scores, word_seq_lens, tags, mask)
loss = unlabed_score - labeled_score
else:
loss = self.compute_nll_loss(lstm_scores, tags, mask, word_seq_lens)
return loss
def decode(self, batchInput: Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]) -> Tuple[torch.Tensor, torch.Tensor]:
"""
Decode the batch input
:param batchInput:
:return:
"""
wordSeqTensor, wordSeqLengths, batch_context_emb, charSeqTensor, charSeqLengths, tagSeqTensor = batchInput
lstm_scores = self.encoder(wordSeqTensor, wordSeqLengths, batch_context_emb,charSeqTensor,charSeqLengths)
if self.inferencer is not None:
bestScores, decodeIdx = self.inferencer.decode(lstm_scores, wordSeqLengths)
else:
bestScores, decodeIdx = torch.max(lstm_scores, dim=2)
return bestScores, decodeIdx
def compute_nll_loss(self, candidate_scores, target, mask, word_seq_lens):
"""
Directly compute the loss right after the linear layer instead of CRF layer.
Partially taken from `masked_cross_entropy.py` (https://gist.github.com/jihunchoi/f1434a77df9db1bb337417854b398df1)
:param candidate_scores:
:param target:
:param mask:
:param word_seq_lens:
:return:
"""
# logits_flat: (batch * max_len, num_classes)
logits_flat = candidate_scores.view(-1, candidate_scores.size(-1))
# log_probs_flat: (batch * max_len, num_classes)
log_probs_flat = torch.log_softmax(logits_flat, dim=1)
# target_flat: (batch * max_len, 1)
target_flat = target.view(-1, 1)
# losses_flat: (batch * max_len, 1)
losses_flat = -torch.gather(log_probs_flat, dim=1, index=target_flat)
# losses: (batch, max_len)
losses = losses_flat.view(*target.size())
# # mask: (batch, max_len)
# mask = _sequence_mask(sequence_length=length, max_len=target.size(1))
losses = losses * mask.float()
# loss = losses.sum() / word_seq_lens.float().sum()
loss = losses.sum()
return loss
class SoftSequence(nn.Module):
def __init__(self, config, softmatcher, encoder=None, print_info=True):
super(SoftSequence, self).__init__()
self.config = config
self.device = config.device
self.encoder = SoftEncoder(self.config)
if encoder is not None:
self.encoder = encoder
self.softmatch_encoder = softmatcher.encoder
self.softmatch_attention = softmatcher.attention
self.label_size = config.label_size
self.inferencer = LinearCRF(config, print_info=print_info)
self.hidden2tag = nn.Linear(config.hidden_dim * 2,
self.label_size).to(self.device)
self.w1 = nn.Linear(config.hidden_dim,
config.hidden_dim // 2).to(self.device)
self.w2 = nn.Linear(config.hidden_dim // 2,
config.hidden_dim // 2).to(self.device)
self.attn1 = nn.Linear(config.hidden_dim // 2, 1).to(self.device)
self.attn2 = nn.Linear(config.hidden_dim + config.hidden_dim // 2,
1).to(self.device)
self.attn3 = nn.Linear(config.hidden_dim // 2, 1).to(self.device)
self.applying = Variable(torch.randn(config.hidden_dim,
config.hidden_dim // 2),
requires_grad=True).to(self.device)
self.tanh = nn.Tanh().to(self.device)
self.perturb = nn.Dropout(config.dropout).to(self.device)
def forward(self, word_seq_tensor: torch.Tensor,
word_seq_lens: torch.Tensor, batch_context_emb: torch.Tensor,
char_inputs: torch.Tensor, char_seq_lens: torch.Tensor,
trigger_position, tags):
batch_size = word_seq_tensor.size(0)
max_sent_len = word_seq_tensor.size(1)
output, sentence_mask, trigger_vec, trigger_mask = \
self.encoder(word_seq_tensor, word_seq_lens, batch_context_emb, char_inputs, char_seq_lens,
trigger_position)
if trigger_vec is not None:
trig_rep, sentence_vec_cat, trigger_vec_cat = self.softmatch_attention(
output, sentence_mask, trigger_vec, trigger_mask)
# attention
weights = []
for i in range(len(output)):
