Create scripts/gte_embedding.py
Browse files- scripts/gte_embedding.py +190 -0
scripts/gte_embedding.py
ADDED
@@ -0,0 +1,190 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import logging
|
2 |
+
from typing import Dict, Optional, List, Tuple
|
3 |
+
import os
|
4 |
+
|
5 |
+
import heapq
|
6 |
+
import json
|
7 |
+
import logging
|
8 |
+
import os
|
9 |
+
import queue
|
10 |
+
import sys
|
11 |
+
import time
|
12 |
+
from tqdm import tqdm
|
13 |
+
|
14 |
+
import torch
|
15 |
+
from collections import defaultdict
|
16 |
+
from torch.utils.data._utils.worker import ManagerWatchdog
|
17 |
+
import numpy as np
|
18 |
+
import torch.distributed as dist
|
19 |
+
from torch import nn, Tensor
|
20 |
+
import torch.nn.functional as F
|
21 |
+
from transformers import AutoModel, AutoTokenizer
|
22 |
+
from transformers.file_utils import ModelOutput
|
23 |
+
|
24 |
+
logger = logging.getLogger(__name__)
|
25 |
+
|
26 |
+
|
27 |
+
class GTEEmbeddidng(nn.Module):
|
28 |
+
def __init__(self,
|
29 |
+
model_name: str = None,
|
30 |
+
normalized: bool = True,
|
31 |
+
pooling_method: str = 'cls',
|
32 |
+
use_fp16: bool = True,
|
33 |
+
device: str = None
|
34 |
+
):
|
35 |
+
super().__init__()
|
36 |
+
self.load_model(model_name)
|
37 |
+
self.vocab_size = self.model.config.vocab_size
|
38 |
+
self.normalized = normalized
|
39 |
+
self.pooling_method = pooling_method
|
40 |
+
if device:
|
41 |
+
self.device = torch.device(device)
|
42 |
+
else:
|
43 |
+
if torch.cuda.is_available():
|
44 |
+
self.device = torch.device("cuda")
|
45 |
+
elif torch.backends.mps.is_available():
|
46 |
+
self.device = torch.device("mps")
|
47 |
+
elif is_torch_npu_available():
|
48 |
+
self.device = torch.device("npu")
|
49 |
+
else:
|
50 |
+
self.device = torch.device("cpu")
|
51 |
+
use_fp16 = False
|
52 |
+
self.model.to(self.device)
|
53 |
+
self.sparse_linear.to(self.device)
|
54 |
+
if use_fp16:
|
55 |
+
self.model.half()
|
56 |
+
self.sparse_linear.half()
|
57 |
+
|
58 |
+
def load_model(self, model_name):
|
59 |
+
if not os.path.exists(model_name):
|
60 |
+
cache_folder = os.getenv('HF_HUB_CACHE')
|
61 |
+
model_name = snapshot_download(repo_id=model_name,
|
62 |
+
cache_dir=cache_folder,
|
63 |
+
ignore_patterns=['flax_model.msgpack', 'rust_model.ot', 'tf_model.h5'])
|
64 |
+
|
65 |
+
self.model = AutoModel.from_pretrained(model_name, trust_remote_code=True)
|
66 |
+
self.sparse_linear = torch.nn.Linear(in_features=self.model.config.hidden_size, out_features=1)
|
67 |
+
self.tokenizer = AutoTokenizer.from_pretrained(model_name)
|
68 |
+
self.model.eval()
|
69 |
+
if os.path.exists(os.path.join(model_name, 'sparse_linear.pt')):
|
70 |
+
logger.info('loading existing sparse_linear---------')
|
71 |
+
self.load_pooler(model_dir=model_name)
|
72 |
+
else:
|
73 |
+
logger.warring('The parameters of sparse linear is not found')
|
74 |
+
|
75 |
+
def dense_embedding(self, hidden_state, mask):
|
76 |
+
if self.pooling_method == 'cls':
|
77 |
+
return hidden_state[:, 0]
|
78 |
+
elif self.pooling_method == 'mean':
|
79 |
+
s = torch.sum(hidden_state * mask.unsqueeze(-1).float(), dim=1)
|
80 |
+
d = mask.sum(axis=1, keepdim=True).float()
|
81 |
+
return s / d
|
82 |
+
|
83 |
+
def sparse_embedding(self, hidden_state, input_ids, return_embedding: bool = True):
|
84 |
+
token_weights = torch.relu(self.sparse_linear(hidden_state))
|
85 |
+
return token_weights
|
86 |
+
|
87 |
+
def _process_token_weights(self, token_weights: np.ndarray, input_ids: list):
|
88 |
+
# conver to dict
|
89 |
+
result = defaultdict(int)
|
90 |
+
unused_tokens = set([self.tokenizer.cls_token_id, self.tokenizer.eos_token_id, self.tokenizer.pad_token_id,
|
91 |
+
self.tokenizer.unk_token_id])
|
92 |
+
# token_weights = np.ceil(token_weights * 100)
|
93 |
+
for w, idx in zip(token_weights, input_ids):
|
94 |
+
if idx not in unused_tokens and w > 0:
|
95 |
+
token = self.tokenizer.decode([int(idx)])
|
96 |
+
if w > result[token]:
|
97 |
+
result[token] = w
|
98 |
+
return result
|
99 |
+
|
100 |
+
@torch.no_grad()
|
101 |
+
def encode(self,
|
102 |
