\u5728\u5730\u7406\u4fe1\u606f\u5904\u7406\u4e0e\u5730\u5740\u6570\u636e\u6cbb\u7406\u9886\u57df，\u5b9e\u4f53\u5bf9\u9f50\u662f\u6784\u5efa\u9ad8\u8d28\u91cf\u5730\u5740\u77e5\u8bc6\u5e93\u7684\u6838\u5fc3\u73af\u8282\u3002\u5c24\u5176\u5728\u7535\u5546\u3001\u7269\u6d41\u3001\u57ce\u5e02\u6cbb\u7406\u7b49\u573a\u666f\u4e2d，\u6765\u81ea\u4e0d\u540c\u7cfb\u7edf\u7684\u5730\u5740\u8bb0\u5f55\u5f80\u5f80\u5b58\u5728\u8868\u8ff0\u5dee\u5f02\u2014\u2014\u5982\u201c\u5317\u4eac\u5e02\u671d\u9633\u533a\u5efa\u56fd\u8def88\u53f7\u201d\u4e0e\u201c\u5317\u4eac\u671d\u9633\u5efa\u56fd\u8def88\u53f7\u201d\u662f\u5426\u4e3a\u540c\u4e00\u5730\u70b9？\u8fd9\u7c7b\u95ee\u9898\u7684\u81ea\u52a8\u5316\u8bc6\u522b\u9700\u6c42\u65e5\u76ca\u8feb\u5207\u3002\u4f20\u7edf\u65b9\u6cd5\u4f9d\u8d56\u89c4\u5219\u5339\u914d\u6216\u8f7b\u91cf\u7ea7\u6587\u672c\u76f8\u4f3c\u5ea6\u7b97\u6cd5（\u5982\u7f16\u8f91\u8ddd\u79bb\u3001Jaccard），\u4f46\u96be\u4ee5\u5e94\u5bf9\u4e2d\u6587\u5730\u5740\u590d\u6742\u7684\u7f29\u5199\u3001\u8bed\u5e8f\u53d8\u5316\u548c\u522b\u540d\u8868\u8fbe\u3002<\/p>\n
\u963f\u91cc\u4e91\u8fd1\u671f\u5f00\u6e90\u7684 MGeo \u5730\u5740\u76f8\u4f3c\u5ea6\u5339\u914d\u6a21\u578b \u6b63\u662f\u4e3a\u6b64\u7c7b\u6311\u6218\u800c\u751f\u3002\u8be5\u6a21\u578b\u4e13\u4e3a\u4e2d\u6587\u5730\u5740\u8bed\u4e49\u7406\u89e3\u8bbe\u8ba1，\u57fa\u4e8e\u5927\u89c4\u6a21\u771f\u5b9e\u5730\u5740\u5bf9\u8fdb\u884c\u8bad\u7ec3，\u80fd\u591f\u7cbe\u51c6\u5224\u65ad\u4e24\u4e2a\u5730\u5740\u63cf\u8ff0\u662f\u5426\u6307\u5411\u540c\u4e00\u5730\u7406\u4f4d\u7f6e\u3002\u66f4\u5173\u952e\u7684\u662f，MGeo \u5728\u63a8\u7406\u9636\u6bb5\u5c55\u73b0\u51fa\u6781\u5f3a\u7684\u786c\u4ef6\u9002\u5e94\u6027\u2014\u2014\u4ec5\u9700\u4e00\u5f20\u6d88\u8d39\u7ea7 GPU（\u5982RTX 4090D）\u5373\u53ef\u9ad8\u6548\u8fd0\u884c，\u6781\u5927\u964d\u4f4e\u4e86\u4e2d\u5c0f\u578b\u4f01\u4e1a\u90e8\u7f72\u9ad8\u7cbe\u5ea6\u5730\u5740\u5339\u914d\u670d\u52a1\u7684\u6280\u672f\u95e8\u69db\u3002<\/p>\n
