support sentence-transformers
Browse files- config_sentence_transformers.json +10 -0
- custom_st.py +189 -0
- modules.json +21 -0
config_sentence_transformers.json
ADDED
@@ -0,0 +1,10 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"__version__": {
|
3 |
+
"sentence_transformers": "3.1.0.dev0",
|
4 |
+
"transformers": "4.41.2",
|
5 |
+
"pytorch": "2.3.1+cu121"
|
6 |
+
},
|
7 |
+
"prompts": {},
|
8 |
+
"default_prompt_name": null,
|
9 |
+
"similarity_fn_name": "cosine"
|
10 |
+
}
|
custom_st.py
ADDED
@@ -0,0 +1,189 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import base64
|
2 |
+
import json
|
3 |
+
import os
|
4 |
+
from io import BytesIO
|
5 |
+
from typing import Any, Dict, List, Optional, Tuple, Union
|
6 |
+
|
7 |
+
import requests
|
8 |
+
import torch
|
9 |
+
from PIL import Image
|
10 |
+
from torch import nn
|
11 |
+
from transformers import AutoConfig, AutoImageProcessor, AutoModel, AutoTokenizer
|
12 |
+
|
13 |
+
|
14 |
+
class Transformer(nn.Module):
|
15 |
+
"""Huggingface AutoModel to generate token embeddings.
|
16 |
+
Loads the correct class, e.g. BERT / RoBERTa etc.
|
17 |
+
|
18 |
+
Args:
|
19 |
+
model_name_or_path: Huggingface models name
|
20 |
+
(https://huggingface.co/models)
|
21 |
+
max_seq_length: Truncate any inputs longer than max_seq_length
|
22 |
+
model_args: Keyword arguments passed to the Huggingface
|
23 |
+
Transformers model
|
24 |
+
tokenizer_args: Keyword arguments passed to the Huggingface
|
25 |
+
Transformers tokenizer
|
26 |
+
config_args: Keyword arguments passed to the Huggingface
|
27 |
+
Transformers config
|
28 |
+
cache_dir: Cache dir for Huggingface Transformers to store/load
|
29 |
+
models
|
30 |
+
do_lower_case: If true, lowercases the input (independent if the
|
31 |
+
model is cased or not)
|
32 |
+
tokenizer_name_or_path: Name or path of the tokenizer. When
|
33 |
+
None, then model_name_or_path is used
|
34 |
+
"""
|
35 |
+
|
36 |
+
def __init__(
|
37 |
+
self,
|
38 |
+
model_name_or_path: str,
|
39 |
+
max_seq_length: int | None = None,
|
40 |
+
model_args: dict[str, Any] | None = None,
|
41 |
+
tokenizer_args: dict[str, Any] | None = None,
|
42 |
+
config_args: dict[str, Any] | None = None,
|
43 |
+
cache_dir: str | None = None,
|
44 |
+
do_lower_case: bool = False,
|
45 |
+
tokenizer_name_or_path: str = None,
|
46 |
+
) -> None:
|
47 |
+
super().__init__()
|
48 |
+
self.config_keys = ["max_seq_length", "do_lower_case"]
|
49 |
+
self.do_lower_case = do_lower_case
|
50 |
+
if model_args is None:
|
51 |
+
model_args = {}
|
52 |
+
if tokenizer_args is None:
|
53 |
+
tokenizer_args = {}
|
54 |
+
if config_args is None:
|
55 |
+
config_args = {}
|
56 |
+
|
57 |
+
config = AutoConfig.from_pretrained(model_name_or_path, **config_args, cache_dir=cache_dir)
|
58 |
+
self._load_model(model_name_or_path, config, cache_dir, **model_args)
|
59 |
+
|
60 |
+
if max_seq_length is not None and "model_max_length" not in tokenizer_args:
|
61 |
+
tokenizer_args["model_max_length"] = max_seq_length
