File size: 10,134 Bytes
dab2163 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 |
#code adapted form https://github.com/Louis-udm/NER-BERT-CRF/blob/master/NER_BERT_CRF.py
import torch
from transformers import BertModel, BertConfig ##### import these guys -important otherwise config error and you spend an hour figuring out!
from transformers.models.bert.modeling_bert import BertPreTrainedModel
from torch import nn
from torch.nn import CrossEntropyLoss, BCELoss, LayerNorm
from transformers.modeling_outputs import TokenClassifierOutput
# Hack to guarantee backward-compatibility.
BertLayerNorm = LayerNorm
def log_sum_exp_batch(log_Tensor, axis=-1): # shape (batch_size,n,m)
return torch.max(log_Tensor, axis)[0]+torch.log(torch.exp(log_Tensor-torch.max(log_Tensor, axis)[0].view(log_Tensor.shape[0],-1,1)).sum(axis))
class BERT_CRF_NER(BertPreTrainedModel):
_keys_to_ignore_on_load_unexpected = [r"pooler"]
def __init__(self, config):
super().__init__(config)
self.hidden_size = 768
self.start_label_id = config.start_label_id
self.stop_label_id = config.stop_label_id
self.num_labels = config.num_classes
# self.max_seq_length = max_seq_length
self.batch_size = config.batch_size
# use pretrainded BertModel
self.bert = BertModel(config, add_pooling_layer=False)
self.dropout = torch.nn.Dropout(0.2)
# Maps the output of the bert into label space.
self.hidden2label = nn.Linear(self.hidden_size, self.num_labels)
# Matrix of transition parameters. Entry i,j is the score of transitioning *to* i *from* j.
self.transitions = nn.Parameter(
torch.randn(self.num_labels, self.num_labels))
# These two statements enforce the constraint that we never transfer *to* the start tag(or label),
# and we never transfer *from* the stop label (the model would probably learn this anyway,
# so this enforcement is likely unimportant)
self.transitions.data[self.start_label_id, :] = -10000
self.transitions.data[:, self.stop_label_id] = -10000
nn.init.xavier_uniform_(self.hidden2label.weight)
nn.init.constant_(self.hidden2label.bias, 0.0)
# self.apply(self.init_bert_weights)
def init_bert_weights(self, module):
""" Initialize the weights.
"""
if isinstance(module, (nn.Linear, nn.Embedding)):
# Slightly different from the TF version which uses truncated_normal for initialization
# cf https://github.com/pytorch/pytorch/pull/5617
module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
elif isinstance(module, BertLayerNorm):
module.bias.data.zero_()
module.weight.data.fill_(1.0)
if isinstance(module, nn.Linear) and module.bias is not None:
module.bias.data.zero_()
def _forward_alg(self, feats):
"""
this also called alpha-recursion or forward recursion, to calculate log_prob of all barX
"""
# T = self.max_seq_length
T = feats.shape[1]
batch_size = feats.shape[0]
# alpha_recursion,forward, alpha(zt)=p(zt,bar_x_1:t)
log_alpha = torch.Tensor(batch_size, 1, self.num_labels).fill_(-10000.).to(self.device)
# normal_alpha_0 : alpha[0]=Ot[0]*self.PIs
# self.start_label has all of the score. it is log,0 is p=1
log_alpha[:, 0, self.start_label_id] = 0
# feats: sentances -> word embedding -> lstm -> MLP -> feats
# feats is the probability of emission, feat.shape=(1,tag_size)
for t in range(1, T):
log_alpha = (log_sum_exp_batch(self.transitions + log_alpha, axis=-1) + feats[:, t]).unsqueeze(1)
# log_prob of all barX
log_prob_all_barX = log_sum_exp_batch(log_alpha)
return log_prob_all_barX
def _get_bert_features(self, input_ids,
attention_mask,
token_type_ids,
position_ids,
head_mask,
inputs_embeds,
output_attentions,
output_hidden_states,
return_dict):
"""
sentences -> word embedding -> lstm -> MLP -> feats
"""
bert_seq_out = self.bert(input_ids,
attention_mask=attention_mask,
token_type_ids=token_type_ids,
position_ids=position_ids,
head_mask=head_mask,
inputs_embeds=inputs_embeds,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
return_dict=return_dict) # output_all_encoded_layers=False removed
bert_seq_out_last = bert_seq_out[0]
bert_seq_out_last = self.dropout(bert_seq_out_last)
bert_feats = self.hidden2label(bert_seq_out_last)
return bert_feats, bert_seq_out
def _score_sentence(self, feats, label_ids):
"""
Gives the score of a provided label sequence
p(X=w1:t,Zt=tag1:t)=...p(Zt=tag_t|Zt-1=tag_t-1)p(xt|Zt=tag_t)...