\u672c\u6587\u5c06\u56f4\u7ed5 MGeo \u7684\u5b9e\u9645\u90e8\u7f72\u4e0e\u5e94\u7528\u5c55\u5f00，\u91cd\u70b9\u4ecb\u7ecd\u5982\u4f55\u5728\u5355\u5361\u73af\u5883\u4e0b\u5feb\u901f\u642d\u5efa\u53ef\u8fd0\u884c\u7684\u670d\u52a1\u539f\u578b，\u5e76\u63d0\u4f9b\u5b8c\u6574\u7684\u64cd\u4f5c\u8def\u5f84\u4e0e\u4f18\u5316\u5efa\u8bae，\u5e2e\u52a9\u5f00\u53d1\u8005\u4ee5\u6700\u4f4e\u6210\u672c\u5b9e\u73b0\u751f\u4ea7\u7ea7\u5730\u5740\u53bb\u91cd\u4e0e\u5bf9\u9f50\u80fd\u529b\u3002<\/p>\n
\u4e2d\u6587\u5730\u5740\u5177\u6709\u9ad8\u5ea6\u7ed3\u6784\u5316\u4f46\u53c8\u7075\u6d3b\u591a\u53d8\u7684\u7279\u70b9：<\/p>\n
\u8fd9\u4e9b\u7279\u6027\u4f7f\u5f97\u901a\u7528\u8bed\u4e49\u6a21\u578b（\u5982BERT）\u5728\u672a\u7ecf\u9886\u57df\u5fae\u8c03\u7684\u60c5\u51b5\u4e0b\u8868\u73b0\u4e0d\u4f73\u3002\u800c MGeo \u7684\u6838\u5fc3\u4f18\u52bf\u5728\u4e8e\u5176\u6df1\u5ea6\u9002\u914d\u4e2d\u6587\u5730\u5740\u8bed\u8a00\u6a21\u5f0f，\u901a\u8fc7\u4ee5\u4e0b\u673a\u5236\u63d0\u5347\u5339\u914d\u51c6\u786e\u7387：<\/p>\n
\u6280\u672f\u4ef7\u503c\u603b\u7ed3：MGeo \u5e76\u975e\u7b80\u5355\u590d\u523b\u901a\u7528NLP\u6a21\u578b，\u800c\u662f\u9488\u5bf9\u4e2d\u6587\u5730\u5740\u8bed\u4e49\u6b67\u4e49\u6027\u5f3a\u3001\u683c\u5f0f\u975e\u6807\u51c6\u5316\u7684\u95ee\u9898，\u63d0\u51fa\u4e86\u4e00\u5957\u201c\u7ed3\u6784+\u8bed\u4e49+\u77e5\u8bc6\u201d\u4e09\u4f4d\u4e00\u4f53\u7684\u89e3\u51b3\u65b9\u6848\u3002<\/p>\n
\u963f\u91cc\u5c06 MGeo \u5f00\u6e90，\u610f\u5473\u7740\u4f01\u4e1a\u4e0d\u518d\u9700\u8981\u4ece\u96f6\u5f00\u59cb\u6536\u96c6\u6570\u767e\u4e07\u6761\u6807\u6ce8\u5730\u5740\u5bf9\u6765\u8bad\u7ec3\u6a21\u578b\u3002\u8fd9\u4e0d\u4ec5\u8282\u7701\u4e86\u6570\u636e\u6807\u6ce8\u6210\u672c，\u4e5f\u907f\u514d\u4e86\u56e0\u6837\u672c\u4e0d\u8db3\u5bfc\u81f4\u7684\u6cdb\u5316\u80fd\u529b\u5dee\u95ee\u9898\u3002\u66f4\u91cd\u8981\u7684\u662f，\u5b98\u65b9\u63d0\u4f9b\u4e86\u5b8c\u6574\u7684\u63a8\u7406\u811a\u672c\u548c\u5bb9\u5668\u955c\u50cf，\u6781\u5927\u7b80\u5316\u4e86\u90e8\u7f72\u6d41\u7a0b\u3002<\/p>\n
\u5bf9\u4e8e\u4e2d\u5c0f\u89c4\u6a21\u4e1a\u52a1（\u65e5\u5747\u5730\u5740\u5339\u914d\u8bf7\u6c42 < 50\u4e07\u6b21），MGeo \u5c55\u73b0\u51fa\u51fa\u8272\u7684\u6027\u4ef7\u6bd4：\u65e0\u9700\u6602\u8d35\u7684\u591a\u5361A100\u96c6\u7fa4，\u4e00\u5f20RTX 4090D\u5373\u53ef\u652f\u6491\u5168\u5929\u5019\u670d\u52a1\u3002\u8fd9\u5bf9\u4e8e\u521d\u521b\u516c\u53f8\u3001\u5730\u65b9\u653f\u52a1\u5e73\u53f0\u6216\u533a\u57df\u6027\u7269\u6d41\u4f01\u4e1a\u800c\u8a00，\u662f\u4e00\u6b21\u771f\u6b63\u610f\u4e49\u4e0a\u7684\u6280\u672f\u666e\u60e0\u3002<\/p>\n