|
62 |
+
self.tokenizer = AutoTokenizer.from_pretrained(
|
63 |
+
tokenizer_name_or_path if tokenizer_name_or_path is not None else model_name_or_path,
|
64 |
+
cache_dir=cache_dir,
|
65 |
+
**tokenizer_args,
|
66 |
+
)
|
67 |
+
|
68 |
+
# No max_seq_length set. Try to infer from model
|
69 |
+
if max_seq_length is None:
|
70 |
+
if (
|
71 |
+
hasattr(self.auto_model, "config")
|
72 |
+
and hasattr(self.auto_model.config, "max_position_embeddings")
|
73 |
+
and hasattr(self.tokenizer, "model_max_length")
|
74 |
+
):
|
75 |
+
max_seq_length = min(self.auto_model.config.max_position_embeddings, self.tokenizer.model_max_length)
|
76 |
+
|
77 |
+
self.max_seq_length = max_seq_length
|
78 |
+
|
79 |
+
if tokenizer_name_or_path is not None:
|
80 |
+
self.auto_model.config.tokenizer_class = self.tokenizer.__class__.__name__
|
81 |
+
|
82 |
+
def forward(
|
83 |
+
self, features: Dict[str, torch.Tensor], task_type: Optional[str] = None
|
84 |
+
) -> Dict[str, torch.Tensor]:
|
85 |
+
"""Returns token_embeddings, cls_token"""
|
86 |
+
if task_type and task_type not in self._lora_adaptations:
|
87 |
+
raise ValueError(
|
88 |
+
f"Unsupported task '{task_type}'. "
|
89 |
+
f"Supported tasks are: {', '.join(self.config.lora_adaptations)}."
|
90 |
+
f"Alternatively, don't pass the `task_type` argument to disable LoRA."
|
91 |
+
)
|
92 |
+
|
93 |
+
adapter_mask = None
|
94 |
+
if task_type:
|
95 |
+
task_id = self._adaptation_map[task_type]
|
96 |
+
num_examples = 1
|
97 |
+
if isinstance(features['input_ids'][0], list):
|
98 |
+
# If input_ids[0] is a list, it means multiple inputs (list of texts)
|
99 |
+
num_examples = len(features['input_ids'])
|
100 |
+
|
101 |
+
adapter_mask = torch.full(
|
102 |
+
(num_examples,), task_id, dtype=torch.int32, device=self.device
|
103 |
+
)
|
104 |
+
|
105 |
+
lora_arguments = (
|
106 |
+
{"adapter_mask": adapter_mask} if adapter_mask is not None else {}
|
107 |
+
)
|
108 |
+
output_states = self.forward(**features, **lora_arguments, return_dict=False)
|
109 |
+
output_tokens = output_states[0]
|
110 |
+
features.update({"token_embeddings": output_tokens, "attention_mask": features["attention_mask"]})
|
111 |
+
return features
|
112 |
+
|
113 |
+
def get_word_embedding_dimension(self) -> int:
|
114 |
+
return self.auto_model.config.hidden_size
|
115 |
+
|
116 |
+
def tokenize(
|
117 |
+
self, texts: list[str] | list[dict] | list[tuple[str, str]], padding: str | bool = True
|
118 |
+
) -> dict[str, torch.Tensor]:
|
119 |
+
"""Tokenizes a text and maps tokens to token-ids"""
|
120 |
+
output = {}
|
121 |
+
if isinstance(texts[0], str):
|
122 |
+
to_tokenize = [texts]
|
123 |
+
elif isinstance(texts[0], dict):
|
124 |
+
to_tokenize = []
|
125 |
+
output["text_keys"] = []
|
126 |
+
for lookup in texts:
|
127 |
+
text_key, text = next(iter(lookup.items()))
|
128 |
+
to_tokenize.append(text)
|
129 |
+
output["text_keys"].append(text_key)
|
130 |
+
to_tokenize = [to_tokenize]
|
131 |
