{"id":223,"date":"2025-12-28T21:06:33","date_gmt":"2025-12-28T13:06:33","guid":{"rendered":"https:\/\/www.xiewuyi.com\/?p=223"},"modified":"2025-12-28T21:31:21","modified_gmt":"2025-12-28T13:31:21","slug":"%e5%a4%9a%e6%a8%a1%e6%80%81%e5%a4%a7%e6%a8%a1%e5%9e%8b%e5%b7%a5%e4%bd%9c%e5%8e%9f%e7%90%86","status":"publish","type":"post","link":"https:\/\/www.xiewuyi.com\/index.php\/2025\/12\/28\/%e5%a4%9a%e6%a8%a1%e6%80%81%e5%a4%a7%e6%a8%a1%e5%9e%8b%e5%b7%a5%e4%bd%9c%e5%8e%9f%e7%90%86\/","title":{"rendered":"\u591a\u6a21\u6001\u5927\u6a21\u578b\u5de5\u4f5c\u539f\u7406"},"content":{"rendered":"\n<p>\u591a\u6a21\u6001\u5927\u6a21\u578b\uff08MLLM\uff09\u5df2\u7ecf\u6210\u4e3a\u5fc5\u7136\u7684\u53d1\u5c55\u8d8b\u52bf\uff0c\u6587\u5b57\u91cc\u7684\u4fe1\u606f\u548c\u77e5\u8bc6\u53ea\u662f\u51b0\u5c71\u4e00\u89d2\uff0c\u6a21\u578b\u4e0d\u4ec5\u8981\u80fd\u7406\u89e3\u6587\u5b57\uff0c\u8fd8\u8981\u80fd\u770b\u61c2\u56fe\u7247\uff0c\u542c\u61c2\u58f0\u97f3\u3002<\/p>\n\n\n\n<p>\u7406\u89e3\u591a\u6a21\u6001\u6a21\u578b\uff0c\u5173\u952e\u5728\u4e8e\u7406\u89e3\u4e24\u79cd\u622a\u7136\u4e0d\u540c\u7684\u67b6\u6784\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u6a21\u5757\u5316\u67b6\u6784\uff08Modular\uff09<\/strong>\uff1a\u50cf\u662f\u4e00\u4f4d\u201c\u6234\u7740\u7ffb\u8bd1\u773c\u955c\u7684\u8bed\u8a00\u5929\u624d\u201d\u3002LLM \u4f5c\u4e3a\u67b6\u6784\u7684\u6838\u5fc3\uff0c\u672c\u8eab\u770b\u4e0d\u89c1\u56fe\u50cf\uff0c\u9700\u8981\u4e00\u4e2a\u201c\u7ffb\u8bd1\u5b98\u201d\uff08\u5373\u201c\u8fde\u63a5\u5668\u201d\uff09\u5c06\u56fe\u50cf\u4fe1\u53f7\u8f6c\u8bd1\u6210\u5b83\u80fd\u542c\u61c2\u7684\u6587\u672c\u8bed\u8a00\u3002<\/li>\n\n\n\n<li><strong>\u539f\u751f\u591a\u6a21\u6001\uff08Native\uff09<\/strong>\uff1a\u50cf\u662f\u4e00\u4f4d\u201c\u5929\u751f\u62e5\u6709\u611f\u5b98\u7684\u6bcd\u8bed\u8005\u201d\u3002\u5b83\u5728\u8bde\u751f\u4e4b\u521d\uff0c\u795e\u7ecf\u7f51\u7edc\u5c31\u540c\u65f6\u53d1\u80b2\u51fa\u4e86\u89c6\u89c9\u548c\u8bed\u8a00\u4e2d\u67a2\uff0c\u6240\u6709\u4fe1\u606f\u5171\u4eab\u540c\u4e00\u5957\u5e95\u5c42\u903b\u8f91\uff0c\u7aef\u5230\u7aef\u7406\u89e3\u591a\u6a21\u6001\u4fe1\u606f\u3002<\/li>\n<\/ul>\n\n\n\n<p>\u672c\u6587\u5c06\u6df1\u5165\u6d45\u51fa\u5730\u8bb2\u89e3\u8fd9\u4e24\u79cd\u67b6\u6784\u7684\u5de5\u4f5c\u539f\u7406\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">\u4e00\u3001\u4e24\u79cd\u67b6\u6784\u7684\u5171\u540c\u57fa\u77f3 \u2014\u2014 \u56fe\u50cf Token \u5316<\/h3>\n\n\n\n<p>\u65e0\u8bba\u54ea\u79cd\u67b6\u6784\uff0c\u5927\u6a21\u578b\u5904\u7406\u4fe1\u606f\u7684\u6838\u5fc3\u90e8\u4ef6\u90fd\u662f <strong>Transformer<\/strong>\u3002\u4f46 Transformer \u53ea\u80fd\u5904\u7406\u79bb\u6563\u7684\u5e8f\u5217\uff08Token\uff09\uff0c\u65e0\u6cd5\u76f4\u63a5\u8bfb\u53d6\u8fde\u7eed\u7684\u50cf\u7d20\u3002\u56e0\u6b64\uff0c\u7b2c\u4e00\u6b65\u5fc5\u987b\u662f<strong>\u56fe\u50cf\u7684 Token \u5316<\/strong>\u3002<\/p>\n\n\n\n<h4 class=\"wp-block-heading\">1. \u4e0e\u6587\u672c Token \u5316\u7684\u7c7b\u6bd4<\/h4>\n\n\n\n<p>\u5728\u5904\u7406\u6587\u672c\u65f6\uff0c\u6a21\u578b\u4f1a\u5c06\u53e5\u5b50\u201c\u6211\u7231\u732b\u201d\u5207\u5206\u4e3a\u8bcd\u5143\uff0c\u5373 Token \uff1a<code>[\u6211]<\/code>\u3001<code>[\u7231]<\/code>\u3001<code>[\u732b]<\/code>\u3002\u6bcf\u4e2a\u8bcd\u5143\u5bf9\u5e94\u8bcd\u8868\u91cc\u7684\u4e00\u4e32\u6570\u5b57 ID\uff0c\u56fe\u50cf\u7684 Token \u5316\u4e5f\u662f\u5c06\u56fe\u50cf\u8f6c\u53d8\u4e3a\u4e00\u4e32\u8fde\u7eed\u7684 Token \u5e8f\u5217 \u3002<\/p>\n\n\n\n<h4 class=\"wp-block-heading\">2. ViT\uff1a\u5c06\u56fe\u7247\u5207\u5206\u4e3a\u5411\u91cf<\/h4>\n\n\n\n<p><strong>ViT\uff08Vision Transformer\uff09<\/strong> \uff0c\u4e00\u79cd\u5c06 Transformer \u67b6\u6784\u76f4\u63a5\u5e94\u7528\u4e8e\u8ba1\u7b97\u673a\u89c6\u89c9\u4efb\u52a1\u7684\u6df1\u5ea6\u5b66\u4e60\u6a21\u578b\u3002\u5de5\u4f5c\u539f\u7406\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u56fe\u50cf\u5206\u5757\uff08Patch Partitioning\uff09<\/strong>\uff1a\u60f3\u8c61\u4e00\u5f20 224*224 \u50cf\u7d20\u7684\u732b\u56fe<em>\u3002<\/em>ViT \u4f1a\u628a\u5b83\u5207\u6210 14*14 = 196 \u4e2a\u5c0f\u65b9\u5757\uff08Patches\uff09\uff0c\u6bcf\u4e2a\u65b9\u5757\u5927\u5c0f\u4e3a 16*16 \u50cf\u7d20\u3002<\/li>\n\n\n\n<li><strong>\u5c55\u5e73\uff08Flattening\uff09<\/strong>\uff1a\u6bcf\u4e2a\u65b9\u5757\u5305\u542b 16*16 \u4e2a\u50cf\u7d20\u70b9\uff0c\u6bcf\u4e2a\u50cf\u7d20\u6709 RGB \u4e09\u4e2a\u989c\u8272\u901a\u9053\u3002\u6211\u4eec\u5c06\u8fd9\u4e9b\u50cf\u7d20\u70b9\u6392\u6210\u4e00\u6392\uff0c\u53d8\u6210\u4e00\u4e2a\u5305\u542b 768 \u4e2a\u6570\u5b57\u7684\u957f\u5217\u8868\u3002<\/li>\n\n\n\n<li><strong>\u7ebf\u6027\u6295\u5f71\uff08Linear Projection\uff09<\/strong>\uff1a\u8fd9\u5c31\u50cf\u662f\u4e00\u4e2a\u201c\u7a7a\u95f4\u53d8\u6362\u5668\u201d\u3002\u6a21\u578b\u901a\u8fc7\u77e9\u9635\u4e58\u6cd5\uff0c\u5c06\u8fd9 768 \u4e2a\u539f\u59cb\u6570\u503c\u538b\u7f29\u6210\u4e00\u4e2a\u66f4\u9ad8\u7ef4\u3001\u8bed\u4e49\u66f4\u4e30\u5bcc\u7684\u6570\u5b66\u7a7a\u95f4\u4e2d\uff08\u4f8b\u5982 1024 \u7ef4\u7684\u5411\u91cf\u7a7a\u95f4\uff09\u7684\u7279\u5f81\u5411\u91cf\u3002<\/li>\n<\/ul>\n\n\n\n<h4 class=\"wp-block-heading\">3. VQ-VAE\uff1a\u5efa\u7acb\u201c\u89c6\u89c9 Token \u5b57\u5178\u201d<\/h4>\n\n\n\n<p>\u4e3a\u4e86\u8ba9\u56fe\u50cf\u66f4\u50cf\u8bed\u8a00\uff0c<strong>VQ-VAE<\/strong>\uff08\u5411\u91cf\u91cf\u5316\u53d8\u5206\u81ea\u7f16\u7801\u5668\uff09\u5f15\u5165\u4e86<strong>\u201c\u7801\u672c\u201d\uff08Codebook\uff09<\/strong>\u3002 \u5b83\u9884\u5148\u5b9a\u4e49\u4e86\u4e00\u672c\u5305\u542b\u51e0\u4e07\u4e2a\u201c\u6807\u51c6\u89c6\u89c9\u7b26\u53f7\u201d\u7684\u5b57\u5178\u3002\u6bcf\u4e2a\u56fe\u50cf\u5207\u7247\u5728\u7ebf\u6027\u6295\u5f71\u540e\uff0c\u6a21\u578b\u4f1a\u53bb\u5b57\u5178\u91cc\u5bfb\u627e\u6700\u50cf\u5b83\u7684\u90a3\u4e2a\u7b26\u53f7\uff0c\u5e76\u7528\u7b26\u53f7\u7684\u7f16\u53f7\uff08\u5982 <code>Token #502<\/code>\uff09\u6765\u4ee3\u66ff\u539f\u59cb\u6570\u636e\uff0c\u5b9e\u73b0\u771f\u6b63\u7684\u7c7b\u6587\u672c\u201c\u5206\u8bcd\u5316\uff08Tokenization\uff09\u201d \uff1a<strong>\u56fe\u7247\u5728\u6570\u5b66\u5f62\u5f0f\u4e0a\u53d8\u5f97\u548c\u6587\u672c\u8bcd\u5143\u4e00\u6837\u3002<\/strong><\/p>\n\n\n\n<h3 class=\"wp-block-heading\">\u4e8c\u3001\u6a21\u5757\u5316\u67b6\u6784\u7684\u5de5\u4f5c\u539f\u7406<\/h3>\n\n\n\n<p>\u4ee5 <strong>LLaVA<\/strong> \u548c <strong>BLIP-2<\/strong> \u4e3a\u4ee3\u8868\u7684\u67b6\u6784\uff0c\u5176\u6838\u5fc3\u903b\u8f91\u662f\u201c\u7ec4\u4ef6\u62fc\u63a5\u201d\u3002<\/p>\n\n\n\n<h4 class=\"wp-block-heading\">1. \u4e09\u4f4d\u4e00\u4f53\u7684\u534f\u540c\u5de5\u4f5c\u6d41<\/h4>\n\n\n\n<p>\u8fd9\u79cd\u6a21\u578b\u7531\u4e09\u4e2a\u76f8\u4e92\u914d\u5408\u7684\u7ec4\u4ef6\u6784\u6210\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u89c6\u89c9\u7f16\u7801\u5668\uff08\u773c\u775b\uff09<\/strong>\uff1a\u901a\u5e38\u662f <strong>CLIP \u6a21\u578b<\/strong>\u3002\u5b83\u8d1f\u8d23\u628a\u539f\u59cb\u50cf\u7d20\u521d\u6b65\u52a0\u5de5\uff0c\u63d0\u53d6\u51fa\u5305\u542b\u7269\u4f53\u5f62\u72b6\u3001\u989c\u8272\u7684<strong>\u539f\u59cb\u89c6\u89c9\u7279\u5f81\u5411\u91cf <\/strong><math data-latex=\"Z\"><semantics><mi>Z<\/mi><annotation encoding=\"application\/x-tex\">Z<\/annotation><\/semantics><\/math><strong><sub>v<\/sub><\/strong>\u3002<\/li>\n\n\n\n<li><strong>\u8fde\u63a5\u5668\uff08\u7ffb\u8bd1\uff09<\/strong>\uff1a\u5b83\u662f\u552f\u4e00\u7684\u201c\u7ffb\u8bd1\u201d\u73af\u8282\u3002\u5728 LLaVA \u4e2d\uff0c\u5b83\u5c31\u662f\u4e00\u4e2a<strong>\u7ebf\u6027\u6295\u5f71\u77e9\u9635 <\/strong>W\uff0c\u8d1f\u8d23\u628a\u89c6\u89c9\u7279\u5f81\u5411\u91cf<strong> <\/strong><math data-latex=\"Z\"><semantics><mi>Z<\/mi><annotation encoding=\"application\/x-tex\">Z<\/annotation><\/semantics><\/math><strong><sub>v<\/sub><\/strong> \u201c\u6295\u5f71\u201d\u5230 LLM \u7684\u7a7a\u95f4\uff0c\u53d8\u6210 <strong>\u201c\u89c6\u89c9 Token\u201d <\/strong>H<strong><sub>v<\/sub><\/strong>\uff0c\u8fd9\u4e00\u6b65\u5b8c\u6210\u56fe\u50cf\u7a7a\u95f4\u5230\u6587\u672c\u7a7a\u95f4\u7684\u5411\u91cf\u5bf9\u9f50\uff0c\u89c6\u89c9 Token \u5728\u6570\u5b66\u7ef4\u5ea6\u4e0a\u4e0e LLM \u7684\u8bcd\u8868\u7a7a\u95f4\u5b8c\u5168\u4e00\u81f4\u3002<\/li>\n\n\n\n<li><strong>\u8bed\u8a00\u6a21\u578b\uff08\u5927\u8111\uff09<\/strong>\uff1a\u5982 Llama\u3002\u5b83\u63a5\u6536\u8fd9\u4e32 