{"id":10021029,"date":"2025-10-13T19:00:07","date_gmt":"2025-10-13T11:00:07","guid":{"rendered":"https:\/\/itmax.cn\/?p=10021029"},"modified":"2025-10-13T19:00:07","modified_gmt":"2025-10-13T11:00:07","slug":"%e4%bd%bf%e7%94%a8-onnx-%e5%b0%86-ai-%e6%8e%a8%e7%90%86%e5%bc%95%e5%85%a5-java%ef%bc%9a%e4%bc%81%e4%b8%9a%e6%9e%b6%e6%9e%84%e5%b8%88%e5%ae%9e%e7%94%a8%e6%8c%87%e5%8d%97","status":"publish","type":"post","link":"https:\/\/itmax.cn\/?p=10021029","title":{"rendered":"\u4f7f\u7528 ONNX \u5c06 AI \u63a8\u7406\u5f15\u5165 Java\uff1a\u4f01\u4e1a\u67b6\u6784\u5e08\u5b9e\u7528\u6307\u5357"},"content":{"rendered":"<h2>\u5f15\u8a00<\/h2>\n<p>\u867d\u7136 Python \u4e3b\u5bfc\u4e86\u673a\u5668\u5b66\u4e60\u751f\u6001\uff0c\u4f46\u5927\u591a\u6570\u4f01\u4e1a\u5e94\u7528\u4ecd\u8fd0\u884c\u5728 Java \u4e0a\u3002\u8fd9\u79cd\u8131\u8282\u9020\u6210\u4e86\u90e8\u7f72\u74f6\u9888\u3002\u7528 <a href=\"https:\/\/link.segmentfault.com\/?enc=8RxYjNh9%2FxD17cixGSTZiA%3D%3D.GIMKpEPuvdWC6zzHiIifTZpOY0OIeFMduT58a0VgG4U%3D\" rel=\"nofollow noopener noreferrer\" target=\"_blank\">PyTorch<\/a> \u6216 <a href=\"https:\/\/link.segmentfault.com\/?enc=5s3TTLQE5CH6XTlXixn5kQ%3D%3D.NtY6UkVqyJE%2FqxZK2%2FjTCMHyycBGwh9m6AgWJCRax4c%3D\" rel=\"nofollow noopener noreferrer\" target=\"_blank\">Hugging Face<\/a> \u8bad\u7ec3\u7684\u6a21\u578b\u5728\u751f\u4ea7\u4e2d\u5f80\u5f80\u9700\u8981 REST \u5c01\u88c5\u3001\u5fae\u670d\u52a1\u6216\u591a\u8bed\u8a00\u53d8\u901a\u65b9\u5f0f\u624d\u80fd\u8fd0\u884c\u3002\u8fd9\u4e9b\u505a\u6cd5\u4f1a\u589e\u52a0\u5ef6\u8fdf\u3001\u63d0\u9ad8\u590d\u6742\u5ea6\uff0c\u5e76\u524a\u5f31\u5bf9\u7cfb\u7edf\u7684\u63a7\u5236\u529b\u3002<\/p>\n<p>\u5bf9\u4f01\u4e1a\u67b6\u6784\u5e08\u800c\u8a00\uff0c\u8fd9\u4e2a\u6311\u6218\u5e76\u4e0d\u964c\u751f\uff1a\u5982\u4f55\u5728\u4e0d\u7834\u574f Java \u7cfb\u7edf\u7684\u7b80\u5355\u6027\u3001\u53ef\u89c2\u6d4b\u6027\u548c\u53ef\u9760\u6027\u7684\u524d\u63d0\u4e0b\u5f15\u5165\u73b0\u4ee3 AI\uff1f\u8fd9\u4e2a\u6311\u6218\u5ef6\u7eed\u4e86\u6211\u4eec\u6b64\u524d\u5173\u4e8e<a href=\"https:\/\/link.segmentfault.com\/?enc=JmOqO6jLjVvntlnU1gp8gQ%3D%3D.BurnLoBS%2F1d%2FOXe2%2Bc6mgCUZRbcPPWW0IhlR7fd4q5E%3D\" rel=\"nofollow noopener noreferrer\" target=\"_blank\">\u5c06 GPU \u7ea7\u6027\u80fd\u5e26\u5165\u4f01\u4e1a Java<\/a>\u7684\u63a2\u7d22\uff0c\u5176\u4e2d\u4fdd\u6301 JVM \u539f\u751f\u7684\u6548\u7387\u4e0e\u53ef\u63a7\u6027\u88ab\u8bc1\u660e\u81f3\u5173\u91cd\u8981\u3002<\/p>\n<p><a href=\"https:\/\/link.segmentfault.com\/?enc=q%2BoayljMAgvzS1wjc1XW4Q%3D%3D.RcUsZTwYYYILAWfkQ1BkFaBKyFmxfKWzkJCYiOMnhwo%3D\" rel=\"nofollow noopener noreferrer\" target=\"_blank\">\u5f00\u653e\u795e\u7ecf\u7f51\u7edc\u4ea4\u6362\uff08ONNX\uff09<\/a> \u6807\u51c6\u7ed9\u51fa\u4e86\u6709\u529b\u7b54\u6848\u3002\u5728\u5fae\u8f6f\u7684\u652f\u6301\u4e0b\u5e76\u88ab\u5404\u5927\u6846\u67b6\u5e7f\u6cdb\u91c7\u7528\uff0cONNX \u80fd\u8ba9\u5305\u542b\u547d\u540d\u5b9e\u4f53\u8bc6\u522b\uff08NER\uff09\u3001\u5206\u7c7b\u3001\u60c5\u611f\u5206\u6790\u7b49\u5728\u5185\u7684 Transformer \u63a8\u7406\u539f\u751f\u5730\u5728 JVM \u4e2d\u8fd0\u884c\u3002\u65e0\u9700 Python \u8fdb\u7a0b\uff0c\u65e0\u9700\u5bb9\u5668\u8513\u5ef6\u3002<\/p>\n<p>\u672c\u6587\u9762\u5411\u5e0c\u671b\u5c06\u673a\u5668\u5b66\u4e60\u63a8\u7406\u5f15\u5165 Java \u751f\u4ea7\u7cfb\u7edf\u7684\u67b6\u6784\u5e08\uff0c\u63d0\u4f9b\u8bbe\u8ba1\u5c42\u9762\u7684\u6307\u5357\u3002\u6211\u4eec\u5c06\u63a2\u8ba8\u5206\u8bcd\u5668\u96c6\u6210\u3001GPU \u52a0\u901f\u3001\u90e8\u7f72\u6a21\u5f0f\u4ee5\u53ca\u5728\u53d7\u76d1\u7ba1\u7684 Java \u73af\u5883\u4e2d\u5b89\u5168\u3001\u53ef\u6269\u5c55\u5730\u8fd0\u884c AI \u7684\u751f\u547d\u5468\u671f\u7b56\u7565\u3002<\/p>\n<h2>\u4e3a\u4ec0\u4e48\u5bf9\u67b6\u6784\u5e08\u91cd\u8981<\/h2>\n<p>\u4f01\u4e1a\u7cfb\u7edf\u65e5\u76ca\u9700\u8981 AI \u6765\u9a71\u52a8\u5ba2\u6237\u4f53\u9a8c\u3001\u81ea\u52a8\u5316\u5de5\u4f5c\u6d41\u3001\u5e76\u4ece\u975e\u7ed3\u6784\u5316\u6570\u636e\u4e2d\u63d0\u53d6\u6d1e\u89c1\u3002\u4f46\u5728\u91d1\u878d\u3001\u533b\u7597\u7b49\u53d7\u76d1\u7ba1\u9886\u57df\uff0c\u751f\u4ea7\u73af\u5883\u66f4\u770b\u91cd\u53ef\u5ba1\u8ba1\u6027\u3001\u8d44\u6e90\u63a7\u5236\u4e0e JVM \u539f\u751f\u5de5\u5177\u94fe\u3002<\/p>\n<p>\u867d\u7136 Python \u5728\u8bd5\u9a8c\u4e0e\u8bad\u7ec3\u9636\u6bb5\u8868\u73b0\u51fa\u8272\uff0c\u4f46\u90e8\u7f72\u5230 Java \u7cfb\u7edf\u65f6\u4f1a\u5e26\u6765\u67b6\u6784\u6469\u64e6\u3002\u5c06\u6a21\u578b\u5305\u88c5\u4e3a Python \u5fae\u670d\u52a1\u4f1a\u5272\u88c2\u53ef\u89c2\u6d4b\u6027\u3001\u6269\u5927\u653b\u51fb\u9762\uff0c\u5e76\u5f15\u5165\u8fd0\u884c\u65f6\u4e0d\u4e00\u81f4\u3002<\/p>\n<p>ONNX \u6539\u53d8\u4e86\u8fd9\u79cd\u5c40\u9762\u3002\u5b83\u4e3a\u5728 Python \u4e2d\u8bad\u7ec3\u5e76\u5728 Java \u5185\u8fd0\u884c\u7684\u6a21\u578b\u63d0\u4f9b\u6807\u51c6\u5316\u7684\u5bfc\u51fa\u683c\u5f0f\uff0c\u539f\u751f\u652f\u6301 GPU \u52a0\u901f\u4e14\u4e0d\u4f9d\u8d56\u5916\u90e8\u8fd0\u884c\u65f6\u3002\u5173\u4e8e\u5728 JVM \u5185\u76f4\u63a5\u5229\u7528 GPU \u52a0\u901f\u7684\u66f4\u591a\u6a21\u5f0f\uff0c\u8bf7\u53c2\u89c1\uff1a<a href=\"https:\/\/link.segmentfault.com\/?enc=%2BUfQSUH4XAZWcVXAGEa%2FJw%3D%3D.FG8GbYGOlRqL7kztNiYi%2FMhnTPWZHGGZV0kt%2B%2Bei8Hc%3D\" rel=\"nofollow noopener noreferrer\" target=\"_blank\">\u5c06 GPU \u7ea7\u6027\u80fd\u5e26\u5165\u4f01\u4e1a Java<\/a>\u3002<\/p>\n<p>\u5bf9\u67b6\u6784\u5e08\u800c\u8a00\uff0cONNX \u89e3\u9501\u56db\u9879\u5173\u952e\u6536\u76ca\uff1a<\/p>\n<ul>\n<li>\u8bed\u8a00\u4e00\u81f4\u6027\uff1a\u63a8\u7406\u8fd0\u884c\u5728 JVM \u5185\uff0c\u800c\u4e0d\u662f\u65c1\u8def\u7684\u8fdb\u7a0b\u3002<\/li>\n<li>\u90e8\u7f72\u7b80\u5316\uff1a\u65e0\u9700\u7ba1\u7406 Python \u8fd0\u884c\u65f6\u6216 REST \u4ee3\u7406\u3002<\/li>\n<li>\u57fa\u7840\u8bbe\u65bd\u590d\u7528\uff1a\u6cbf\u7528\u73b0\u6709\u7684\u57fa\u4e8e Java \u7684\u76d1\u63a7\u3001\u8ffd\u8e2a\u4e0e\u5b89\u5168\u63a7\u5236\u3002<\/li>\n<li>\u53ef\u6269\u5c55\u6027\uff1a\u5728\u9700\u8981\u65f6\u542f\u7528 GPU \u6267\u884c\uff0c\u65e0\u9700\u91cd\u6784\u6838\u5fc3\u903b\u8f91\u3002<\/li>\n<\/ul>\n<p>\u901a\u8fc7\u6d88\u9664\u8bad\u7ec3\u4e0e\u90e8\u7f72\u4e4b\u95f4\u7684\u8fd0\u884c\u65f6\u4e0d\u5339\u914d\uff0cONNX \u8ba9\u6211\u4eec\u53ef\u4ee5\u50cf\u5bf9\u5f85\u4efb\u4f55\u53ef\u590d\u7528\u7684 Java \u6a21\u5757\u4e00\u6837\u5bf9\u5f85 AI \u63a8\u7406\uff1a\u6a21\u5757\u5316\u3001\u53ef\u89c2\u6d4b\u3001\u5e76\u53ef\u7ecf\u53d7\u751f\u4ea7\u73af\u5883\u7684\u8003\u9a8c\u3002<\/p>\n<h2>\u8bbe\u8ba1\u76ee\u6807<\/h2>\n<p>\u5728 Java \u4e2d\u8bbe\u8ba1 AI \u63a8\u7406\u4e0d\u53ea\u662f\u6a21\u578b\u51c6\u786e\u5ea6\u95ee\u9898\uff0c\u66f4\u662f\u5c06\u673a\u5668\u5b66\u4e60\u5d4c\u5165\u4f01\u4e1a\u7cfb\u7edf\u7684\u67b6\u6784\u3001\u8fd0\u8425\u4e0e\u5b89\u5168\u808c\u7406\u4e2d\u3002\u826f\u597d\u7684\u8bbe\u8ba1\u4f1a\u8bbe\u5b9a\u7cfb\u7edf\u7ea7\u76ee\u6807\uff0c\u786e\u4fdd AI \u7684\u91c7\u7528\u5728\u5404\u73af\u5883\u4e2d\u53ef\u6301\u7eed\u3001\u53ef\u6d4b\u8bd5\u3001\u4e14\u7b26\u5408\u6cd5\u89c4\u7ea6\u675f\u3002<\/p>\n<p>\u4ee5\u4e0b\u8bbe\u8ba1\u76ee\u6807\u6765\u81ea\u5728 Java \u4e2d\u6784\u5efa\u673a\u5668\u5b66\u4e60\u670d\u52a1\u7684\u9ad8\u7ee9\u6548\u4f01\u4e1a\u56e2\u961f\u7684\u6210\u529f\u5b9e\u8df5\uff1a<\/p>\n<p><strong>\u6d88\u9664\u751f\u4ea7\u4e2d\u7684 Python<\/strong><br \/>ONNX \u8ba9\u56e2\u961f\u53ef\u4ee5\u5c06 Python \u4e2d\u8bad\u7ec3\u7684\u6a21\u578b\u5bfc\u51fa\u5e76\u5728 Java \u4e2d\u539f\u751f\u8fd0\u884c\uff0c\u907f\u514d\u5d4c\u5165 Python \u8fd0\u884c\u65f6\u3001gRPC \u6865\u63a5\u6216\u5bb9\u5668\u5316\u7684 Python \u63a8\u7406\u670d\u52a1\uff0c\u8fd9\u4e9b\u90fd\u4f1a\u589e\u52a0\u8fd0\u7ef4\u6469\u64e6\u5e76\u590d\u6742\u5316\u5b89\u5168\u90e8\u7f72\u3002<\/p>\n<p><strong>\u652f\u6301\u53ef\u63d2\u62d4\u7684\u5206\u8bcd\u4e0e\u63a8\u7406<\/strong><br \/>\u5206\u8bcd\u5668\u4e0e\u6a21\u578b\u5e94\u5f53\u6a21\u5757\u5316\u4e14\u53ef\u914d\u7f6e\u3002\u5206\u8bcd\u5668\u6587\u4ef6\uff08\u5982 <code>tokenizer.json<\/code>\uff09\u548c\u6a21\u578b\u6587\u4ef6\uff08\u5982 <code>model.onnx<\/code>\uff09\u5e94\u53ef\u6309\u7528\u4f8b\u4e92\u6362\u3002\u901a\u8fc7\u5206\u8bcd\u5668\u4e0e\u6a21\u578b\u7684\u7ec4\u5408\uff0c\u53ef\u4ee5\u9002\u914d NER\u3001\u5206\u7c7b\u3001\u6458\u8981\u7b49\u4efb\u52a1\uff0c\u800c\u65e0\u9700\u91cd\u5199\u4ee3\u7801\u6216\u8fdd\u53cd\u6574\u6d01\u67b6\u6784\u539f\u5219\u3002<\/p>\n<p><strong>\u786e\u4fdd CPU \u2013 