+
else:
|
132 |
+
batch1, batch2 = [], []
|
133 |
+
for text_tuple in texts:
|
134 |
+
batch1.append(text_tuple[0])
|
135 |
+
batch2.append(text_tuple[1])
|
136 |
+
to_tokenize = [batch1, batch2]
|
137 |
+
|
138 |
+
# strip
|
139 |
+
to_tokenize = [[str(s).strip() for s in col] for col in to_tokenize]
|
140 |
+
|
141 |
+
# Lowercase
|
142 |
+
if self.do_lower_case:
|
143 |
+
to_tokenize = [[s.lower() for s in col] for col in to_tokenize]
|
144 |
+
|
145 |
+
output.update(
|
146 |
+
self.tokenizer(
|
147 |
+
*to_tokenize,
|
148 |
+
padding=padding,
|
149 |
+
truncation="longest_first",
|
150 |
+
return_tensors="pt",
|
151 |
+
max_length=self.max_seq_length,
|
152 |
+
)
|
153 |
+
)
|
154 |
+
return output
|
155 |
+
|
156 |
+
def save(self, output_path: str, safe_serialization: bool = True) -> None:
|
157 |
+
self.auto_model.save_pretrained(output_path, safe_serialization=safe_serialization)
|
158 |
+
self.tokenizer.save_pretrained(output_path)
|
159 |
+
|
160 |
+
with open(os.path.join(output_path, "sentence_bert_config.json"), "w") as fOut:
|
161 |
+
json.dump(self.get_config_dict(), fOut, indent=2)
|
162 |
+
|
163 |
+
|
164 |
+
@classmethod
|
165 |
+
def load(cls, input_path: str) -> "Transformer":
|
166 |
+
# Old classes used other config names than 'sentence_bert_config.json'
|
167 |
+
for config_name in [
|
168 |
+
"sentence_bert_config.json",
|
169 |
+
"sentence_roberta_config.json",
|
170 |
+
"sentence_distilbert_config.json",
|
171 |
+
"sentence_camembert_config.json",
|
172 |
+
"sentence_albert_config.json",
|
173 |
+
"sentence_xlm-roberta_config.json",
|
174 |
+
"sentence_xlnet_config.json",
|
175 |
+
]:
|
176 |
+
sbert_config_path = os.path.join(input_path, config_name)
|
177 |
+
if os.path.exists(sbert_config_path):
|
178 |
+
break
|
179 |
+
|
180 |
+
with open(sbert_config_path) as fIn:
|
181 |
+
config = json.load(fIn)
|
182 |
+
# Don't allow configs to set trust_remote_code
|
183 |
+
if "model_args" in config and "trust_remote_code" in config["model_args"]:
|
184 |
+
config["model_args"].pop("trust_remote_code")
|
185 |
+
if "tokenizer_args" in config and "trust_remote_code" in config["tokenizer_args"]:
|
186 |
+
config["tokenizer_args"].pop("trust_remote_code")
|
187 |
+
if "config_args" in config and "trust_remote_code" in config["config_args"]:
|
188 |
+
config["config_args"].pop("trust_remote_code")
|
189 |
+
return cls(model_name_or_path=input_path, **config)
|
modules.json
ADDED
@@ -0,0 +1,21 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
[
|
2 |
+
{
|
3 |
+
"idx": 0,
|
4 |
+
"name": "0",
|
5 |
+
"path": "",
|
6 |
+
"type": "custom_st.Transformer",
|
7 |
+
"kwargs": ["task_type"]
|
8 |
+
},
|
9 |
+
{
|
10 |
+
"idx": 1,
|
11 |
+
"name": "1",
|
12 |
+
"path": "1_Pooling",
|
13 |
+
"type": "sentence_transformers.models.Pooling"
|
14 |
+
},
|
15 |
+
{
|
16 |
+
"idx": 2,
|
17 |
+
"name": "2",
|
18 |
+
"path": "2_Normalize",
|
19 |
+
"type": "sentence_transformers.models.Normalize"
|
20 |
+
}
|
21 |
+
]
|