\u672c\u8282\u5c06\u624b\u628a\u624b\u6f14\u793a\u5982\u4f55\u5728\u914d\u5907 RTX 4090D \u7684\u670d\u52a1\u5668\u4e0a\u90e8\u7f72 MGeo \u6a21\u578b，\u5b8c\u6210\u4ece\u73af\u5883\u51c6\u5907\u5230\u9996\u6b21\u63a8\u7406\u7684\u5168\u8fc7\u7a0b\u3002\u6574\u4e2a\u8fc7\u7a0b\u63a7\u5236\u572810\u5206\u949f\u5185，\u9002\u5408\u5feb\u901f\u9a8c\u8bc1\u4e0e\u539f\u578b\u5f00\u53d1\u3002<\/p>\n
\u5047\u8bbe\u4f60\u5df2\u62e5\u6709\u4e00\u53f0\u5b89\u88c5\u4e86 NVIDIA \u9a71\u52a8\u548c Docker \u7684 Linux \u670d\u52a1\u5668（\u63a8\u8350 Ubuntu 20.04+），\u4e14 GPU \u652f\u6301 CUDA 11.8 \u6216\u4ee5\u4e0a\u7248\u672c\u3002<\/p>\n
docker pull registry.cn-beijing.aliyuncs.com\/mgeo\/mgeo-inference:latest<\/p>\n
docker run -it \\\\
\n –gpus all \\\\
\n -p 8888:8888 \\\\
\n -v \/your\/local\/workspace:\/root\/workspace \\\\
\n –name mgeo-container \\\\
\n registry.cn-beijing.aliyuncs.com\/mgeo\/mgeo-inference:latest<\/p>\n
-v \u6302\u8f7d\u672c\u5730\u76ee\u5f55\u7528\u4e8e\u6301\u4e45\u5316\u4ee3\u7801\u4e0e\u7ed3\u679c<\/p>\n<\/li>\n
\u8fdb\u5165\u5bb9\u5668\u7ec8\u7aef<\/p>\n<\/li>\n<\/ul>\n
docker exec -it mgeo-container \/bin\/bash<\/p>\n
\u5bb9\u5668\u5185\u5df2\u9884\u88c5 Conda \u73af\u5883\u4e0e\u4f9d\u8d56\u5e93，\u6309\u7167\u4ee5\u4e0b\u6b65\u9aa4\u6267\u884c\u5373\u53ef：<\/p>\n
conda activate py37testmaas<\/p>\n
\u8be5\u73af\u5883\u5305\u542b PyTorch 1.12 + Transformers + CUDA \u652f\u6301，\u4e13\u4e3a MGeo \u63a8\u7406\u4f18\u5316\u3002<\/p>\n
python \/root\/\u63a8\u7406.py<\/p>\n
\u6b64\u811a\u672c\u4f1a\u52a0\u8f7d\u9884\u8bad\u7ec3\u6a21\u578b，\u5e76\u5bf9 \/root\/test_cases.json \u4e2d\u7684\u5730\u5740\u5bf9\u8fdb\u884c\u6279\u91cf\u76f8\u4f3c\u5ea6\u6253\u5206\u3002\u8f93\u51fa\u683c\u5f0f\u5982\u4e0b：<\/p>\n
[
\n {
\n "addr1": "\u5317\u4eac\u5e02\u6d77\u6dc0\u533a\u4e2d\u5173\u6751\u5927\u88571\u53f7",
\n "addr2": "\u5317\u4eac\u6d77\u6dc0\u4e2d\u5173\u6751\u5927\u88571\u53f7\u5927\u53a6",
\n "score": 0.96,
\n "is_match": true
\n },
\n …
\n]<\/p>\n
\u5176\u4e2d score \u4e3a [0,1] \u533a\u95f4\u5185\u7684\u76f8\u4f3c\u5ea6\u5206\u6570，is_match \u7531\u9608\u503c（\u9ed8\u8ba40.85）\u51b3\u5b9a\u3002<\/p>\n