H<strong><sub>v<\/sub><\/strong> \u4ee5\u53ca\u4eba\u7c7b\u7684\u6587\u672c\u6307\u4ee4\uff0c\u50cf\u5904\u7406\u666e\u901a\u6587\u672c\u4e00\u6837\uff0c\u5229\u7528\u6ce8\u610f\u529b\u673a\u5236\uff08Attention\uff09\u8ba1\u7b97\u8fd9\u4e9b\u201c\u89c6\u89c9 Token\u201d\u4e0e\u201c\u4eba\u7c7b\u95ee\u9898\u5bf9\u5e94\u7684\u201c\u6587\u672c Token\u201d\u4e4b\u95f4\u7684\u903b\u8f91\u8054\u7cfb\uff0c\u6700\u7ec8\u8f93\u51fa\u7b54\u6848\u3002<\/li>\n<\/ul>\n\n\n\n<h4 class=\"wp-block-heading\">2. BLIP-2 \u67b6\u6784\u7684\u7279\u522b\u4e4b\u5904<\/h4>\n\n\n\n<p><strong>BLIP-2<\/strong> \u5f15\u5165\u4e86 <strong>Q-Former<\/strong> \u6a21\u5757\u4f5c\u4e3a\u5176\u201c\u8fde\u63a5\u5668\u201d\u3002<\/p>\n\n\n\n<p>Q-Former \u4e0d\u518d\u50cf LLaVA \u505a\u5168\u91cf\u89c6\u89c9\u7279\u5f81\u7ffb\u8bd1\uff0c\u5b83\u7684\u4f5c\u7528\u662f\u89e3\u51b3<strong>\u4fe1\u606f\u74f6\u9888<\/strong>\u3002<\/p>\n\n\n\n<p>\u5b83\u5c06\u6d77\u91cf\u7684\u3001\u5206\u8fa8\u7387\u65e0\u5173\u7684\u89c6\u89c9\u7279\u5f81\u538b\u7f29\u6210\u56fa\u5b9a\u957f\u5ea6\u7684\u8bed\u4e49\u7247\u6bb5\uff08\u5982 32 \u4e2a Token\uff09\uff0c\u6781\u5927\u964d\u4f4e\u4e86\u540e\u7eed LLM \u7684\u8ba1\u7b97\u538b\u529b \u3002<\/p>\n\n\n\n<p>\u901a\u8fc7 32 \u4e2a<strong>\u67e5\u8be2\u5411\u91cf\uff08Queries\uff09<\/strong>\uff0c\u53bb\u89c6\u89c9\u7279\u5f81\u91cc\u201c\u63d0\u95ee\u201d\uff0c\u5b83\u4eec\u50cf 32 \u4e2a\u5e26\u7740\u4e0d\u540c\u4efb\u52a1\u7684\u4e13\u5bb6\uff0c\u53bb\u56fe\u50cf\u4e2d\u5bfb\u627e\u7b54\u6848\uff0c\u53ea\u5bfb\u627e\u4e0e\u5f53\u524d\u6587\u672c\u6700\u76f8\u5173\u7684\u89c6\u89c9\u7ec6\u8282\uff08\u5982\u201c\u56fe\u4e2d\u732b\u662f\u4ec0\u4e48\u989c\u8272\uff1f\u201d\u3001\u201d\u80cc\u666f\u662f\u4ec0\u4e48\uff1f\u201c\uff09\uff0cQueries \u4f1a\u628a\u4e0a\u767e\u4e07\u50cf\u7d20\u7684\u5197\u4f59\u4fe1\u606f\u538b\u7f29\u6210 32 \u4e2a\u7cbe\u534e Token \u5582\u7ed9 LLM\uff0c\u6781\u5927\u5730\u51cf\u5c11\u4e86\u5197\u4f59\u4fe1\u606f\u7684\u5e72\u6270\u3002<\/p>\n\n\n\n<h4 