【Coding】从BertForSequenceClassification分类出pooled_output作为final feature

从BertForSequenceClassification分类出pooled_output作为final feature

BertForSequenceClassification函数在设计时并未返回pooled_output参数：

class BertForSequenceClassification(BertPreTrainedModel):
    	    #....
	        #....
            outputs = 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,
        )
        #....
        return SequenceClassifierOutput(
            loss=loss,
            logits=logits,
            hidden_states=outputs.hidden_states,
            attentions=outputs.attentions,
        )

但有时我们又想用pooled_output的做后续的操作或分析其性质。一种方法是直接在huggingface的底层文件中更改BertForSequenceClassification返回的参数，但是比较麻烦，尤其不适用于多机器平台，不易于统一。

相对简便耦合性高的方法是先用BertForSequenceClassification对model进行训练，随后将model save起来，再用BertModel 将存储好的model 加载，传递参数给该model再返回pooled_output就可以。

class BertModel(BertPreTrainedModel):
		#...
        return BaseModelOutputWithPoolingAndCrossAttentions(
            last_hidden_state=sequence_output,
            pooler_output=pooled_output,
            past_key_values=encoder_outputs.past_key_values,
            hidden_states=encoder_outputs.hidden_states,
            attentions=encoder_outputs.attentions,
            cross_attentions=encoder_outputs.cross_attentions,
        )

样例pseudocode：

model = BertForSequenceClassification.from_pretrained(checkpoint, num_labels=num_class, 

model.save_pretrained('saved_path',state_dict=True)
model_eval = BertModel.from_pretrained(state_dict = 'saved_path').to(device)
model_eval.eval()
                                                      
for batch in dataloader:
    batch = {k: v.to(device) for k, v in batch.items()}
    with torch.no_grad():
            pooled_out = model_eval(input_ids=batch['input_ids'],
                                token_type_ids=batch['token_type_ids'],
                                attention_mask=batch['attention_mask'],
                                labels=batch['labels'])['pooler_out'] 

参考资料

Extracting Features from BertForSequenceClassification