GPU \u7684\u7075\u6d3b\u6027<\/strong><br \/>\u76f8\u540c\u7684\u63a8\u7406\u903b\u8f91\u5e94\u5f53\u65e2\u80fd\u5728\u5f00\u53d1\u8005\u7b14\u8bb0\u672c\uff08CPU\uff09\u4e0a\u8fd0\u884c\uff0c\u4e5f\u80fd\u65e0\u6539\u52a8\u5730\u6269\u5c55\u5230\u751f\u4ea7 GPU \u96c6\u7fa4\u3002ONNX Runtime \u539f\u751f\u652f\u6301\u901a\u8fc7 CPU \u4e0e CUDA \u6267\u884c\u63d0\u4f9b\u8005\u8fd0\u884c\u63a8\u7406\uff0c\u4f7f\u8de8\u73af\u5883\u4e00\u81f4\u6027\u65e2\u53ef\u884c\u4e5f\u5177\u6210\u672c\u6548\u76ca\u3002<\/p>\n<p><strong>\u4f18\u5316\u4e3a\u53ef\u9884\u6d4b\u5ef6\u8fdf\u4e0e\u7ebf\u7a0b\u5b89\u5168<\/strong><br \/>\u63a8\u7406\u5fc5\u987b\u50cf\u4efb\u4f55\u4f01\u4e1a\u7ea7\u670d\u52a1\u4e00\u6837\uff1a\u786e\u5b9a\u6027\u3001\u7ebf\u7a0b\u5b89\u5168\u3001\u8d44\u6e90\u9ad8\u6548\u3002\u5e72\u51c0\u7684\u591a\u7ebf\u7a0b\u3001\u6a21\u578b\u9884\u52a0\u8f7d\u4e0e\u663e\u5f0f\u5185\u5b58\u63a7\u5236\u5bf9\u4e8e\u6ee1\u8db3 SLA\u3001\u589e\u5f3a\u53ef\u89c2\u6d4b\u6027\u5e76\u907f\u514d\u5e76\u53d1\u7cfb\u7edf\u4e2d\u7684\u7ade\u6001\u6761\u4ef6\u81f3\u5173\u91cd\u8981\u3002<\/p>\n<p><strong>\u8bbe\u8ba1\u4e3a\u8de8\u6808\u590d\u7528<\/strong><br \/>\u57fa\u4e8e ONNX \u7684\u63a8\u7406\u6a21\u5757\u5e94\u80fd\u591f\u6574\u6d01\u5730\u96c6\u6210\u5230 REST API\u3001\u6279\u5904\u7406\u7ba1\u9053\u3001\u4e8b\u4ef6\u9a71\u52a8\u5904\u7406\u5668\u4e0e\u5d4c\u5165\u5f0f\u5206\u6790\u5c42\u3002\u5c06\u9884\u5904\u7406\u3001\u6a21\u578b\u6267\u884c\u4e0e\u540e\u5904\u7406\u89e3\u8026\uff0c\u662f\u5b9e\u73b0\u590d\u7528\u3001\u53ef\u6d4b\u8bd5\u4e0e\u957f\u671f\u7ef4\u62a4\u5408\u89c4\u6027\u7684\u5173\u952e\u3002<\/p>\n<p>\u8fd9\u4e9b\u76ee\u6807\u5e2e\u52a9\u4f01\u4e1a\u56e2\u961f\u5728\u4e0d\u727a\u7272\u67b6\u6784\u5b8c\u6574\u6027\u3001\u5f00\u53d1\u654f\u6377\u6027\u6216\u5408\u89c4\u8981\u6c42\u7684\u60c5\u51b5\u4e0b\u91c7\u7528\u673a\u5668\u5b66\u4e60\u3002<\/p>\n<h2>\u7cfb\u7edf\u67b6\u6784\u6982\u89c8<\/h2>\n<p>\u5c06\u673a\u5668\u5b66\u4e60\u63a8\u7406\u5f15\u5165\u4f01\u4e1a\u7ea7 Java \u7cfb\u7edf\u4e0d\u4ec5\u4ec5\u662f\u6a21\u578b\u96c6\u6210\u7684\u95ee\u9898\uff0c\u66f4\u9700\u8981\u6e05\u6670\u7684\u67b6\u6784\u5206\u5c42\u4e0e\u6a21\u5757\u5316\u3002\u4e00\u4e2a\u5065\u58ee\u7684\u57fa\u4e8e ONNX \u7684\u63a8\u7406\u7cfb\u7edf\u5e94\u88ab\u8bbe\u8ba1\u4e3a\u4e00\u7ec4\u677e\u8026\u5408\u7ec4\u4ef6\uff0c\u6bcf\u4e2a\u7ec4\u4ef6\u8d1f\u8d23\u63a8\u7406\u751f\u547d\u5468\u671f\u4e2d\u7684\u7279\u5b9a\u73af\u8282\u3002<\/p>\n<p>\u6838\u5fc3\u6d41\u7a0b\u4ece\u63a5\u6536\u6765\u81ea REST \u7aef\u70b9\u3001Kafka \u6d41\u6216\u57fa\u4e8e\u6587\u4ef6\u7684\u96c6\u6210\u7b49\u591a\u79cd\u6765\u6e90\u7684\u8f93\u5165\u5f00\u59cb\u3002\u539f\u59cb\u8f93\u5165\u4f1a\u4ea4\u7531\u5206\u8bcd\u5668\u7ec4\u4ef6\u5904\u7406\uff0c\u5c06\u5176\u8f6c\u6362\u4e3a Transformer \u6a21\u578b\u6240\u9700\u7684\u6570\u503c\u683c\u5f0f\u3002\u5206\u8bcd\u5668\u901a\u8fc7\u4e00\u4e2a\u517c\u5bb9 <a href=\"https:\/\/link.segmentfault.com\/?enc=137U0vVcdbkpX%2FMBT5WbVg%3D%3D.bmA87jLy1ZgT8FZIhe8ASiToyQwD56mc%2FPvVWM%2Br8kA%3D\" rel=\"nofollow noopener noreferrer\" target=\"_blank\">Hugging Face<\/a> \u7684 <code>tokenizer.json<\/code> \u6587\u4ef6\u8fdb\u884c\u914d\u7f6e\uff0c\u786e\u4fdd\u4e0e\u8bad\u7ec3\u65f6\u4f7f\u7528\u7684\u8bcd\u8868\u4e0e\u7f16\u7801\u65b9\u5f0f\u4e00\u81f4\u3002<\/p>\n<p>\u5b8c\u6210\u5206\u8bcd\u540e\uff0c\u8f93\u5165\u6d41\u5165 ONNX \u63a8\u7406\u5f15\u64ce\u3002\u8be5\u7ec4\u4ef6\u8c03\u7528 ONNX Runtime\uff0c\u5728 CPU \u6216 GPU \u540e\u7aef\u6267\u884c\u6a21\u578b\u63a8\u7406\u3002\u5982\u679c\u53ef\u7528 GPU \u8d44\u6e90\uff0cONNX Runtime \u53ef\u65e0\u7f1d\u59d4\u6d3e\u7ed9\u57fa\u4e8e CUDA \u7684\u6267\u884c\u63d0\u4f9b\u8005\uff0c\u800c\u65e0\u9700\u66f4\u6539\u5e94\u7528\u903b\u8f91\u3002\u63a8\u7406\u5f15\u64ce\u8fd4\u56de\u4e00\u7ec4\u9884\u6d4b\uff0c\u901a\u5e38\u662f<a href=\"https:\/\/link.segmentfault.com\/?enc=0CaLtuaeTeUq5SuMdmAx5Q%3D%3D.LfbNYvWzNmQZYyXAnHNVYhK8zUteS3yvDEuWcKt0zXXqF9n7LptSZsyMM7TQCiZm\" rel=\"nofollow noopener noreferrer\" target=\"_blank\">logits<\/a>\uff08\u6a21\u578b\u7684\u539f\u59cb\u3001\u672a\u7ecf\u8fc7 softmax \u7684\u8f93\u51fa\u5206\u6570\uff09\u6216\u7c7b\u522b