cp \/root\/\u63a8\u7406.py \/root\/workspace<\/p>\n
\u6b64\u4e3e\u5c06\u811a\u672c\u590d\u5236\u5230\u6302\u8f7d\u7684\u5de5\u4f5c\u76ee\u5f55，\u4f60\u53ef\u4ee5\u5728\u5bbf\u4e3b\u673a\u4e0a\u4f7f\u7528 VS Code \u7b49\u5de5\u5177\u76f4\u63a5\u7f16\u8f91，\u5b9e\u73b0\u53ef\u89c6\u5316\u5f00\u53d1\u3002<\/p>\n
\u4ee5\u4e0b\u662f \u63a8\u7406.py \u7684\u6838\u5fc3\u4ee3\u7801\u7247\u6bb5\u89e3\u6790，\u4f9b\u4e8c\u6b21\u5f00\u53d1\u53c2\u8003：<\/p>\n
# -*- coding: utf-8 -*-
\nimport torch
\nfrom transformers import AutoTokenizer, AutoModelForSequenceClassification<\/p>\n
# \u52a0\u8f7d tokenizer \u548c\u6a21\u578b
\nmodel_path = "\/root\/models\/mgeo-chinese-address-v1"
\ntokenizer = AutoTokenizer.from_pretrained(model_path)
\nmodel = AutoModelForSequenceClassification.from_pretrained(model_path)<\/p>\n
# \u79fb\u52a8\u6a21\u578b\u5230 GPU
\ndevice = torch.device("cuda" if torch.cuda.is_available() else "cpu")
\nmodel.to(device)
\nmodel.eval()<\/p>\n
def compute_similarity(addr1, addr2):
\n inputs = tokenizer(
\n addr1, addr2,
\n padding=True,
\n truncation=True,
\n max_length=128,
\n return_tensors="pt"
\n ).to(device)<\/p>\n
with torch.no_grad():
\n outputs = model(**inputs)
\n probs = torch.softmax(outputs.logits, dim=-1)
\n match_prob = probs[0][1].item() # \u6b63\u7c7b\u6982\u7387<\/p>\n
return match_prob<\/p>\n
# \u793a\u4f8b\u8c03\u7528
\naddr_a = "\u4e0a\u6d77\u5e02\u6d66\u4e1c\u65b0\u533a\u5f20\u6c5f\u9ad8\u79d1\u6280\u56ed\u533a\u79d1\u82d1\u8def88\u53f7"
\naddr_b = "\u4e0a\u6d77\u6d66\u4e1c\u5f20\u6c5f\u79d1\u82d1\u8def88\u53f7"
\nscore = compute_similarity(addr_a, addr_b)
\nprint(f"\u76f8\u4f3c\u5ea6\u5f97\u5206: {score:.3f}")<\/p>\n
| \u53c2\u6570 | \u503c | \u8bf4\u660e |
\n|——|—–|——-|
\n| max_length | 128 | \u8986\u76d6\u7edd\u5927\u591a\u6570\u4e2d\u6587\u5730\u5740\u957f\u5ea6 |
\n| padding\/truncation | True | \u6279\u91cf\u63a8\u7406\u65f6\u81ea\u52a8\u5bf9\u9f50\u5f20\u91cf\u5c3a\u5bf8 |
\n| return_tensors | "pt" | \u8fd4\u56de PyTorch \u5f20\u91cf |
\n| softmax(logits)[:,1] | \u53d6\u6b63\u7c7b\u6982\u7387 | \u5206\u7c7b\u5934\u8f93\u51fa[\u4e0d\u5339\u914d, \u5339\u914d]，\u53d6\u7b2c\u4e8c\u9879 |<\/p>\n
