本项目是 text2vec 项目的 onnxruntime 推理版本,实现了向量获取和文本匹配搜索。为了保证项目的轻量,只使用了 onnxruntime
、 tokenizers
和 numpy
三个库。
主要在 GanymedeNil/text2vec-base-chinese-onnx 模型上进行测试,理论上支持 BERT 系列模型。
pip install text2vec2onnx[cpu]
pip install text2vec2onnx[gpu]
以下载 GanymedeNil/text2vec-base-chinese-onnx 为例,下载模型到本地。
- huggingface 模型下载
huggingface-cli download --resume-download GanymedeNil/text2vec-base-chinese-onnx --local-dir text2vec-base-chinese-onnx
from text2vec2onnx import SentenceModel
embedder = SentenceModel(model_dir_path='local-dir')
emb = embedder.encode("你好")
from text2vec2onnx import SentenceModel, semantic_search
embedder = SentenceModel(model_dir_path='local-dir')
corpus = [
"谢谢观看 下集再见",
"感谢您的观看",
"请勿模仿",
"记得订阅我们的频道哦",
"The following are sentences in English.",
"Thank you. Bye-bye.",
"It's true",
"I don't know.",
"Thank you for watching!",
]
corpus_embeddings = embedder.encode(corpus)
queries = [
'Thank you. Bye.',
'你干啥呢',
'感谢您的收听']
for query in queries:
query_embedding = embedder.encode(query)
hits = semantic_search(query_embedding, corpus_embeddings, top_k=1)
print("\n\n======================\n\n")
print("Query:", query)
print("\nTop 5 most similar sentences in corpus:")
hits = hits[0] # Get the hits for the first query
for hit in hits:
print(corpus[hit['corpus_id']], "(Score: {:.4f})".format(hit['score']))