ID\uff0c\u968f\u540e\u7531\u540e\u5904\u7406\u6a21\u5757\u8fdb\u884c\u89e3\u91ca\u3002<\/p>\n<p>\u540e\u5904\u7406\u5668\u5c06\u539f\u59cb\u8f93\u51fa\u8f6c\u5316\u4e3a\u6709\u610f\u4e49\u7684\u9886\u57df\u5b9e\u4f53\uff0c\u5982\u6807\u7b7e\u3001\u7c7b\u522b\u6216\u62bd\u53d6\u5b57\u6bb5\u3002\u6700\u7ec8\u7ed3\u679c\u4f1a\u8def\u7531\u5230\u4e0b\u6e38\u6d88\u8d39\u8005\uff0c\u65e0\u8bba\u662f\u4e1a\u52a1\u6d41\u7a0b\u5f15\u64ce\u3001\u5173\u7cfb\u578b\u6570\u636e\u5e93\u8fd8\u662f HTTP \u54cd\u5e94\u7ba1\u9053\u3002<\/p>\n<p>\u7cfb\u7edf\u9075\u5faa\u6574\u6d01\u7684\u67b6\u6784\u6d41\uff1a\u9002\u914d\u5668\u5230\u5206\u8bcd\u5668\u3001\u5206\u8bcd\u5668\u5230\u63a8\u7406\u5f15\u64ce\u3001\u63a8\u7406\u5f15\u64ce\u5230\u540e\u5904\u7406\u5668\u3001\u540e\u5904\u7406\u5668\u5230\u6d88\u8d39\u8005\u3002\u6bcf\u4e2a\u6a21\u5757\u90fd\u53ef\u4ee5\u72ec\u7acb\u5f00\u53d1\u3001\u6d4b\u8bd5\u4e0e\u90e8\u7f72\uff0c\u4f7f\u6574\u4e2a\u7ba1\u9053\u9ad8\u5ea6\u53ef\u590d\u7528\u4e0e\u53ef\u7ef4\u62a4\u3002<\/p>\n<p><img decoding=\"async\"   referrerpolicy=\"no-referrer\" src=\"https:\/\/segmentfault.com\/img\/remote\/1460000047318623\" alt=\"\" title=\"\"><\/p>\n<p><strong>\u56fe 1\uff1aJava \u4e2d\u53ef\u63d2\u62d4\u7684 ONNX \u63a8\u7406\u67b6\u6784<\/strong><\/p>\n<p>\u901a\u8fc7\u5c06\u63a8\u7406\u89c6\u4e3a\u4e00\u6761\u5b9a\u4e49\u6e05\u6670\u7684\u8f6c\u6362\u6d41\u6c34\u7ebf\uff0c\u800c\u975e\u5d4c\u5165\u5230\u5355\u4f53\u670d\u52a1\u5185\u90e8\u7684\u903b\u8f91\uff0c\u67b6\u6784\u5e08\u53ef\u4ee5\u5bf9\u6027\u80fd\u3001\u53ef\u89c2\u6d4b\u6027\u4e0e\u90e8\u7f72\u8fdb\u884c\u7cbe\u7ec6\u5316\u63a7\u5236\u3002\u8fd9\u79cd\u6a21\u5757\u5316\u65b9\u6cd5\u4e5f\u652f\u6301\u6a21\u578b\u7684\u6f14\u8fdb\uff0c\u80fd\u591f\u5728\u4e0d\u7834\u574f\u7cfb\u7edf\u7a33\u5b9a\u6027\u7684\u60c5\u51b5\u4e0b\u66f4\u65b0\u5206\u8bcd\u5668\u6216 ONNX \u6a21\u578b\u3002<\/p>\n<h2>\u6a21\u578b\u751f\u547d\u5468\u671f<\/h2>\n<p>\u5728\u591a\u6570\u4f01\u4e1a\u573a\u666f\u4e2d\uff0c\u673a\u5668\u5b66\u4e60\u6a21\u578b\u5728 Java \u751f\u6001\u4e4b\u5916\u8fdb\u884c\u8bad\u7ec3\uff0c\u901a\u5e38\u4f7f\u7528 Hugging Face Transformers \u6216 PyTorch \u7b49 Python \u6846\u67b6\u3002\u6a21\u578b\u5b9a\u7a3f\u540e\uff0c\u4f1a\u8fde\u540c\u5176\u5206\u8bcd\u5668\u914d\u7f6e\u4e00\u8d77\u5bfc\u51fa\u4e3a ONNX \u683c\u5f0f\uff0c\u751f\u6210 <code>model.onnx<\/code> \u6587\u4ef6\u4e0e\u517c\u5bb9\u7684 <code>tokenizer.json<\/code> \u6587\u4ef6\u3002<\/p>\n<p>\u5bf9\u4e8e\u57fa\u4e8e Java \u7684\u63a8\u7406\u7cfb\u7edf\uff0c\u8fd9\u4e9b\u5de5\u4ef6\u76f8\u5f53\u4e8e\u7248\u672c\u5316\u8f93\u5165\uff0c\u7c7b\u4f3c\u5916\u90e8 JAR \u6216\u6a21\u5f0f\u6587\u4ef6\u3002\u67b6\u6784\u5e08\u5e94\u5c06\u5176\u89c6\u4e3a\u53d7\u63a7\u7684\u90e8\u7f72\u8d44\u4ea7\uff1a\u7ecf\u8fc7\u9a8c\u8bc1\u3001\u6d4b\u8bd5\uff0c\u5e76\u5728\u5404\u73af\u5883\u95f4\u6709\u5e8f\u53d1\u5e03\uff0c\u9075\u5faa\u4e0e\u4ee3\u7801\u6216\u6570\u636e\u5e93\u8fc1\u79fb\u540c\u6837\u7684\u7eaa\u5f8b\u3002<\/p>\n<p>\u53ef\u91cd\u590d\u7684\u6a21\u578b\u751f\u547d\u5468\u671f\u5305\u62ec\u5bfc\u51fa\u6a21\u578b\u4e0e\u5206\u8bcd\u5668\u3001\u7528\u4ee3\u8868\u6027\u7528\u4f8b\u8fdb\u884c\u6d4b\u8bd5\uff0c\u5e76\u5b58\u50a8\u5230\u5185\u90e8\u7684\u6ce8\u518c\u8868\u6216\u5236\u54c1\u5e93\u3002\u5728\u8fd0\u884c\u65f6\uff0c\u63a8\u7406\u5f15\u64ce\u4e0e\u5206\u8bcd\u5668\u6a21\u5757\u901a\u8fc7\u914d\u7f6e\u52a0\u8f7d\u8fd9\u4e9b\u6587\u4ef6\uff0c\u5b9e\u73b0\u5b89\u5168\u66f4\u65b0\u800c\u65e0\u9700\u91cd\u542f\u6216\u5b8c\u6574\u91cd\u65b0\u90e8\u7f72\u5e94\u7528\u3002<\/p>\n<p>\u901a\u8fc7\u5c06\u6a21\u578b\u4e0e\u5206\u8bcd\u5668\u63d0\u5347\u4e3a\u4e00\u7b49\u90e8\u7f72\u7ec4\u4ef6\uff0c\u56e2\u961f\u53ef\u4ee5\u83b7\u5f97\u53ef\u8ffd\u6eaf\u6027\u4e0e\u7248\u672c\u63a7\u5236\u3002\u5728\u53d7\u76d1\u7ba1\u73af\u5883\u4e2d\uff0c\u8fd9\u79cd\u63d0\u5347\u5bf9\u4e8e\u53ef\u590d\u73b0\u6027\u3001\u53ef\u89e3\u91ca\u6027\u4e0e\u56de\u6eda\u80fd\u529b\u81f3\u5173\u91cd\u8981\u3002<\/p>\n<h2>\u5206\u8bcd\u5668\u67b6\u6784<\/h2>\n<p>\u5206\u8bcd\u5668\u662f Transformer \u63a8\u7406\u7cfb\u7edf\u4e2d\u6700\u5bb9\u6613\u88ab\u5ffd\u89c6\u5374\u81f3\u5173\u91cd\u8981\u7684\u7ec4\u4ef6\u4e4b\u4e00\u3002\u4eba\u4eec\u7684\u6ce8\u610f\u529b\u5f80\u5f80\u96c6\u4e2d\u5728\u6a21\u578b\u4e0a\uff0c\u4f46\u5206\u8bcd\u5668\u8d1f\u8d23\u5c06\u4eba\u7c7b\u53ef\u8bfb\u6587\u672c\u8f6c\u6362\u4e3a\u6a21\u578b\u9700\u8981\u7684\u8f93\u5165 ID \u4e0e\u6ce8\u610f\u529b\u63a9\u7801\u3002\u53ea\u8981\u8fd9\u4e00\u8f6c\u6362\u8fc7\u7a0b\u5b58\u5728\u4efb\u4f55\u4e0d\u5339\u914d\uff0c\u5c31\u4f1a\u51fa\u73b0\u9759\u9ed8\u5931\u8d25\u2014\u2014\u9884\u6d4b\u5728\u8bed\u6cd5\u4e0a\u770b\u4f3c\u5408\u7406\uff0c\u8bed\u4e49\u4e0a\u5374\u4e0d\u6b63\u786e\u3002<\/p>\n<p>\u5728 Hugging Face \u751f\u6001\u4e2d\uff0c\u5206\u8bcd\u903b\u8f91\u88ab\u5e8f\u5217\u5316\u5230 <code>tokenizer.json<\/code> \u6587\u4ef6\u4e2d\u3002\u8be5\u5de5\u4ef6\u7f16\u7801\u4e86\u8bcd\u8868\u3001\u5206\u8bcd\u7b56\u7565\uff08\u4f8b\u5982<a href=\"https:\/\/link.segmentfault.com\/?enc=Vl7gXwywMVzarbHZB43oWQ%3D%3D.YM5zXvyXvSnwPqOJwuP5IIX5%2F4%2B4qdOMhDX88D%2FsPmyJ%2BGBK2Z73NOVGHPPJSzqEt05ucdm%2FGnLsFQ8u4j6vWQ%3D%3D\" rel=\"nofollow noopener noreferrer\" target=\"_blank\">\u5b57\u8282\u5bf9\u7f16\u7801\u6216 WordPiece<\/a>\uff09\u3001\u7279\u6b8a token \u5904\u7406\u4e0e\u914d\u7f6e\u8bbe\u7f6e\u3002\u5b83\u5fc5\u987b\u4f7f\u7528\u8bad\u7ec3\u65f6\u5b8c\u5168\u76f8\u540c\u7684\u5206\u8bcd\u5668\u7c7b\u4e0e\u53c2\u6570\u751f\u6210\u3002\u5373\u4fbf\u662f\u7ec6\u5fae\u5dee\u5f02\uff0c\u5982\u7f3a\u5c11 [CLS] token \u6216\u8bcd\u8868\u7d22\u5f15\u504f\u79fb\uff0c\u90fd\u4f1a\u964d\u4f4e\u6027\u80fd\u6216\u7834\u574f\u63a8\u7406\u8f93\u51fa\u3002<\/p>\n<p>\u4ece\u67b6\u6784\u4e0a\u770b\uff0c\u5206\u8bcd\u5668\u5e94\u4ee5\u72ec\u7acb\u3001\u7ebf\u7a0b\u5b89\u5168\u7684 Java \u6a21\u5757\u5b58\u5728\uff0c\u6d88\u8d39 <code>tokenizer.json<\/code> \u6587\u4ef6\u5e76\u751f\u6210\u53ef\u7528\u4e8e\u63a8\u7406\u7684\u7ed3\u6784\u3002\u5b83\u5fc5\u987b\u63a5\u53d7\u539f\u59cb\u5b57\u7b26\u4e32\u5e76\u8fd4\u56de\u7ed3\u6784\u5316\u8f93\u51fa\uff0c\u5305\u542b token ID\u3001\u6ce8\u610f\u529b\u63a9\u7801\u4ee5\u53ca\uff08\u53ef\u9009\uff09\u504f\u79fb\u6620\u5c04\uff0c\u4ee5\u4fbf\u4e0b\u6e38\u89e3\u91ca\u3002\u5728 Java \u670d\u52a1\u4e2d\u76f4\u63a5\u5d4c\u5165\u8fd9\u5c42\u903b\u8f91\uff0c\u800c\u4e0d\u662f\u4f9d\u8d56 Python \u5fae\u670d\u52a1\uff0c\u53ef\u4ee5\u964d\u4f4e\u5ef6\u8fdf\u5e76\u907f\u514d\u8106\u5f31\u7684\u57fa\u7840\u8bbe\u65bd\u4f9d\u8d56\u3002<\/p>\n<p>\u5728 Java \u4e2d\u6784\u5efa\u5206\u8bcd\u5c42\u8fd8\u80fd\u5b9e\u73b0\u76d1\u63a7\u3001\u5355\u5143\u6d4b\u8bd5\uff0c\u5e76\u5b8c\u6574\u96c6\u6210\u5230\u4f01\u4e1a CI\/CD \u6d41\u6c34\u7ebf\uff0c\u540c\u65f6\u4fbf\u4e8e\u5728\u7981\u6b62 Python \u8fd0\u884c\u65f6\u7684\u5b89\u5168\u6216\u53d7\u76d1\u7ba1\u73af\u5883\u4e2d\u90e8\u7f72\u3002\u5728\u6211\u4eec\u7684\u67b6\u6784\u4e2d\uff0c\u5206\u8bcd\u5668\u662f\u4e00\u4e2a\u6a21\u5757\u5316\u7684\u8fd0\u884c\u65f6\u7ec4\u4ef6\uff0c\u80fd\u591f\u52a8\u6001\u52a0\u8f7d <code>tokenizer.json<\/code> \u6587\u4ef6\u5e76\u5728\u4e0d\u540c\u6a21\u578b\u4e0e\u56e2\u961f\u4e4b\u95f4\u590d\u7528\u3002<\/p>\n<h2>\u63a8\u7406\u5f15\u64ce<\/h2>\n<p>\u4e00\u65e6\u8f93\u5165\u6587\u672c\u88ab\u8f6c\u6362\u4e3a token ID \u4e0e\u6ce8\u610f\u529b\u63a9\u7801\uff0c\u63a8\u7406\u5f15\u64ce\u7684\u6838\u5fc3\u4efb\u52a1\u5c31\u662f\u5c06\u8fd9\u4e9b\u5f20\u91cf\u4f20\u5165 ONNX \u6a21\u578b\u5e76\u8fd4\u56de\u6709\u610f\u4e49\u7684\u8f93\u51fa\u3002\u5728 Java \u4e2d\uff0c\u8fd9\u4e00\u8fc7\u7a0b\u7531 <a href=\"https:\/\/link.segmentfault.com\/?enc=VST7lAUKxWci3CRCwf%2B5wQ%3D%3D.XnAR9YyrrEgWGLc1Fxvn3UPP32upfrsMutL2Ab9U8QTJSpiBBaxlwectf5bbw%2BgN\" rel=\"nofollow noopener noreferrer\" target=\"_blank\">ONNX Runtime \u7684 Java API<\/a> \u8d1f\u8d23\uff0c\u63d0\u4f9b\u6210\u719f\u7684\u7ed1\u5b9a\u4ee5\u52a0\u8f7d\u6a21\u578b\u3001\u6784\u9020\u5f20\u91cf\u3001\u6267\u884c\u63a8\u7406\u4e0e\u83b7\u53d6\u7ed3\u679c\u3002<\/p>\n<p>\u63a8\u7406\u5f15\u64ce\u7684\u6838\u5fc3\u662f <code>OrtSession<\/code> \u7c7b\uff0c\u5b83\u662f ONNX \u6a21\u578b\u7684\u7f16\u8bd1\u3001\u521d\u59cb\u5316\u540e\u7684\u8868\u793a\uff0c\u53ef\u5728\u8bf7\u6c42\u4e4b\u95f4\u590d\u7528\u3002\u8be5\u4f1a\u8bdd\u5e94\u5728\u5e94\u7528\u542f\u52a8\u65f6\u521d\u59cb\u5316\uff0c\u5e76\u5728\u5404\u7ebf\u7a0b\u4e4b\u95f4\u5171\u4eab\u3002\u6bcf\u6b21\u8bf7\u6c42\u90fd\u91cd\u5efa\u4f1a\u8bdd\u4f1a\u5f15\u5165\u4e0d\u5fc5\u8981\u7684\u5ef6\u8fdf\u4e0e\u5185\u5b58\u538b\u529b\u3002<\/p>\n<p>\u51c6\u5907\u8f93\u5165\u6d89\u53ca\u521b\u5efa <code>NDArray<\/code> \u5f20\u91cf\uff0c\u5982 <code>input_ids<\/code>\u3001<code>attention_mask<\/code>\uff0c\u4ee5\u53ca\u53ef\u9009\u7684 <code>token_type_ids<\/code>\uff0c\u8fd9\u4e9b\u90fd\u662f Transformer \u6a21\u578b\u901a\u5e38\u671f\u671b\u7684\u6807\u51c6\u8f93\u5165\u3002\u5f20\u91cf\u7531 Java \u539f\u751f\u6570\u636e\u7ed3\u6784\u6784\u9020\uff0c\u518d\u4f20\u5165 ONNX \u4f1a\u8bdd\u3002\u4f1a\u8bdd\u8fd0\u884c\u63a8\u7406\u5e76\u751f\u6210\u8f93\u51fa\uff0c\u901a\u5e38\u5305\u62ec logits\u3001\u7c7b\u522b\u6982\u7387\u6216\u7ed3\u6784\u5316\u6807\u7b7e\uff0c\u5177\u4f53\u53d6\u51b3\u4e8e\u6a21\u578b\u3002<\/p>\n<p>\u5728 Java \u4e2d\uff0c\u63a8\u7406\u8c03\u7528\u901a\u5e38\u5982\u4e0b\uff1a<\/p>\n<pre><code class=\"java\">OrtSession.Result result = session.run(inputs);<\/code><\/pre>\n<p>ONNX Runtime \u8fd8\u652f\u6301\u6267\u884c\u63d0\u4f9b\u8005\uff0c\u7528\u4ee5\u51b3\u5b9a\u63a8\u7406\u5728 CPU \u8fd8\u662f GPU \u4e0a\u8fd0\u884c\u3002\u5728\u542f\u7528\u4e86 <a href=\"https:\/\/link.segmentfault.com\/?enc=%2FNh9YqhEIph86%2B%2BslRyvFQ%3D%3D.N5SccbGZk2jrLs3MFV3LMwPo89PxiDxaVGIPb36KYcH8hXnIcj80fWkZjU3hgWWK\" rel=\"nofollow noopener noreferrer\" target=\"_blank\">CUDA<\/a> \u7684\u7cfb\u7edf\u4e0a\uff0c\u63a8\u7406\u53ef\u4ee5\u4ee5\u6700\u5c0f\u914d\u7f6e\u5378\u8f7d\u5230 GPU\u3002\u5982\u679c GPU \u8d44\u6e90\u4e0d\u53ef\u7528\uff0c\u5219\u4f18\u96c5\u5730\u56de\u9000\u5230 CPU\uff0c\u4f7f\u5f97\u8de8\u73af\u5883\u884c\u4e3a\u4fdd\u6301\u4e00\u81f4\u3002\u8fd9\u79cd\u7075\u6d3b\u6027\u8ba9\u5355\u4e00 Java \u4ee3\u7801\u5e93\u53ef\u4ee5\u4ece\u5f00\u53d1\u8005\u7b14\u8bb0\u672c\u6269\u5c55\u5230\u751f\u4ea7 GPU \u96c6\u7fa4\uff0c\u65e0\u9700\u5206\u652f\u903b\u8f91\uff0c\u4e0e\u6211\u4eec\u5728\u201c\u5c06 GPU \u7ea7\u6027\u80fd\u5e26\u5165\u4f01\u4e1a Java\u201d\u4e00\u6587\u4e2d\u8ba8\u8bba\u7684\u7406\u5ff5\u4e00\u8109\u76f8\u627f\u3002<\/p>\n<p>\u4ece\u67b6\u6784\u89d2\u5ea6\u770b\uff0c\u63a8\u7406\u5f15\u64ce\u5fc5\u987b\u4fdd\u6301\u65e0\u72b6\u6001\u3001\u7ebf\u7a0b\u5b89\u5168\u4e0e\u8d44\u6e90\u9ad8\u6548\u3002\u5b83\u5e94\u63d0\u4f9b\u5e72\u51c0\u7684\u53ef\u89c2\u6d4b\u6027\u63a5\u53e3\uff1a\u65e5\u5fd7\u3001\u8ffd\u8e2a\u4e0e\u7ed3\u6784\u5316\u9519\u8bef\u5904\u7406\u3002\u5bf9\u4e8e\u9ad8\u541e\u5410\u573a\u666f\uff0c\u6c60\u5316\u4e0e\u5fae\u6279\u5904\u7406\u6709\u52a9\u4e8e\u4f18\u5316\u6027\u80fd\uff1b\u5728\u4f4e\u5ef6\u8fdf\u8bed\u5883\u4e0b\uff0c\u5185\u5b58\u590d\u7528\u4e0e\u4f1a\u8bdd\u8c03\u4f18\u5bf9\u4e8e\u4fdd\u6301\u53ef\u9884\u6d4b\u7684\u63a8\u7406\u6210\u672c\u81f3\u5173\u91cd\u8981\u3002<\/p>\n<p>\u5c06\u63a8\u7406\u89c6\u4e3a\u5177\u6709\u6e05\u6670\u5951\u7ea6\u4e0e\u53ef\u8fb9\u754c\u5316\u6027\u80fd\u7279\u5f81\u7684\u6a21\u5757\u5316\u670d\u52a1\uff0c\u67b6\u6784\u5e08\u5373\u53ef\u5c06 AI \u903b\u8f91\u5b8c\u5168\u4e0e\u4e1a\u52a1\u5de5\u4f5c\u6d41\u89e3\u8026\uff0c\u5b9e\u73b0\u72ec\u7acb\u6f14\u8fdb\u4e0e\u53ef\u9760\u6269\u5c55\u3002<\/p>\n<h2>\u90e8\u7f72\u6a21\u5f0f<\/h2>\n<p>\u8bbe\u8ba1\u63a8\u7406\u5f15\u64ce\u53ea\u662f\u6311\u6218\u7684\u4e00\u534a\uff1b\u5982\u4f55\u5728\u4f01\u4e1a\u73af\u5883\u4e2d\u90e8\u7f72\u540c\u6837\u91cd\u8981\u3002Java \u7cfb\u7edf\u6db5\u76d6\u4ece REST API \u5230 ETL \u7ba1\u9053\u4e0e\u5b9e\u65f6\u5f15\u64ce\u7684\u5e7f\u6cdb\u573a\u666f\uff0c\u56e0\u6b64\u57fa\u4e8e ONNX \u7684\u63a8\u7406\u5fc5\u987b\u5728\u4e0d\u91cd\u590d\u903b\u8f91\u6216\u5272\u88c2\u914d\u7f6e\u7684\u524d\u63d0\u4e0b\u8fdb\u884c\u9002\u914d\u3002<br \/>\u5728\u591a\u6570\u60c5\u51b5\u4e0b\uff0c\u5206\u8bcd\u5668\u4e0e\u63a8\u7406\u5f15\u64ce\u4f5c\u4e3a Java \u5e93\u76f4\u63a5\u5d4c\u5165\uff0c\u4ece\u800c\u907f\u514d\u8fd0\u884c\u65f6\u4f9d\u8d56\uff0c\u5e76\u4e0e\u65e5\u5fd7\u3001\u76d1\u63a7\u4e0e\u5b89\u5168\u6846\u67b6\u6574\u6d01\u96c6\u6210\u3002\u5728 Spring Boot \u4e0e Quarkus \u7b49\u6846\u67b6\u4e2d\uff0c\u63a8\u7406\u5c31\u662f\u53e6\u4e00\u4e2a\u53ef\u6ce8\u5165\u7684\u670d\u52a1\u3002<\/p>\n<p>\u66f4\u5927\u7684\u56e2\u961f\u5e38\u5c06\u8fd9\u5c42\u903b\u8f91\u5916\u7f6e\u4e3a\u5171\u4eab\u6a21\u5757\uff0c\u8d1f\u8d23\u5206\u8bcd\u5668\u4e0e\u6a21\u578b\u52a0\u8f7d\u3001\u5f20\u91cf\u51c6\u5907\u4e0e