\u6027\u80fd\u63d0\u793a：\u5728 RTX 4090D \u4e0a，\u5355\u6761\u5730\u5740\u5bf9\u63a8\u7406\u8017\u65f6\u7ea6 18ms（\u542b tokenize），\u6279\u91cf\u5904\u7406（batch_size=32）\u53ef\u8fbe 1200 QPS，\u5b8c\u5168\u6ee1\u8db3\u4e2d\u5c0f\u4e1a\u52a1\u5b9e\u65f6\u54cd\u5e94\u9700\u6c42\u3002<\/p>\n
\u867d\u7136 MGeo \u9ed8\u8ba4\u914d\u7f6e\u5df2\u5728\u5355\u5361\u4e0a\u8868\u73b0\u4f18\u5f02，\u4f46\u5728\u5b9e\u9645\u751f\u4ea7\u73af\u5883\u4e2d\u4ecd\u53ef\u901a\u8fc7\u4ee5\u4e0b\u65b9\u5f0f\u8fdb\u4e00\u6b65\u63d0\u5347\u6548\u7387\u4e0e\u7a33\u5b9a\u6027\u3002<\/p>\n
\u539f\u59cb\u811a\u672c\u91c7\u7528\u9010\u6761\u63a8\u7406\u65b9\u5f0f，\u672a\u5145\u5206\u53d1\u6325 GPU \u5e76\u884c\u8ba1\u7b97\u80fd\u529b\u3002\u5efa\u8bae\u6539\u9020\u6210\u6279\u91cf\u5904\u7406\u6a21\u5f0f：<\/p>\n
def batch_similarity(address_pairs, batch_size=32):
\n results = []
\n for i in range(0, len(address_pairs), batch_size):
\n batch = address_pairs[i:i+batch_size]
\n addr1_list = [pair[0] for pair in batch]
\n addr2_list = [pair[1] for pair in batch]<\/p>\n
inputs = tokenizer(addr1_list, addr2_list, … , return_tensors="pt").to(device)<\/p>\n
with torch.no_grad():
\n logits = model(**inputs).logits
\n probs = torch.softmax(logits, dim=1)[:,1].cpu().numpy()<\/p>\n
results.extend(probs)
\n return results<\/p>\n
\u5728\u5b9e\u9645\u4e1a\u52a1\u4e2d，\u90e8\u5206\u5730\u5740（\u5982\u70ed\u95e8\u5546\u5708\u3001\u603b\u90e8\u5927\u697c）\u4f1a\u88ab\u9891\u7e41\u67e5\u8be2\u3002\u5f15\u5165 Redis \u7f13\u5b58\u53ef\u663e\u8457\u964d\u4f4e\u91cd\u590d\u8ba1\u7b97\u5f00\u9500：<\/p>\n
import hashlib
\nimport redis<\/p>\n
r = redis.Redis(host='localhost', port=6379, db=0)<\/p>\n
def get_cached_score(addr1, addr2):
\n key = hashlib.md5(f"{addr1}_{addr2}".encode()).hexdigest()
\n cached = r.get(key)
\n if cached:
\n return float(cached)
\n score = compute_similarity(addr1, addr2)
\n r.setex(key, 3600, str(score)) # \u7f13\u5b581\u5c0f\u65f6
\n return score<\/p>\n
\u5c06\u63a8\u7406\u903b\u8f91\u5c01\u88c5\u4e3a REST API，\u4fbf\u4e8e\u7cfb\u7edf\u96c6\u6210：<\/p>\n
from flask import Flask, request, jsonify<\/p>\n
app = Flask(__name__)<\/p>\n
@app.route('\/match', methods=['POST'])
