#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, )