:gem: [Feature] Enable onnx model for embedder
Browse files- transforms/embed.py +70 -7
transforms/embed.py
CHANGED
@@ -1,10 +1,16 @@
|
|
1 |
import os
|
2 |
|
|
|
|
|
|
|
|
|
3 |
from typing import Union
|
4 |
|
5 |
-
from
|
6 |
-
from transformers import AutoModel
|
7 |
from numpy.linalg import norm
|
|
|
|
|
|
|
8 |
|
9 |
from configs.envs import ENVS
|
10 |
from configs.constants import AVAILABLE_MODELS
|
@@ -18,6 +24,59 @@ def cosine_similarity(a, b):
|
|
18 |
return (a @ b.T) / (norm(a) * norm(b))
|
19 |
|
20 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
21 |
class JinaAIEmbedder:
|
22 |
def __init__(self, model_name: str = AVAILABLE_MODELS[0]):
|
23 |
self.model_name = model_name
|
@@ -44,9 +103,13 @@ class JinaAIEmbedder:
|
|
44 |
|
45 |
|
46 |
if __name__ == "__main__":
|
47 |
-
embedder = JinaAIEmbedder()
|
48 |
-
|
49 |
-
|
50 |
-
embeddings =
|
|
|
|
|
51 |
logger.success(embeddings)
|
52 |
-
|
|
|
|
|
|
1 |
import os
|
2 |
|
3 |
+
import numpy as np
|
4 |
+
import torch
|
5 |
+
|
6 |
+
from pathlib import Path
|
7 |
from typing import Union
|
8 |
|
9 |
+
from huggingface_hub import hf_hub_download
|
|
|
10 |
from numpy.linalg import norm
|
11 |
+
from onnxruntime import InferenceSession
|
12 |
+
from tclogger import logger
|
13 |
+
from transformers import AutoTokenizer, AutoModel
|
14 |
|
15 |
from configs.envs import ENVS
|
16 |
from configs.constants import AVAILABLE_MODELS
|
|
|
24 |
return (a @ b.T) / (norm(a) * norm(b))
|
25 |
|
26 |
|
27 |
+
class JinaAIOnnxEmbedder:
|
28 |
+
"""https://huggingface.co/jinaai/jina-embeddings-v2-base-zh/discussions/6#65bc55a854ab5eb7b6300893"""
|
29 |
+
|
30 |
+
def __init__(self):
|
31 |
+
self.repo_name = "jinaai/jina-embeddings-v2-base-zh"
|
32 |
+
self.download_model()
|
33 |
+
self.load_model()
|
34 |
+
|
35 |
+
def download_model(self):
|
36 |
+
self.onnx_folder = Path(__file__).parent
|
37 |
+
self.onnx_filename = "onnx/model_quantized.onnx"
|
38 |
+
self.onnx_path = self.onnx_folder / self.onnx_filename
|
39 |
+
if not self.onnx_path.exists():
|
40 |
+
logger.note("> Downloading ONNX model")
|
41 |
+
hf_hub_download(
|
42 |
+
repo_id=self.repo_name,
|
43 |
+
filename=self.onnx_filename,
|
44 |
+
local_dir=self.onnx_folder,
|
45 |
+
local_dir_use_symlinks=False,
|
46 |
+
)
|
47 |
+
logger.success(f"+ ONNX model downloaded: {self.onnx_path}")
|
48 |
+
else:
|
49 |
+
logger.success(f"+ ONNX model loaded: {self.onnx_path}")
|
50 |
+
|
51 |
+
def load_model(self):
|
52 |
+
self.tokenizer = AutoTokenizer.from_pretrained(
|
53 |
+
self.repo_name, trust_remote_code=True
|
54 |
+
)
|
55 |
+
self.session = InferenceSession(self.onnx_path)
|
56 |
+
|
57 |
+
def mean_pooling(self, model_output, attention_mask):
|
58 |
+
token_embeddings = model_output
|
59 |
+
input_mask_expanded = (
|
60 |
+
attention_mask.unsqueeze(-1).expand(token_embeddings.size()).float()
|
61 |
+
)
|
62 |
+
return torch.sum(token_embeddings * input_mask_expanded, 1) / torch.clamp(
|
63 |
+
input_mask_expanded.sum(1), min=1e-9
|
64 |
+
)
|
65 |
+
|
66 |
+
def encode(self, text: str):
|
67 |
+
inputs = self.tokenizer(text, return_tensors="np")
|
68 |
+
inputs = {
|
69 |
+
name: np.array(tensor, dtype=np.int64) for name, tensor in inputs.items()
|
70 |
+
}
|
71 |
+
outputs = self.session.run(
|
72 |
+
output_names=["last_hidden_state"], input_feed=dict(inputs)
|
73 |
+
)
|
74 |
+
embeddings = self.mean_pooling(
|
75 |
+
torch.from_numpy(outputs[0]), torch.from_numpy(inputs["attention_mask"])
|
76 |
+
)
|
77 |
+
return embeddings
|
78 |
+
|
79 |
+
|
80 |
class JinaAIEmbedder:
|
81 |
def __init__(self, model_name: str = AVAILABLE_MODELS[0]):
|
82 |
self.model_name = model_name
|
|
|
103 |
|
104 |
|
105 |
if __name__ == "__main__":
|
106 |
+
# embedder = JinaAIEmbedder()
|
107 |
+
embedder = JinaAIOnnxEmbedder()
|
108 |
+
texts = ["How is the weather today?", "今天天气怎么样?"]
|
109 |
+
embeddings = []
|
110 |
+
for text in texts:
|
111 |
+
embeddings.append(embedder.encode(text))
|
112 |
logger.success(embeddings)
|
113 |
+
print(cosine_similarity(embeddings[0], embeddings[1]))
|
114 |
+
|
115 |
+
# python -m transforms.embed
|