\ndef match():
\n data = request.json
\n addr1 = data['address1']
\n addr2 = data['address2']
\n threshold = data.get('threshold', 0.85)<\/p>\n
score = compute_similarity(addr1, addr2)
\n is_match = score >= threshold<\/p>\n
return jsonify({
\n 'score': round(score, 4),
\n 'is_match': is_match
\n })<\/p>\n
if __name__ == '__main__':
\n app.run(host='0.0.0.0', port=5000)<\/p>\n
\u914d\u5408 Gunicorn + Nginx \u53ef\u8f7b\u677e\u652f\u6301\u9ad8\u5e76\u53d1\u8bbf\u95ee\u3002<\/p>\n
\u4e3a\u4e86\u66f4\u6e05\u6670\u5730\u5c55\u793a MGeo \u7684\u5b9a\u4f4d\u4f18\u52bf，\u6211\u4eec\u5c06\u5176\u4e0e\u4e09\u79cd\u5e38\u89c1\u65b9\u6848\u8fdb\u884c\u6a2a\u5411\u5bf9\u6bd4：<\/p>\n
| \u65b9\u6848 | \u51c6\u786e\u7387（F1） | \u5355\u6b21\u63a8\u7406\u8017\u65f6 | \u786c\u4ef6\u8981\u6c42 | \u662f\u5426\u652f\u6301\u4e2d\u6587 |
\n|——|————-|————–|———-|————-|
\n| MGeo（\u672c\u6a21\u578b） | 0.94 | 18ms | \u5355\u5361GPU（\u59824090D） | \u2705 \u6df1\u5ea6\u4f18\u5316 |
\n| \u7f16\u8f91\u8ddd\u79bb（Levenshtein） | 0.62 | <1ms | CPU\u5373\u53ef | \u274c \u5ffd\u7565\u8bed\u4e49 |
\n| SimHash + Jaccard | 0.68 | 2ms | CPU\u5373\u53ef | \u26a0\ufe0f \u4ec5\u8868\u5c42\u5339\u914d |
\n| \u5fae\u8c03BERT-base | 0.87 | 45ms | \u591a\u5361\u8bad\u7ec3，\u5355\u5361\u63a8\u7406 | \u2705 \u53ef\u9002\u914d |<\/p>\n
\u6ce8：\u6d4b\u8bd5\u96c6\u4e3a\u516c\u5f00\u4e2d\u6587\u5730\u5740\u5bf9\u9f50\u6570\u636e\u96c6 CADEC（Chinese Address Deduplication Corpus）<\/p>\n
| \u4e1a\u52a1\u573a\u666f | \u63a8\u8350\u65b9\u6848 | \u7406\u7531 |
\n|———|———-|——|
\n| \u9ad8\u7cbe\u5ea6\u5339\u914d（\u5982\u91d1\u878d\u5f00\u6237\u6838\u9a8c） | MGeo | \u6700\u9ad8F1\u503c，\u652f\u6301\u590d\u6742\u522b\u540d |
\n| \u4f4e\u5ef6\u8fdf\u8fc7\u6ee4（\u5982\u65e5\u5fd7\u6e05\u6d17\u9884\u7b5b） | SimHash | \u901f\u5ea6\u5feb，\u9002\u5408\u7c97\u7b5b |
\n| \u5df2\u6709NLP\u5e73\u53f0\u7684\u4f01\u4e1a | \u5fae\u8c03BERT | \u53ef\u590d\u7528\u73b0\u6709\u67b6\u6784 |
\n| \u65e0GPU\u8d44\u6e90\u7684\u5c0f\u578b\u9879\u76ee | \u7f16\u8f91\u8ddd\u79bb+\u89c4\u5219 | \u96f6\u4f9d\u8d56，\u6613\u90e8\u7f72 |<\/p>\n
