背景
用BERT對句子進行向量化
實施
TensorFlow版直接用肖涵博士的bert-as-service。使用方法真的很小白,簡單概括爲2點:server和client安裝。
pip install bert-serving-server # server
pip install bert-serving-client # client, independent of `bert-serving-server`
在server安裝完後,啓動服務,比如:bert-serving-start -model_dir /home/pretained_models/chinese_wwm_ext_L-12_H-768_A-12 -num_worker=4
通過model_dir
參數可以自行指定不同類型的BERT的模型路徑,我這裏使用的是哈工大發布的WWM-EXT版。在client上的測試代碼:
def test_bert_tf(string):
from bert_serving.client import BertClient
bc = BertClient()
s_encode = bc.encode([string])
print(s_encode[0])
上述方案雖然簡單易於上手,但是個人還是覺自己動手更香,比如基於huggingface的transformers。如何驗證呢?就以bert-as-service編碼得到的句向量作爲標準值。將相同的文本輸入到transformers試圖得到與bert-as-service方案相同的句向量。
由於bert-as-service默認的句向量構造方案是取倒數第二層的隱狀態值在token上的均值,即選用的層數是倒數第2層,池化策略是REDUCE_MEAN
。
import torch
import pdb
from transformers import AutoConfig
from transformers import BertTokenizer, BertModel, BertConfig
UNCASE = "/home/pretained_models/chinese_wwm_ext_pytorch"
VOCAB = "vocab.txt"
tokenizer = BertTokenizer.from_pretrained(UNCASE + "/" + VOCAB)
model = BertModel.from_pretrained(UNCASE, output_hidden_states = True) # 如果想要獲取到各個隱層值需要如此設置
model.eval()
string = '寫代碼不香嗎'
string1 = "[CLS]" + string + "[SEP]"
# Convert token to vocabulary indices
tokenized_string = tokenizer.tokenize(string1)
tokens_ids = tokenizer.convert_tokens_to_ids(tokenized_string)
# Convert inputs to PyTorch tensors
tokens_tensor = torch.tensor([tokens_ids])
outputs = model(tokens_tensor) # encoded_layers, pooled_output
if model.config.output_hidden_states:
hidden_states = outputs[2]
# last_layer = outputs[-1]
second_to_last_layer = hidden_states[-2]
# 由於只要一個句子,所以尺寸爲[1, 10, 768]
token_vecs = second_to_last_layer[0]
print(token_vecs.shape)
# Calculate the average of all input token vectors.
sentence_embedding = torch.mean(token_vecs, dim=0)
print(sentence_embedding.shape)
print(sentence_embedding[0:10])
print("tf version-----")
from bert_serving.client import BertClient
bc = BertClient()
s_encode = bc.encode([string])
print(s_encode[0].shape)
# pdb.set_trace()
print(s_encode[0][0:10])
結果如下:
從向量的前10維可以看出,兩者向量是相同的。那麼進一步計算二者的餘弦相似度的結果:
tf_tensor = torch.tensor(s_encode[0])
similarity = torch.cosine_similarity(sentence_embedding, tf_tensor, dim=0)
print(similarity)
餘弦相似度爲1,所以兩個向量相同。