ONNX \u4f1a\u8bdd\u6267\u884c\u3002\u5916\u7f6e\u5316\u4fc3\u8fdb\u590d\u7528\u3001\u7b80\u5316\u6cbb\u7406\u3001\u5e76\u5728\u5404\u670d\u52a1\u4e4b\u95f4\u63d0\u4f9b\u4e00\u81f4\u7684 AI \u884c\u4e3a\u3002<\/p>\n<p>\u5728\u5177\u5907 GPU \u7684\u73af\u5883\u4e2d\uff0c\u53ef\u4ee5\u901a\u8fc7\u914d\u7f6e\u542f\u7528 ONNX Runtime \u7684 CUDA \u63d0\u4f9b\u8005\uff0c\u65e0\u9700\u66f4\u6539\u4ee3\u7801\u3002\u76f8\u540c\u7684 Java \u5e94\u7528\u65e2\u80fd\u5728 CPU \u96c6\u7fa4\u4e0a\u8fd0\u884c\uff0c\u4e5f\u80fd\u5728 GPU \u96c6\u7fa4\u4e0a\u8fd0\u884c\uff0c\u4f7f\u90e8\u7f72\u65e2\u53ef\u79fb\u690d\u53c8\u8d44\u6e90\u611f\u77e5\u3002<\/p>\n<p>\u6a21\u578b\u5236\u54c1\u65e2\u53ef\u968f\u5e94\u7528\u6253\u5305\uff0c\u4e5f\u53ef\u4ece\u6a21\u578b\u6ce8\u518c\u8868\u6216\u6302\u8f7d\u5377\u52a8\u6001\u52a0\u8f7d\u3002\u540e\u8005\u652f\u6301\u70ed\u66ff\u6362\u3001\u56de\u6eda\u4e0e A\/B \u6d4b\u8bd5\uff0c\u4f46\u9700\u8981\u8c28\u614e\u7684\u9a8c\u8bc1\u4e0e\u7248\u672c\u7ba1\u7406\u3002\u5173\u952e\u5728\u4e8e\u7075\u6d3b\u6027\u3002\u4e00\u4e2a\u53ef\u63d2\u62d4\u3001\u5bf9\u73af\u5883\u654f\u611f\u7684\u90e8\u7f72\u6a21\u578b\uff0c\u65e0\u8bba\u662f\u5d4c\u5165\u5f0f\u3001\u5171\u4eab\u5f0f\u8fd8\u662f\u5bb9\u5668\u5316\uff0c\u90fd\u80fd\u786e\u4fdd\u63a8\u7406\u65e0\u7f1d\u878d\u5165\u73b0\u6709\u7684 CI\/CD \u4e0e\u8fd0\u884c\u7b56\u7565\u3002<\/p>\n<h2>\u4e0e\u6846\u67b6\u5c42\u62bd\u8c61\u7684\u6bd4\u8f83<\/h2>\n<p>Spring AI \u7b49\u6846\u67b6\u901a\u8fc7\u4e3a OpenAI\u3001Azure \u6216 AWS Bedrock \u7b49\u63d0\u4f9b\u8005\u63d0\u4f9b\u5ba2\u6237\u7aef\u62bd\u8c61\uff0c\u7b80\u5316\u4e86\u8c03\u7528\u5916\u90e8\u5927\u578b\u8bed\u8a00\u6a21\u578b\u7684\u8fc7\u7a0b\u3002\u8fd9\u4e9b\u6846\u67b6\u5728\u539f\u578b\u5316\u5bf9\u8bdd\u754c\u9762\u4e0e\u68c0\u7d22\u589e\u5f3a\u751f\u6210\uff08RAG\uff09\u7ba1\u9053\u65b9\u9762\u5f88\u6709\u4ef7\u503c\uff0c\u4f46\u5176\u5de5\u4f5c\u5c42\u6b21\u4e0e\u57fa\u4e8e ONNX \u7684\u63a8\u7406\u6839\u672c\u4e0d\u540c\u3002\u524d\u8005\u5c06\u63a8\u7406\u59d4\u6258\u7ed9\u8fdc\u7a0b\u670d\u52a1\uff0c\u540e\u8005\u5219\u5728 JVM \u5185\u76f4\u63a5\u6267\u884c\u6a21\u578b\uff0c\u4f7f\u63a8\u7406\u4fdd\u6301\u786e\u5b9a\u6027\u3001\u53ef\u5ba1\u8ba1\u4e14\u5b8c\u5168\u53d7\u4f01\u4e1a\u63a7\u5236\u3002<\/p>\n<p>\u8fd9\u79cd\u533a\u522b\u5177\u6709\u5b9e\u9645\u5f71\u54cd\u3002\u5916\u90e8\u6846\u67b6\u7684\u8f93\u51fa\u4e0d\u53ef\u91cd\u590d\uff0c\u5e76\u4f9d\u8d56\u7b2c\u4e09\u65b9\u63d0\u4f9b\u8005\u7684\u53ef\u7528\u6027\u4e0e\u4e0d\u65ad\u6f14\u8fdb\u7684 API\u3002\u76f8\u6bd4\u4e4b\u4e0b\uff0cONNX \u63a8\u7406\u4f7f\u7528\u7248\u672c\u5316\u7684\u5de5\u4ef6\uff0c<code>model.onnx<\/code> \u4e0e <code>tokenizer.json<\/code>\uff0c\u5728\u5404\u73af\u5883\u4e2d\u8868\u73b0\u4e00\u81f4\uff0c\u4ece\u5f00\u53d1\u8005\u7b14\u8bb0\u672c\u5230\u751f\u4ea7 GPU \u96c6\u7fa4\u90fd\u4e0d\u4f8b\u5916\u3002\u8fd9\u79cd\u53ef\u91cd\u590d\u6027\u5bf9\u5408\u89c4\u4e0e\u56de\u5f52\u6d4b\u8bd5\u81f3\u5173\u91cd\u8981\uff0c\u56e0\u4e3a\u6a21\u578b\u884c\u4e3a\u7684\u7ec6\u5fae\u5dee\u5f02\u53ef\u80fd\u5e26\u6765\u663e\u8457\u7684\u4e0b\u6e38\u5f71\u54cd\u3002\u5b83\u4e5f\u786e\u4fdd\u654f\u611f\u6570\u636e\u4e0d\u79bb\u5f00\u4f01\u4e1a\u8fb9\u754c\uff0c\u8fd9\u5728\u91d1\u878d\u4e0e\u533b\u7597\u7b49\u9886\u57df\u662f\u57fa\u672c\u8981\u6c42\u3002<\/p>\n<p>\u4e5f\u8bb8\u66f4\u4e3a\u91cd\u8981\u7684\u662f\uff0cONNX \u4fdd\u6301\u4f9b\u5e94\u5546\u4e2d\u7acb\u3002\u4f5c\u4e3a\u88ab\u5404\u8bad\u7ec3\u6846\u67b6\u5e7f\u6cdb\u652f\u6301\u7684\u5f00\u653e\u6807\u51c6\uff0c\u7ec4\u7ec7\u53ef\u4ee5\u81ea\u7531\u4f7f\u7528\u504f\u597d\u7684\u751f\u6001\u8fdb\u884c\u8bad\u7ec3\uff0c\u5e76\u5728 Java \u4e2d\u90e8\u7f72\uff0c\u800c\u4e0d\u5fc5\u62c5\u5fc3\u4f9b\u5e94\u5546\u9501\u5b9a\u6216 API \u6f02\u79fb\u3002\u7531\u6b64\uff0cONNX \u4e0e Spring AI \u7b49\u6846\u67b6\u662f\u4e92\u8865\u5173\u7cfb\u800c\u975e\u7ade\u4e89\u5173\u7cfb\u3002\u524d\u8005\u4e3a\u5408\u89c4\u5173\u952e\u7684\u5de5\u4f5c\u8d1f\u8f7d\u63d0\u4f9b\u7a33\u5b9a\u7684\u3001\u8fdb\u7a0b\u5185\u7684\u57fa\u7840\uff1b\u540e\u8005\u5e2e\u52a9\u5f00\u53d1\u8005\u5728\u5e94\u7528\u8fb9\u7f18\u5feb\u901f\u63a2\u7d22\u751f\u6210\u5f0f\u7528\u4f8b\u3002\u5bf9\u67b6\u6784\u5e08\u800c\u8a00\uff0c\u80fd\u591f\u6e05\u6670\u5212\u5b9a\u8fd9\u6761\u8fb9\u754c\uff0c\u624d\u80fd\u786e\u4fdd AI \u91c7\u7528\u65e2\u5177\u521b\u65b0\u6027\u53c8\u5728\u8fd0\u8425\u4e0a\u53ef\u6301\u7eed\u3002<\/p>\n<h2>\u4e0b\u4e00\u6b65<\/h2>\n<p>\u65e2\u7136\u6211\u4eec\u5df2\u7ecf\u9610\u660e\u5982\u4f55\u901a\u8fc7\u539f\u751f\u5206\u8bcd\u4e0e\u65e0\u72b6\u6001\u63a8\u7406\u5c42\u5c06 ONNX \u6a21\u578b\u96c6\u6210\u5230 Java \u7cfb\u7edf\u4e2d\uff0c\u4e0b\u4e00\u6b65\u7684\u6311\u6218\u662f\u5728\u751f\u4ea7\u73af\u5883\u4e2d\u5b89\u5168\u3001\u53ef\u9760\u5730\u6269\u5c55\u8fd9\u5957\u67b6\u6784\u3002<\/p>\n<p>\u5728\u4e0b\u4e00\u7bc7\u6587\u7ae0\u4e2d\uff0c\u6211\u4eec\u5c06\u63a2\u8ba8\uff1a<\/p>\n<ul>\n<li>Java \u4e2d\u7684\u5b89\u5168\u4e0e\u5ba1\u8ba1\uff1a\u5b9e\u73b0\u53ef\u8ffd\u8e2a\u3001\u53ef\u89e3\u91ca\u4e14\u7b26\u5408\u6cd5\u6cbb\u7684 AI \u7ba1\u9053\u3002<\/li>\n<li>\u53ef\u6269\u5c55\u7684\u63a8\u7406\u6a21\u5f0f\uff1a\u5728 CPU\/GPU \u7ebf\u7a0b\u3001\u5f02\u6b65\u4f5c\u4e1a\u961f\u5217\u4e0e\u9ad8\u541e\u5410\u7ba1\u9053\u95f4\u8d1f\u8f7d\u5747\u8861\uff0c\u4f7f\u7528 Java \u539f\u751f\u6784\u9020\u3002<\/li>\n<li>\u5185\u5b58\u7ba1\u7406\u4e0e\u53ef\u89c2\u6d4b\u6027\uff1a\u5256\u6790\u63a8\u7406\u5185\u5b58\u5360\u7528\u3001\u8ffd\u8e2a\u6162\u8def\u5f84\uff0c\u5e76\u4f7f\u7528 JVM \u539f\u751f\u5de5\u5177\u8c03\u4f18\u5ef6\u8fdf\u3002<\/li>\n<li>\u8d85\u8d8a JNI \u7684\u6f14\u8fdb\uff1a\u52a8\u624b\u89e3\u8bfb\u5916\u90e8\u51fd\u6570\u4e0e\u5185\u5b58 API\uff08JEP 454\uff09\u4f5c\u4e3a\u672a\u6765\u53ef\u9760\u63a8\u7406\u7ba1\u9053\u5bf9 JNI \u7684\u66ff\u4ee3\u65b9\u6848\u3002<\/li>\n<\/ul>\n<p><strong>\u4f5c\u8005\u6ce8<\/strong>\uff1a\u672c\u6587\u5b9e\u73b0\u57fa\u4e8e\u72ec\u7acb\u7684\u6280\u672f\u7814\u7a76\uff0c\u4e0d\u4ee3\u8868\u4efb\u4f55\u7279\u5b9a\u7ec4\u7ec7\u7684\u67b6\u6784\u3002<\/p>\n<blockquote><p>\u672c\u6587\u7ffb\u8bd1\u81ea\uff1a<a href=\"https:\/\/link.segmentfault.com\/?enc=thiv%2BndZrRAK4f3N34AKEg%3D%3D.RTC4wjqGLs%2FRv23hec5dHxnyCnUon%2FlDTcfJw8ocvvVNmwO7EUcrdXpy%2FD8N6%2BWJEYG4cNaPUFIKArfdLA7iog%3D%3D\" rel=\"nofollow noopener noreferrer\" target=\"_blank\">https:\/\/www.infoq.com\/articles\/onnx-ai-inference-with-java\/<\/a>\uff0c\u4f5c\u8005\uff1aSyed Danish Ali<\/p><\/blockquote>\n<p><a href=\"https:\/\/segmentfault.com\/a\/1190000047318621\">https:\/\/segmentfault.com\/a\/1190000047318621<\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u5f15\u8a00 \u867d\u7136 Python \u4e3b\u5bfc\u4e86\u673a\u5668\u5b66\u4e60\u751f\u6001\uff0c\u4f46\u5927\u591a\u6570\u4f01\u4e1a\u5e94\u7528\u4ecd\u8fd0\u884c\u5728 Java \u4e0a\u3002\u8fd9\u79cd\u8131\u8282\u9020\u6210\u4e86\u90e8\u7f72\u74f6\u9888\u3002 [&hellip;]<\/p>\n","protected":false},"author":8,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"_jetpack_memberships_contains_paid_content":false,"footnotes":""},"categories":[4],"tags":[],"class_list":["post-10021029","post","type-post","status-publish","format-standard","hentry","category-share"],"jetpack_sharing_enabled":true,"jetpack_featured_media_url":"","_links":{"self":[{"href":"https:\/\/itmax.cn\/index.php?rest_route=\/wp\/v2\/posts\/10021029","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/itmax.cn\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/itmax.cn\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/itmax.cn\/index.php?rest_route=\/wp\/v2\/users\/8"}],"replies":[{"embeddable":true,"href":"https:\/\/itmax.cn\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=10021029"}],"version-history":[{"count":0,"href":"https:\/\/itmax.cn\/index.php?rest_route=\/wp\/v2\/posts\/10021029\/revisions"}],"wp:attachment":[{"href":"https:\/\/itmax.cn\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=10021029"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/itmax.cn\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=10021029"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/itmax.cn\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=10021029"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}