\u53ef\u4ee5\u770b\u51fa，MGeo \u5728\u201c\u7cbe\u5ea6\u4f18\u5148\u201d\u4e14\u5177\u5907\u57fa\u7840GPU\u6761\u4ef6\u7684\u573a\u666f\u4e2d\u5177\u6709\u7edd\u5bf9\u4f18\u52bf\u3002\u800c\u5bf9\u4e8e\u7eafCPU\u73af\u5883，\u53ef\u8003\u8651\u5148\u7528 SimHash \u505a\u521d\u6b65\u805a\u7c7b，\u518d\u5c06\u5019\u9009\u5bf9\u9001\u5165 MGeo \u7cbe\u6392，\u5f62\u6210\u6df7\u5408\u6d41\u6c34\u7ebf\u3002<\/p>\n
MGeo \u7684\u5f00\u6e90\u6807\u5fd7\u7740\u4e2d\u6587\u5730\u5740\u7406\u89e3\u6280\u672f\u6b63\u4ece\u201c\u5927\u5382\u4e13\u5c5e\u201d\u8d70\u5411\u201c\u666e\u60e0\u53ef\u7528\u201d\u3002\u901a\u8fc7\u672c\u6587\u7684\u5b9e\u8df5\u53ef\u4ee5\u770b\u51fa，\u5373\u4fbf\u53ea\u6709\u4e00\u5f20 RTX 4090D \u8fd9\u6837\u7684\u6d88\u8d39\u7ea7\u663e\u5361，\u4e5f\u80fd\u6784\u5efa\u51fa\u7a33\u5b9a\u9ad8\u6548\u7684\u5730\u5740\u5339\u914d\u670d\u52a1\u3002<\/p>\n
\u6700\u7ec8\u7ed3\u8bba：MGeo \u4e0d\u4ec5\u662f\u4e00\u4e2a\u6a21\u578b，\u66f4\u662f\u4e00\u79cd\u201c\u5c0f\u800c\u7f8e\u201d\u7684\u5730\u7406\u667a\u80fd\u843d\u5730\u8303\u5f0f\u2014\u2014\u65e0\u9700\u8ffd\u9010\u9876\u7ea7\u7b97\u529b，\u4e5f\u80fd\u83b7\u5f97\u5de5\u4e1a\u7ea7\u6548\u679c\u3002\u5bf9\u4e8e\u5e7f\u5927\u4e2d\u5c0f\u4f01\u4e1a\u800c\u8a00，\u8fd9\u6b63\u662f AI \u8d4b\u80fd\u4e1a\u52a1\u7684\u771f\u5b9e\u8d77\u70b9\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"
\u6587\u7ae0\u6d4f\u89c8\u9605\u8bfb780\u6b21\uff0c\u70b9\u8d5e27\u6b21\uff0c\u6536\u85cf13\u6b21\u3002\u4ee5\u4e0b\u662f\u63a8\u7406.py# \u52a0\u8f7d tokenizer \u548c\u6a21\u578b# \u79fb\u52a8\u6a21\u578b\u5230 GPUmatch_prob = probs[0][1].item() # \u6b63\u7c7b\u6982\u7387# \u793a\u4f8b\u8c03\u7528addr_a = "\u4e0a\u6d77\u5e02\u6d66\u4e1c\u65b0\u533a\u5f20\u6c5f\u9ad8\u79d1\u6280\u56ed\u533a\u79d1\u82d1\u8def88\u53f7"addr_b = "\u4e0a\u6d77\u6d66\u4e1c\u5f20\u6c5f\u79d1\u82d1\u8def88\u53f7"print(f"\u76f8\u4f3c\u5ea6\u5f97\u5206: {score:.3f}")MGeo \u7684\u5f00\u6e90\u6807\u5fd7\u7740\u4e2d\u6587\u5730\u5740\u7406\u89e3\u6280\u672f\u6b63\u4ece\u201c\u5927\u5382\u4e13\u5c5e\u201d\u8d70\u5411\u201c\u666e\u60e0\u53ef\u7528\u201d\u3002<\/p>\n","protected":false},"author":2,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[],"topic":[],"class_list":["post-59307","post","type-post","status-publish","format-standard","hentry","category-server"],"yoast_head":"\n