simbert文本相似语义召回；保存及在线服务/weixin_42357472/article/details/116205077
SimBERT（基于UniLM思想、融检索与生成于一体的BERT模型）【主要应用场景：相似文本生成、相似文本检索】
/u013250861/article/details/123649047
import numpy as npimport osfrom collections import Counteros.environ['TF_KERAS'] = '1'from bert4keras.backend import keras, Kfrom bert4keras.models import build_transformer_modelfrom bert4keras.tokenizers import Tokenizerfrom bert4keras.snippets import sequence_paddingfrom bert4keras.snippets import unioutfrom keras.models import Modelmaxlen = 32# bert配置# bert配置config_path = r'D:***t\chinese_simbert_L-6_H-384_A-12\bert_config.json'checkpoint_path = r'D:\*****rt\chinese_simbert_L-6_H-384_A-12\bert_model.ckpt'dict_path = r'D:\****rt\chinese_simbert_L-6_H-384_A-12\vocab.txt'# 建立分词器tokenizer = Tokenizer(dict_path, do_lower_case=True) # 建立分词器# 建立加载模型bert = build_transformer_model(config_path,checkpoint_path,with_pool='linear',application='unilm',return_keras_model=False,)encoder = keras.models.Model(bert.model.inputs, bert.model.outputs[0])import pandas as pddatas1 = pd.read_csv(r'D:****raw_datas150.csv')datas_all = list(datas1["title"])# 测试相似度效果data = datas_alla_token_ids, b_token_ids, labels = [], [], []texts = []for d in data:token_ids = tokenizer.encode(d, maxlen=maxlen)[0]a_token_ids.append(token_ids)#token_ids = tokenizer.encode(d[1], maxlen=maxlen)[0]#b_token_ids.append(token_ids)#labels.append(d[2])texts.append(d)a_token_ids = sequence_padding(a_token_ids)# b_token_ids = sequence_padding(b_token_ids)a_vecs = encoder.predict([a_token_ids, np.zeros_like(a_token_ids)],verbose=True)# b_vecs = encoder.predict([b_token_ids, np.zeros_like(b_token_ids)],#verbose=True)# labels = np.array(labels)a_vecs = a_vecs / (a_vecs**2).sum(axis=1, keepdims=True)**0.5print(type(a_vecs))np.save("sim_all_datas.npy",a_vecs)#import numpy as np#a_vecsss = np.load(r"D:\tcl\simbert\sim_all_datas.npy")def most_similar(text, topn=10):"""检索最相近的topn个句子"""token_ids, segment_ids = tokenizer.encode(text, max_length=maxlen)print(token_ids, segment_ids )vec = encoder.predict([[token_ids], [segment_ids]])[0]vec /= (vec**2).sum()**0.5sims = np.dot(a_vecsss, vec)return [(i, datas_all[i], sims[i]) for i in sims.argsort()[::-1][:topn]]kk=["海绵宝宝"]mmm = []for i in kk:results = most_similar(i, 10)mmm.append([i,results])print(i,results)
from paddlenlp import Taskflowsimilarity = Taskflow("text_similarity")[-03-22 15:17:18,306] [ INFO] - Downloading model_state.pdparams from [/paddlenlp/taskflow/text_similarity/simbert-base-chinese/model_state.pdparams](/paddlenlp/taskflow/text_similarity/simbert-base-chinese/model_state.pdparams)100%|█████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 615M/615M [00:29<00:00, 22.1MB/s][-03-22 15:17:51,977] [ INFO] - Downloading model_config.json from [/paddlenlp/taskflow/text_similarity/simbert-base-chinese/model_config.json](/paddlenlp/taskflow/text_similarity/simbert-base-chinese/model_config.json)100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 334/334 [00:00<00:00, 197kB/s][-03-22 15:17:52,154] [ INFO] - Downloading /paddlenlp/models/transformers/simbert/vocab.txt and saved to /root/.paddlenlp/models/simbert-base-chinese[-03-22 15:17:52,154] [ INFO] - Downloading vocab.txt from [/paddlenlp/models/transformers/simbert/vocab.txt](/paddlenlp/models/transformers/simbert/vocab.txt)100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 63.4k/63.4k [00:00<00:00, 744kB/s][-03-22 15:18:10,818] [ INFO] - Weights from pretrained model not used in BertModel: ['cls.predictions.decoder_bias', 'cls.predictions.transform.weight', 'cls.predictions.transform.bias', 'cls.predictions.transform.LayerNorm.weight', 'cls.predictions.transform.LayerNorm.bias', 'cls.predictions.decoder_weight', 'cls.predictions.decoder.bias', 'cls.seq_relationship.weight', 'cls.seq_relationship.bias'][-03-22 15:18:12,113] [ INFO] - Converting to the inference model cost a little time.[-03-22 15:18:30,093] [ INFO] - The inference model save in the path:/root/.paddlenlp/taskflow/text_similarity/simbert-base-chinese/static/inference
