pytorch模型变压器库针对GPU的NLP任务的有效预测

如何解决pytorch模型变压器库针对GPU的NLP任务的有效预测

我拥有大量的推文文本数据集（近30亿条推文）。我对带有注释数据集的BERT模型进行了分类。

有没有一种方法可以使预测更有效，更快？我只能访问一个GPU。

此刻，我使用以下代码：

import pandas as pd

import sys

from transformers import pipeline

judge = pipeline(
task="sentiment-analysis",model="/trained_disasterlabels",tokenizer='bert-base-uncased',device=0)

id_disaster=sys.argv[1]

path="/disaster_data/ids_"+id_disaster+"_text"

open(path+"_labelpred",'w').close()

n=200
with open(path,"rb") as f:
    batch=[]
    for line in f:
        batch.append(line.decode().rstrip("\n"))
        if len(batch)==n:

            preds=judge(batch)
            preds_labels=[x['label'].replace("LABEL_","") for x in preds]
            preds_probs=[round(x['score'],4) for x in preds]
            
            valsdf=pd.DataFrame({"labels":preds_labels,"probs":preds_probs})
            
            valsdf.to_csv(path+"_labelpred",mode='a',header=False,index=False)
            
            batch=[]

    preds=judge(batch)
    preds_labels=[x['label'].replace("LABEL_","") for x in preds]
    preds_probs=[round(x['score'],4) for x in preds]
    
    valsdf=pd.DataFrame({"labels":preds_labels,"probs":preds_probs})
    
    valsdf.to_csv(path+"_labelpred",index=False)

在我还尝试以下方法之前：

model=BertForSequenceClassification.from_pretrained("/trained_disasterlabels")

model.eval()

model.to('cuda')

tokenizer=BertTokenizer.from_pretrained('bert-base-uncased')

import torch.nn.functional as F 

path="/disaster_data/ids_"+id_disaster+"_text"

open(path+"_labelpred",'w').close()

n=100
 
with open(path,"rb") as f:
    batch=[]
    for line in f:
        batch.append(line.decode().rstrip("\n"))
        if len(batch)==n:
           input_ids = tokenizer.batch_encode_plus(batch,add_special_tokens=True,return_tensors = 'pt',padding=True,truncation=True)
            
            input_ids.to('cuda')
            
            with torch.no_grad():
                last_hidden_states = model(**input_ids)
            
            temp_cpu = F.softmax(last_hidden_states[0],dim=1).detach().cpu().numpy()    

            valsdf=pd.DataFrame(temp_cpu.tolist(),columns=["prob","labelpred"])
                        
            batch=[]
            
    input_ids = tokenizer.batch_encode_plus(batch,truncation=True)

    with torch.no_grad():
        last_hidden_states = model(**input_ids)  
        
    temp_cpu = F.softmax(last_hidden_states[0],dim=1).detach().cpu().numpy()
            
    valsdf=pd.DataFrame(temp_cpu.tolist(),columns=["labelpred","prob"])
    
    valsdf.to_csv(path+"_labelpred",index=False)
           
    batch=[]