trig_applied = self.tanh(
self.w1(output[i].unsqueeze(0)) +
self.w2(trig_rep[i].unsqueeze(0).unsqueeze(0)))
x = self.attn1(trig_applied) #63,1
x = torch.mul(x.squeeze(0), sentence_mask[i].unsqueeze(1))
x[x == 0] = float('-inf')
weights.append(x)
normalized_weights = F.softmax(torch.stack(weights), 1)
attn_applied1 = torch.mul(
normalized_weights.repeat(1, 1, output.size(2)), output)
else:
weights = []
for i in range(len(output)):
trig_applied = self.tanh(
self.w1(output[i].unsqueeze(0)) +
self.w1(output[i].unsqueeze(0)))
x = self.attn1(trig_applied) # 63,1
x = torch.mul(x.squeeze(0), sentence_mask[i].unsqueeze(1))
x[x == 0] = float('-inf')
weights.append(x)
normalized_weights = F.softmax(torch.stack(weights), 1)
attn_applied1 = torch.mul(
normalized_weights.repeat(1, 1, output.size(2)), output)
output = torch.cat([output, attn_applied1], dim=2)
lstm_scores = self.hidden2tag(output)
maskTemp = torch.arange(1, max_sent_len + 1, dtype=torch.long).view(
1, max_sent_len).expand(batch_size, max_sent_len).to(self.device)
mask = torch.le(
maskTemp,
word_seq_lens.view(batch_size,
1).expand(batch_size,
max_sent_len)).to(self.device)
if self.inferencer is not None:
unlabeled_score, labeled_score = self.inferencer(
lstm_scores, word_seq_lens, tags, mask)
sequence_loss = unlabeled_score - labeled_score
else:
sequence_loss = self.compute_nll_loss(lstm_scores, tags, mask,
word_seq_lens)
return sequence_loss
def decode_top(self, word_seq_tensor: torch.Tensor,
word_seq_lens: torch.Tensor,
batch_context_emb: torch.Tensor, char_inputs: torch.Tensor,
char_seq_lens: torch.Tensor, trig_rep):
output, sentence_mask, _, _ = \
self.encoder(word_seq_tensor, word_seq_lens, batch_context_emb, char_inputs, char_seq_lens, None)
soft_output, soft_sentence_mask, _, _ = \
self.softmatch_encoder(word_seq_tensor, word_seq_lens, batch_context_emb, char_inputs, char_seq_lens, None)
soft_sent_rep = self.softmatch_attention.attention(
soft_output, soft_sentence_mask)
trig_vec = trig_rep[0]
trig_key = trig_rep[1]
n = soft_sent_rep.size(0)
m = trig_vec.size(0)
d = soft_sent_rep.size(1)
soft_sent_rep_dist = soft_sent_rep.unsqueeze(1).expand(n, m, d)
trig_vec_dist = trig_vec.unsqueeze(0).expand(n, m, d)
dist = torch.pow(soft_sent_rep_dist - trig_vec_dist, 2).sum(2).sqrt()
dvalue, dindices = torch.min(dist, dim=1)
trigger_list = []
for i in dindices.tolist():
trigger_list.append(trig_vec[i])
trig_rep = torch.stack(trigger_list)
# attention
weights = []
for i in range(len(output)):
trig_applied = self.tanh(
self.w1(output[i].unsqueeze(0)) +
self.w2(trig_rep[i].unsqueeze(0).unsqueeze(0)))
x = self.attn1(trig_applied)
x = torch.mul(x.squeeze(0), sentence_mask[i].unsqueeze(1))
x[x == 0] = float('-inf')
weights.append(x)
normalized_weights = F.softmax(torch.stack(weights), 1)
attn_applied1 = torch.mul(
normalized_weights.repeat(1, 1, output.size(2)), output)
output = torch.cat([output, attn_applied1], dim=2)
lstm_scores = self.hidden2tag(output)
bestScores, decodeIdx = self.inferencer.decode(lstm_scores,
word_seq_lens, None)
return bestScores, decodeIdx
class SoftSequenceNaive(nn.Module):
def __init__(self, config, encoder=None, print_info=True):
super(SoftSequenceNaive, self).__init__()
self.config = config
self.device = config.device
self.encoder = SoftEncoder(self.config)
self.label_size = config.label_size
self.inferencer = LinearCRF(config, print_info=print_info)
self.hidden2tag = nn.Linear(config.hidden_dim,
self.label_size).to(self.device)
self.dsc_loss = DSCLoss(gamma=2)
self.bert = AutoModel.from_pretrained(self.config.bert_path).to(
