Update README.md
Browse files
README.md
CHANGED
|
@@ -30,3 +30,78 @@ transformers>=4.39.2
|
|
| 30 |
flash_attn>=2.5.6
|
| 31 |
```
|
| 32 |
## Usage
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 30 |
flash_attn>=2.5.6
|
| 31 |
```
|
| 32 |
## Usage
|
| 33 |
+
|
| 34 |
+
Get Dense Embeddings with Transformers
|
| 35 |
+
```
|
| 36 |
+
# Requires transformers>=4.36.0
|
| 37 |
+
|
| 38 |
+
import torch.nn.functional as F
|
| 39 |
+
from transformers import AutoModel, AutoTokenizer
|
| 40 |
+
|
| 41 |
+
input_texts = [
|
| 42 |
+
"what is the capital of China?",
|
| 43 |
+
"how to implement quick sort in python?",
|
| 44 |
+
"北京",
|
| 45 |
+
"快排算法介绍"
|
| 46 |
+
]
|
| 47 |
+
|
| 48 |
+
model_path = 'Alibaba-NLP/gte-multilingual-base'
|
| 49 |
+
tokenizer = AutoTokenizer.from_pretrained(model_path)
|
| 50 |
+
model = AutoModel.from_pretrained(model_path, trust_remote_code=True)
|
| 51 |
+
|
| 52 |
+
# Tokenize the input texts
|
| 53 |
+
batch_dict = tokenizer(input_texts, max_length=8192, padding=True, truncation=True, return_tensors='pt')
|
| 54 |
+
|
| 55 |
+
outputs = model(**batch_dict)
|
| 56 |
+
|
| 57 |
+
dimension=768 # The output dimension of the output embedding, should be in [128, 768]
|
| 58 |
+
embeddings = outputs.last_hidden_state[:, 0][:dimension]
|
| 59 |
+
|
| 60 |
+
embeddings = F.normalize(embeddings, p=2, dim=1)
|
| 61 |
+
scores = (embeddings[:1] @ embeddings[1:].T) * 100
|
| 62 |
+
print(scores.tolist())
|
| 63 |
+
```
|
| 64 |
+
|
| 65 |
+
Use with sentence-transformers
|
| 66 |
+
```
|
| 67 |
+
from sentence_transformers import SentenceTransformer
|
| 68 |
+
from sentence_transformers.util import cos_sim
|
| 69 |
+
|
| 70 |
+
input_texts = [
|
| 71 |
+
"what is the capital of China?",
|
| 72 |
+
"how to implement quick sort in python?",
|
| 73 |
+
"北京",
|
| 74 |
+
"快排算法介绍"
|
| 75 |
+
]
|
| 76 |
+
|
| 77 |
+
model = SentenceTransformer('Alibaba-NLP/gte-multilingual-base', trust_remote_code=True)
|
| 78 |
+
embeddings = model.encode(input_texts)
|
| 79 |
+
```
|
| 80 |
+
|
| 81 |
+
Use with custom code to get dense embeddigns and sparse token weights
|
| 82 |
+
```
|
| 83 |
+
# You can find the gte_embeddings.py in https://huggingface.co/Alibaba-NLP/gte-multilingual-base/blob/main/scripts/gte_embedding.py
|
| 84 |
+
from gte_embeddings import GTEEmbeddidng
|
| 85 |
+
|
| 86 |
+
model_path = 'Alibaba-NLP/gte-multilingual-base'
|
| 87 |
+
model = GTEEmbeddidng(model_path)
|
| 88 |
+
query = "中国的首都在哪儿"
|
| 89 |
+
|
| 90 |
+
docs = [
|
| 91 |
+
"what is the capital of China?",
|
| 92 |
+
"how to implement quick sort in python?",
|
| 93 |
+
"北京",
|
| 94 |
+
"快排算法介绍"
|
| 95 |
+
]
|
| 96 |
+
|
| 97 |
+
embs = model.encode(docs, return_dense=True,return_sparse=True)
|
| 98 |
+
print('dense_embeddings vecs', embs['dense_embeddings'])
|
| 99 |
+
print('token_weights', embs['token_weights'])
|
| 100 |
+
pairs = [(query, doc) for doc in docs]
|
| 101 |
+
dense_scores = model.compute_scores(pairs, dense_weight=1.0, sparse_weight=0.0)
|
| 102 |
+
sparse_scores = model.compute_scores(pairs, dense_weight=0.0, sparse_weight=1.0)
|
| 103 |
+
hybird_scores = model.compute_scores(pairs, dense_weight=1.0, sparse_weight=0.3)
|
| 104 |
+
print('dense_scores', dense_scores)
|
| 105 |
+
print('sparse_scores', sparse_scores)
|
| 106 |
+
print('hybird_scores', hybird_scores)
|
| 107 |
+
```
|