+
texts: None,
|
103 |
+
dimension: int = None,
|
104 |
+
max_length: int = 8192,
|
105 |
+
batch_size: int = 16,
|
106 |
+
return_dense: bool = True,
|
107 |
+
return_sparse: bool = False):
|
108 |
+
if dimension is None:
|
109 |
+
dimension = self.model.config.hidden_size
|
110 |
+
if isinstance(texts, str):
|
111 |
+
texts = [texts]
|
112 |
+
num_texts = len(texts)
|
113 |
+
all_dense_vecs = []
|
114 |
+
all_token_weights = []
|
115 |
+
for n, i in enumerate(range(0, num_texts, batch_size)):
|
116 |
+
batch = texts[i: i + batch_size]
|
117 |
+
resulst = self._encode(batch, dimension, max_length, batch_size, return_dense, return_sparse)
|
118 |
+
if return_dense:
|
119 |
+
all_dense_vecs.append(resulst['dense_embeddings'])
|
120 |
+
if return_sparse:
|
121 |
+
all_token_weights.extend(resulst['token_weights'])
|
122 |
+
all_dense_vecs = torch.cat(all_dense_vecs, dim=0)
|
123 |
+
return {
|
124 |
+
"dense_embeddings": all_dense_vecs,
|
125 |
+
"token_weights": all_token_weights
|
126 |
+
}
|
127 |
+
|
128 |
+
@torch.no_grad()
|
129 |
+
def _encode(self,
|
130 |
+
texts: Dict[str, Tensor] = None,
|
131 |
+
dimension: int = None,
|
132 |
+
max_length: int = 1024,
|
133 |
+
batch_size: int = 16,
|
134 |
+
return_dense: bool = True,
|
135 |
+
return_sparse: bool = False):
|
136 |
+
|
137 |
+
text_input = self.tokenizer(texts, padding=True, truncation=True, return_tensors='pt', max_length=max_length)
|
138 |
+
text_input = {k: v.to(self.model.device) for k,v in text_input.items()}
|
139 |
+
last_hidden_state = self.model(**text_input, return_dict=True).last_hidden_state
|
140 |
+
|
141 |
+
output = {}
|
142 |
+
if return_dense:
|
143 |
+
dense_vecs = self.dense_embedding(last_hidden_state, text_input['attention_mask'])
|
144 |
+
dense_vecs = dense_vecs[:, :dimension]
|
145 |
+
if self.normalized:
|
146 |
+
dense_vecs = torch.nn.functional.normalize(dense_vecs, dim=-1)
|
147 |
+
output['dense_embeddings'] = dense_vecs
|
148 |
+
if return_sparse:
|
149 |
+
token_weights = self.sparse_embedding(last_hidden_state, text_input['input_ids']).squeeze(-1)
|
150 |
+
token_weights = list(map(self._process_token_weights, token_weights.detach().cpu().numpy().tolist(),
|
151 |
+
text_input['input_ids'].cpu().numpy().tolist()))
|
152 |
+
output['token_weights'] = token_weights
|
153 |
+
|
154 |
+
return output
|
155 |
+
|
156 |
+
def load_pooler(self, model_dir):
|
157 |
+
sparse_state_dict = torch.load(os.path.join(model_dir, 'sparse_linear.pt'), map_location='cpu')
|
158 |
+
self.sparse_linear.load_state_dict(sparse_state_dict)
|
159 |
+
|
160 |
+
def _compute_sparse_scores(self, embs1, embs2):
|
161 |
+
scores = 0
|
162 |
+
for token, weight in embs1.items():
|
163 |
+
if token in embs2:
|
164 |
+
scores += weight * embs2[token]
|
165 |
+
return scores
|
166 |
+
|
167 |
+
def compute_sparse_scores(self, embs1, embs2):
|
168 |
+
scores = [self._compute_sparse_scores(emb1, emb2) for emb1, emb2 in zip(embs1, embs2)]
|
169 |
+
return np.array(scores)
|
170 |
+
|
171 |
+
def compute_dense_scores(self, embs1, embs2):
|
172 |
+
scores = torch.sum(embs1*embs2, dim=-1).cpu().detach().numpy()
|
173 |
+
return scores
|
174 |
+
|
175 |
+
@torch.no_grad()
|
176 |
+
def compute_scores(self,
|
177 |
+
text_pairs: List[Tuple[str, str]],
|
178 |
+
dimension: int = None,
|
179 |
+
max_length: int = 1024,
|
180 |
+
batch_size: int = 16,
|
181 |
+
dense_weight=1.0,
|
182 |
+
sparse_weight=0.1):
|
183 |
+
text1_list = [text_pair[0] for text_pair in text_pairs]
|
184 |
+
text2_list = [text_pair[1] for text_pair in text_pairs]
|
185 |
+
embs1 = self.encode(text1_list, dimension, max_length, batch_size, return_dense=True, return_sparse=True)
|
186 |
+
embs2 = self.encode(text2_list, dimension, max_length, batch_size, return_dense=True, return_sparse=True)
|
187 |
+
scores = self.compute_dense_scores(embs1['dense_embeddings'], embs2['dense_embeddings']) * dense_weight + \
|
188 |
+
self.compute_sparse_scores(embs1['token_weights'], embs2['token_weights']) * sparse_weight
|
189 |
+
scores = scores.tolist()
|
190 |
+
return scores
|