sentence_transformers 微调模型

模型使用的是：uer/sbert-base-chinese-nli · Hugging Face

使用sentencetransformer微调模型

sentence_transformers 官网：SentenceTransformers Documentation — Sentence-Transformers documentation

使用自己的数据微调模型

sentence_transformer 使用自己训练数据微调模型的代码如下所示。(模型为计算句子相似度)

from sentence_transformers import SentenceTransformer,SentenceDataset,InputExample,evaluation,losses,util
class sbert():
    def build_train_data(self,o1,o2,n1,n2,train_size):
        train_data = []
        for i in range(train_size):
            train_data.append(InputExample(tests=[o1[i],o2[i]],label=1.0))
            train_data.append(InputExample(tests=[n1[i],n2[i]],label=1.0))
        return train_data
    def build_evaluation_data(o1,o2,n1,n2,train_size,eval_size):
        s1 = o1[train_size:]
        s2 = o2[train_size:]
        s1.extend(list(n1[train_size:]))
        s2.extend(list(n2[train_size:]))
        score = [1.0]*eval_size + [0.0]*eval_size
        evaluator = evaluation.EmbeddingSimilarityEvaluator(s1,s2,score)
        return evaluator
    def callback(self,score,epoch,steps)
        print('score:{},epoch:{},steps:{}'.format(score,epoch,steps))

    def train(self):
        #1.获取正、负样本,o1是标准问，O2是相似问
        o1,o2 = self.get_act_data()
        n1,n2 = self.get_neg_data()

        #2.定义训练集、测试集大小 + 构造训练数据
        train_size = int(len(o1)*0.8)
        eval_size = len(o1) - train_size
        train_data = self.build_train_data(o1,o2,n1,n2,train_size)
        
        #3.定义测试数据
        evaluator = self.build_evaluation_data(o1,o2,n1,n2,train_size,eval_size)

        #4.需要训练的模型
        mode = SentenceTransformer('模型地址')

        #5
        train_dataset = SentenceDataset(train_data,model)
        train_dataloader = DataLoader(train_dataset,shuffle =true, batch_size = 8)
        train_loss = losses.CosineSimilarityLoss(model)

        #6.调试模型
        model.fit(train_objectives = [(train_dataloader,train_loss)],epochs = 1,warmup_steps = 100,evaluator = evaluator,evaluation_steps = 100,output_path = '存调试后模型的地址',save_best_model = True,callback = self.callback)
        

sentence_transformer使用自己微调后的模型的代码如下所示：

#1. 定义模型
model = SentenceTransformer('模型地址')
#2.编码向量
o1_emb = model.encode(['数据list','求求一定要好运啊'])
o2_emb = model.encode(['一定要是列表','我绝对可以好运']
#计算相似度
cosine_score0 = util.cos_sim(o1_emb,o2_emb)
cosine_score = []
for i in range(len(cosine_score0)):
    cosine_score.append(cosine_score0[i][i].numpy().tolist()

增加模型层数

from sentence_transformers import SentenceTransformer, models
from torch import nn

word_embedding_model = models.Transformer('bert-base-uncased', max_seq_length=256)
pooling_model = models.Pooling(word_embedding_model.get_word_embedding_dimension())
dense_model = models.Dense(in_features=pooling_model.get_sentence_embedding_dimension(), out_features=256, activation_function=nn.Tanh())

model = SentenceTransformer(modules=[word_embedding_model, pooling_model, dense_model])