self.device)
self.tokenizer = AutoTokenizer.from_pretrained(self.config.bert_path)
def forward(self, word_seq_tensor: torch.Tensor,
word_seq_lens: torch.Tensor, batch_context_emb: torch.Tensor,
char_inputs: torch.Tensor, char_seq_lens: torch.Tensor, tags,
one_batch_insts):
word_seq_tensor, word_seq_lens = self.load_bert_embedding(
one_batch_insts)
batch_size = word_seq_tensor.size(0)
max_sent_len = word_seq_tensor.size(1)
output, sentence_mask = self.encoder(word_seq_tensor, word_seq_lens,
batch_context_emb, char_inputs,
char_seq_lens, one_batch_insts)
lstm_scores = self.hidden2tag(output)
maskTemp = torch.arange(1, max_sent_len + 1, dtype=torch.long).view(1, max_sent_len)\
.expand(batch_size, max_sent_len).to(self.device)
mask = torch.le(
maskTemp,
word_seq_lens.view(batch_size,
1).expand(batch_size,
max_sent_len)).to(self.device)
unlabeled_score, labeled_score = self.inferencer(
lstm_scores, word_seq_lens, tags, mask)
sequence_loss = unlabeled_score - labeled_score
return sequence_loss
def decode(self, word_seq_tensor: torch.Tensor,
word_seq_lens: torch.Tensor, batch_context_emb: torch.Tensor,
char_inputs: torch.Tensor, char_seq_lens: torch.Tensor,
one_batch_insts):
soft_output, soft_sentence_mask = \
self.encoder(word_seq_tensor, word_seq_lens, batch_context_emb, char_inputs, char_seq_lens, one_batch_insts)
lstm_scores = self.hidden2tag(soft_output)
bestScores, decodeIdx = self.inferencer.decode(lstm_scores,
word_seq_lens, None)
return bestScores, decodeIdx
def load_bert_embedding(self, insts):
# sentence_list = []
for sent in insts:
# sentence = " ".join(str(w) for w in sent.input.words)
# sentence_list.append(sentence)
words = sent.input.words
sent.word_ids = self.tokenizer.convert_tokens_to_ids(words)
# sentence_list = tuple(sentence_list)
# bert_embedding = self.get_bert_embedding(sentence_list)
batch_size = len(insts)
batch_data = insts
# 统计这批数据的序列长度
word_seq_len = torch.LongTensor(
list(map(lambda inst: len(inst.input.words), batch_data)))
max_seq_len = word_seq_len.max()
word_seq_tensor = torch.zeros((batch_size, max_seq_len),
dtype=torch.long)
for idx in range(batch_size):
word_seq_tensor[idx, :word_seq_len[idx]] = torch.LongTensor(
batch_data[idx].word_ids)
word_seq_tensor = word_seq_tensor.to(self.device)
word_seq_len = word_seq_len.to(self.device)
return word_seq_tensor, word_seq_len
def get_bert_embedding(self, batch):
final_dataset = []
for sentence in batch:
tokenized_sentence = [
"[CLS]"
] + self.tokenizer.tokenize(sentence) + ["[SEP]"]
# pooling operation (BERT - first)
isSubword = False
firstSubwordList = []
for t_id, token in enumerate(tokenized_sentence):
if token.startswith("#") == False:
isSubword = False
firstSubwordList.append(t_id)
if isSubword:
continue
if token.startswith("#"):
isSubword = True
input_ids = torch.tensor(
self.tokenizer.convert_tokens_to_ids(tokenized_sentence))
final_dataset.append(input_ids)
word_seq_lens = torch.LongTensor(
list(map(lambda inst: inst.size(), final_dataset))).reshape(-1)
# print(word_seq_lens)
max_seq_len = word_seq_lens.max()
word_seq_tensor = torch.zeros((self.config.batch_size, max_seq_len),
dtype=torch.long)
for idx in range(len(final_dataset)):
tmp = torch.LongTensor(final_dataset[idx])
word_seq_tensor[idx, :word_seq_lens[idx]] = tmp
# embeddings = embeddings[0][0]
# size0 = len(final_dataset)
# final_dataset = torch.cat(final_dataset, dim=0).view(size0, -1, 768)
word_seq_tensor = word_seq_tensor.to(self.device)
word_seq_lens = word_seq_lens.to(self.device)
return word_seq_tensor, word_seq_lens