利用pytorch_pretrained_bert得到wordEmbedding (batch_size，1024),保存为稀疏矩阵sparse.dok_matrix( )

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
//st为一行文本，多行文本加个for循环即可
from pytorch_pretrained_bert import BertTokenizer,BertModel

test_input_ids=[]
tokenizer = BertTokenizer.from_pretrained('bert-large-uncased', do_lower_case=True)

text = tokenizer.tokenize(st)

text_id=tokenizer.convert_tokens_to_ids(text)

test_input_ids.append(text_id)

model=BertModel.from_pretrained('bert-large-uncased')

Input_ids = torch.tensor(test_input_ids).view(1, len(test_input_ids[i]))
//pooled_output即为embedding= [1，分词长度,1024]

all_encoder_layers, pooled_output = model(Input_ids)

//然后将pooled_output降维成[分词长度,1024],再按行累加得到一个embedding=[1,1024]
//传入多行文本,即得到embedding=[batch_size,1024]

1
2
3
4
5
6
7
8
9
10
11
12
13
import scipy.sparse as sparse


//注意通过Bert之后得到Embedding我处理成了(batch_size,1024)的形式
//每一行代表一个文档向量
//embedding是一个(batch_size,1024)的list

sparse_raw_matrix = sparse.dok_matrix((batch_size,1024))
for i in range(0,batch_size):
    for j in range(1024):
        sparse_raw_matrix[i,j]=embedding[i][j]
sparse.save_npz('embedding.npz', sparse_raw_matrix.tocsr())        

1
2
3
4
5
import scipy
from scipy import sparse
test=scipy.sparse.load_npz('embedding.npz')
test=test.toarray()