michaelfeil commited on
Commit
b459537
1 Parent(s): 68e1553

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +47 -0
README.md CHANGED
@@ -2866,6 +2866,53 @@ sentence_embeddings = torch.nn.functional.normalize(sentence_embeddings, p=2, di
2866
  print("Sentence embeddings:", sentence_embeddings)
2867
  ```
2868
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2869
  ### Usage for Reranker
2870
 
2871
  Different from embedding model, reranker uses question and document as input and directly output similarity instead of embedding.
 
2866
  print("Sentence embeddings:", sentence_embeddings)
2867
  ```
2868
 
2869
+
2870
+ #### Usage of the ONNX files
2871
+
2872
+ ```python
2873
+ from optimum.onnxruntime import ORTModelForFeatureExtraction # type: ignore
2874
+
2875
+ import torch
2876
+ from transformers import AutoModel, AutoTokenizer
2877
+
2878
+ tokenizer = AutoTokenizer.from_pretrained('BAAI/bge-large-en-v1.5')
2879
+ model = AutoModel.from_pretrained('BAAI/bge-large-en-v1.5', revision="refs/pr/13")
2880
+ model_ort = ORTModelForFeatureExtraction.from_pretrained('BAAI/bge-large-en-v1.5', revision="refs/pr/13",file_name="onnx/model.onnx")
2881
+
2882
+ # Sentences we want sentence embeddings for
2883
+ sentences = ["样例数据-1", "样例数据-2"]
2884
+
2885
+ # Tokenize sentences
2886
+ encoded_input = tokenizer(sentences, padding=True, truncation=True, return_tensors='pt')
2887
+ # for s2p(short query to long passage) retrieval task, add an instruction to query (not add instruction for passages)
2888
+ # encoded_input = tokenizer([instruction + q for q in queries], padding=True, truncation=True, return_tensors='pt')
2889
+
2890
+ model_output_ort = model_ort(**encoded_input)
2891
+ # Compute token embeddings
2892
+ with torch.no_grad():
2893
+ model_output = model(**encoded_input)
2894
+
2895
+ # model_output and model_output_ort are identical
2896
+
2897
+ ```
2898
+
2899
+ #### Usage via infinity
2900
+ Its also possible to deploy the onnx files with the [infinity_emb](https://github.com/michaelfeil/infinity) pip package.
2901
+ ```python
2902
+ import asyncio
2903
+ from infinity_emb import AsyncEmbeddingEngine, EngineArgs
2904
+
2905
+ sentences = ["Embed this is sentence via Infinity.", "Paris is in France."]
2906
+ engine = AsyncEmbeddingEngine.from_args(
2907
+ EngineArgs(model_name_or_path = "BAAI/bge-large-en-v1.5", device="cpu", engine="optimum" # or engine="torch"
2908
+ ))
2909
+
2910
+ async def main():
2911
+ async with engine:
2912
+ embeddings, usage = await engine.embed(sentences=sentences)
2913
+ asyncio.run(main())
2914
+ ```
2915
+
2916
  ### Usage for Reranker
2917
 
2918
  Different from embedding model, reranker uses question and document as input and directly output similarity instead of embedding.