"""
# T = self.max_seq_length
T = feats.shape[1]
batch_size = feats.shape[0]
batch_transitions = self.transitions.expand(batch_size, self.num_labels, self.num_labels)
batch_transitions = batch_transitions.flatten(1)
score = torch.zeros((feats.shape[0], 1)).to(self.device)
# the 0th node is start_label->start_word, the probability of them=1. so t begins with 1.
for t in range(1, T):
score = score + \
batch_transitions.gather(-1, (label_ids[:, t] * self.num_labels + label_ids[:, t-1]).view(-1, 1)) + \
feats[:, t].gather(-1, label_ids[:, t].view(-1, 1)).view(-1, 1)
return score
def _viterbi_decode(self, feats):
"""
Max-Product Algorithm or viterbi algorithm, argmax(p(z_0:t|x_0:t))
"""
# T = self.max_seq_length
# feats=feats[0]#added
T = feats.shape[1]
batch_size = feats.shape[0]
# batch_transitions=self.transitions.expand(batch_size,self.num_labels,self.num_labels)
log_delta = torch.Tensor(batch_size, 1, self.num_labels).fill_(-10000.).to(self.device)
log_delta[:, 0, self.start_label_id] = 0
# psi is for the value of the last latent that make P(this_latent) maximum.
psi = torch.zeros((batch_size, T, self.num_labels), dtype=torch.long).to(self.device) # psi[0]=0000 useless
for t in range(1, T):
# delta[t][k]=max_z1:t-1( p(x1,x2,...,xt,z1,z2,...,zt-1,zt=k|theta) )
# delta[t] is the max prob of the path from z_t-1 to z_t[k]
log_delta, psi[:, t] = torch.max(self.transitions + log_delta, -1)
# psi[t][k]=argmax_z1:t-1( p(x1,x2,...,xt,z1,z2,...,zt-1,zt=k|theta) )
# psi[t][k] is the path chosen from z_t-1 to z_t[k],the value is the z_state(is k) index of z_t-1
log_delta = (log_delta + feats[:, t]).unsqueeze(1)
# trace back
path = torch.zeros((batch_size, T), dtype=torch.long).to(self.device)
# max p(z1:t,all_x|theta)
max_logLL_allz_allx, path[:, -1] = torch.max(log_delta.squeeze(), -1)
for t in range(T-2, -1, -1):
# choose the state of z_t according the state chosen of z_t+1.
path[:, t] = psi[:, t+1].gather(-1, path[:, t+1].view(-1, 1)).squeeze()
return max_logLL_allz_allx, path
def neg_log_likelihood(self, input_ids,
attention_mask,
token_type_ids,
position_ids,
head_mask,
inputs_embeds,
output_attentions,
output_hidden_states,
return_dict,
label_ids):
bert_feats, _ = self._get_bert_features(input_ids,
attention_mask,
token_type_ids,
position_ids,
head_mask,
inputs_embeds,
output_attentions,
output_hidden_states,
return_dict)
forward_score = self._forward_alg(bert_feats)
# p(X=w1:t,Zt=tag1:t)=...p(Zt=tag_t|Zt-1=tag_t-1)p(xt|Zt=tag_t)...
gold_score = self._score_sentence(bert_feats, label_ids)
# - log[ p(X=w1:t,Zt=tag1:t)/p(X=w1:t) ] = - log[ p(Zt=tag1:t|X=w1:t) ]
return torch.mean(forward_score - gold_score)
# this forward is just for predict, not for train
# dont confuse this with _forward_alg above.
def forward(
self,
input_ids=None,
attention_mask=None,
token_type_ids=None,
position_ids=None,
head_mask=None,
inputs_embeds=None,
labels=None,
output_attentions=None,
output_hidden_states=None,
return_dict=None,
inference_mode=False,
):
# Get the emission scores from the BiLSTM
bert_feats, bert_out = self._get_bert_features(input_ids,
attention_mask,
token_type_ids,
position_ids,
head_mask,
inputs_embeds,
output_attentions,
output_hidden_states,
return_dict)
# Find the best path, given the features.
score, label_seq_ids = self._viterbi_decode(bert_feats)
if not inference_mode:
neg_log_likelihood = self.neg_log_likelihood(input_ids,
attention_mask,
token_type_ids,
position_ids,
head_mask,
inputs_embeds,
output_attentions,
output_hidden_states,
return_dict,
labels)
return TokenClassifierOutput(
loss=neg_log_likelihood,
logits=label_seq_ids,
hidden_states=bert_out.hidden_states,
attentions=bert_out.attentions,
)
else:
neg_log_likelihood = None
return TokenClassifierOutput(
loss=neg_log_likelihood,
logits=label_seq_ids,
hidden_states=bert_out.hidden_states,
attentions=bert_out.attentions,
)
|