class=\"wp-block-heading\">3. \u8bad\u7ec3\u7684\u4e24\u9636\u6bb5\uff1a\u4ece\u201c\u8bc6\u5b57\u201d\u5230\u201c\u5bf9\u8bdd\u201d<\/h4>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u9636\u6bb5\u4e00\uff1a\u7279\u5f81\u5bf9\u9f50\uff08\u5bf9\u9f50\uff09<\/strong>\u3002\u51bb\u7ed3\u201c\u773c\u775b\u201d\u548c\u201c\u5927\u8111\u201d\uff0c\u53ea\u8bad\u7ec3\u201c\u8fde\u63a5\u5668\u201d\u3002\u76ee\u6807\u662f\u8ba9\u6a21\u578b\u5b66\u4f1a\u201c\u770b\u56fe\u8bc6\u5b57\u201d\uff0c\u5efa\u7acb\u8d77\u56fe\u50cf\u5411\u91cf\u4e0e\u6587\u5b57\u5750\u6807\u7684\u57fa\u672c\u5173\u8054\u3002<\/li>\n\n\n\n<li><strong>\u9636\u6bb5\u4e8c\uff1a\u591a\u6a21\u6001\u76d1\u7763\u5fae\u8c03\uff08SFT\uff09<\/strong>\u3002\u89e3\u51bb\u8fde\u63a5\u5668\u548c LLM\uff0c\u5582\u7ed9\u5b83\u4f8b\u5982\u7531 GPT-4 \u6a21\u62df\u751f\u6210\u7684\u3001\u5305\u542b\u590d\u6742\u903b\u8f91\u7684 Q&amp;A \u6570\u636e\u3002\u6b64\u65f6\uff0c\u5927\u8111\u624d\u771f\u6b63\u5f00\u59cb\u5b66\u4e60\u5982\u4f55\u9075\u5faa\u6307\u4ee4\u89c2\u5bdf\u56fe\u7247\u7ec6\u8282\uff0c\u5e76\u5b66\u4f1a\u5728\u6ca1\u89c1\u8fc7\u7684\u590d\u6742\u73af\u5883\u4e0b\u4e3e\u4e00\u53cd\u4e09\u3002<\/li>\n<\/ul>\n\n\n\n<h3 class=\"wp-block-heading\">\u4e09\u3001\u539f\u751f\u67b6\u6784\u7684\u5de5\u4f5c\u539f\u7406<\/h3>\n\n\n\n<p>2024 \u5e74\uff0c<strong>GPT-4o<\/strong>\u3001<strong>Emu3<\/strong> \u548c <strong>Baichuan-omni<\/strong> \u6a21\u578b\u7684\u8bde\u751f\uff0c\u6807\u5fd7\u7740\u539f\u751f\u591a\u6a21\u6001\uff08Native Multimodality\uff09\u65f6\u4ee3\u7684\u5230\u6765\u3002<\/p>\n\n\n\n<p>\u5728\u8fd9\u79cd\u67b6\u6784\u4e2d\uff0c\u4e0d\u518d\u9700\u8981\u590d\u6742\u7684\u201c\u8fde\u63a5\u5668\uff08\u7ffb\u8bd1\uff09\u201d\u73af\u8282\u3002<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li><strong>\u7edf\u4e00\u7684\u795e\u7ecf\u7f51\u7edc<\/strong>\uff1a\u6a21\u578b\u662f<strong>\u8de8\u6a21\u6001\u7aef\u5230\u7aef\u9884\u8bad\u7ec3<\/strong>\u7684\uff0c\u6240\u6709\u6a21\u6001\u5728\u540c\u4e00\u4e2a\u795e\u7ecf\u7f51\u7edc\u91cc\u88ab\u5904\u7406\uff0c\u56fe\u50cf\u3001\u6587\u672c\u751a\u81f3\u97f3\u9891\u5728\u6a21\u578b\u8bde\u751f\u7684\u7b2c\u4e00\u5929\uff0c\u5c31\u88ab\u6df7\u5408\u5728\u540c\u4e00\u4e2a Transformer \u5e8f\u5217\u4e2d\u5904\u7406\u3002\u8fd9\u79cd\u201c\u6bcd\u8bed\u7ea7\u201d\u7684\u878d\u5408\u610f\u5473\u7740\u6a21\u578b\u4e0d\u518d\u9700\u8981\u50cf\u6a21\u5757\u5316\u67b6\u6784\u90a3\u6837\u53bb\u201c\u5bf9\u9f50\u201d\u4e24\u4e2a\u9884\u8bad\u7ec3\u597d\u7684\u7a7a\u95f4\uff0c\u800c\u662f\u5728\u751f\u957f\u8fc7\u7a0b\u4e2d\u81ea\u53d1\u5f62\u6210\u4e86\u591a\u7ef4\u611f\u5b98\u7684\u5173\u8054\u3002<\/li>\n\n\n\n<li><strong>Next-Token Prediction\uff08\u9884\u6d4b\u4e0b\u4e00\u4e2a Token\uff09<\/strong>\uff1a\u8fd9\u662f\u539f\u751f\u6a21\u578b\u7684\u6838\u5fc3\u4fe1\u4ef0\u3002\u5728 Emu3 \u6216 GPT-4o \u770b\u6765\uff0c\u751f\u6210\u4e00\u6bb5\u6587\u672c\u8fd8\u662f\u751f\u6210\u4e00\u4e2a\u50cf\u7d20\u5757\uff0c\u672c\u8d28\u4e0a\u90fd\u662f\u9884\u6d4b\u5e8f\u5217\u4e2d\u4e0b\u4e00\u4e2a Token \u51fa\u73b0\u7684\u6982\u7387\u3002\u8fd9\u79cd\u9ad8\u5ea6\u7684\u4e00\u81f4\u6027\u8ba9\u6a21\u578b\u83b7\u5f97\u4e86\u7c7b\u4f3c\u4eba\u7c7b\u7684\u201c\u611f\u5b98\u76f4\u89c9\u201d\u3002<\/li>\n<\/ol>\n\n\n\n<h3 class=\"wp-block-heading\">\u56db\u3001\u4e3e\u4f8b\u8bb2\u89e3\u4e3a\u4ec0\u4e48\u6a21\u578b\u80fd\u8bfb\u61c2\u4e00\u5f20\u56fe\u7247<\/h3>\n\n\n\n<p>\u5047\u8bbe\u6709\u8fd9\u6837\u4e00\u5f20\u56fe\u7247\uff1a\u4e00\u4e2a\u91d1\u53d1\u5973\u5b69\uff0c\u8eab\u7a7f\u7c89\u8272\u788e\u82b1\u767e\u8936\u88d9\uff0c\u624b\u62ff\u4e00\u676f\u5192\u70ed\u6c14\u7684\u62ff\u94c1\uff0c\u7ad9\u5728\u5199\u6709\u201c\u9999\u69ad\u4e3d\u820d\u201d\u7684\u8def\u724c\u4e0b\u3002\u90a3\u4e48\uff0c<strong>\u6a21\u578b\u662f\u5982\u4f55\u7406\u89e3\u56fe\u7247\u4e2d\u7684\u7ec6\u8282\u5462\uff1f<\/strong><\/p>\n\n\n\n<h4 class=\"wp-block-heading\">1. \u4f20\u7edf\u6a21\u5757\u5316\u67b6\u6784\u7684\u5de5\u4f5c\u6d41<\/h4>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u611f\u77e5<\/strong>\uff1a\u89c6\u89c9\u7f16\u7801\u5668\u63d0\u53d6\u51fa\u6d77\u91cf\u7684\u50cf\u7d20\u7279\u5f81\uff08\u5982\u788e\u82b1\u7eb9\u7406\u3001\u6587\u5b57\u7b14\u753b\uff09\u3002<\/li>\n\n\n\n<li><strong>\u8f6c\u8bd1<\/strong>\uff1a\u8fde\u63a5\u5668\u5c06\u8fd9\u4e9b\u7279\u5f81\u6253\u5305\u6210 LLM \u719f\u6089\u7684\u5411\u91cf\u3002<\/li>\n\n\n\n<li><strong>\u63a8\u7406<\/strong>\uff1aLLM \u63a5\u6536\u8fd9\u4e9b\u5411\u91cf\uff0c\u7ed3\u5408\u9884\u8bad\u7ec3\u4e2d\u5b66\u5230\u7684\u77e5\u8bc6\uff0c\u5728\u5927\u8111\u4e2d\u5c06\u5176\u201c\u805a\u7c7b\u201d\u5230\u5bf9\u5e94\u7684\u8bed\u4e49\u70b9\uff1a\u91d1\u8272\u7684\u5411\u91cf \u2248 \u201c\u53d1\u8272\u201d\uff0c\u788e\u82b1\u5411\u91cf \u2248 \u201c\u88d9\u5b50\u6b3e\u5f0f\u201d\u3002<\/li>\n\n\n\n<li><strong>\u4f4d\u7f6e\u4e0e\u6587\u5b57<\/strong>\uff1a\u5229\u7528<strong>\u4f4d\u7f6e\u7f16\u7801<\/strong>\u8bc6\u522b\u7269\u4f53\u5750\u6807\uff0c\u901a\u8fc7<strong>\u5bf9\u9f50\u9884\u8bad\u7ec3<\/strong>\u8bc6\u522b\u51fa\u8def\u724c\u7b14\u753b\u5bf9\u5e94\u7279\u5b9a\u7684\u5730\u540d\u8bcd\u6c47\u3002<\/li>\n<\/ul>\n\n\n\n<blockquote class=\"wp-block-quote is-layout-flow wp-block-quote-is-layout-flow\">\n<ol class=\"wp-block-list\">\n<li>\u4e3a\u4ec0\u4e48\u80fd\u770b\u6e05\u53d1\u578b\u548c\u6b3e\u5f0f\uff1f<\/li>\n<\/ol>\n\n\n\n<ul class=\"wp-block-list\">\n<li>LLaVA-1.5-HD \u6280\u672f\u4f1a\u5c06\u56fe\u7247\u5207\u6210\u66f4\u7ec6\u7684\u7f51\u683c\u3002\u539f\u672c\u6a21\u7cca\u7684\u50cf\u7d20\u56e2\u5757\u5728\u653e\u5927\u540e\uff0c\u5c55\u73b0\u51fa\u4e86\u53d1\u4e1d\u7684\u6ce2\u6d6a\u548c\u88d9\u5b50\u7684\u8936\u76b1 \u3002\u901a\u8fc7\u5927\u89c4\u6a21\u8bad\u7ec3\uff0c\u6a21\u578b\u5185\u90e8\u5df2\u7ecf\u5efa\u7acb\u8d77\u4e00\u79cd\u201c\u805a\u7c7b\u201d\uff1a\u5f53\u8fd9\u5806\u7279\u5b9a\u7684\u7eb9\u7406\u5411\u91cf\u51fa\u73b0\u65f6\uff0c\u5b83\u5728\u6570\u5b66\u5750\u6807\u4e0a\u4f1a\u7cbe\u51c6\u5730\u843d\u5411\u201c\u6ce2\u897f\u7c73\u4e9a\u98ce\u201d\u6216\u201c\u767e\u8936\u88d9\u201d\u7684\u8bed\u4e49\u533a\u57df\u3002<\/li>\n<\/ul>\n\n\n\n<ol start=\"2\" class=\"wp-block-list\">\n<li>\u4e3a\u4ec0\u4e48\u77e5\u9053\u8c01\u5728\u524d\u8c01\u5728\u540e\uff1f<\/li>\n<\/ol>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u5728\u6a21\u5757\u5316\u67b6\u6784\uff08\u5982 LLaVA\uff09\u4e2d\uff0c\u4f4d\u7f6e\u4fe1\u606f\u4e3b\u8981\u901a\u8fc7\u89c6\u89c9\u7f16\u7801\u5668\uff08ViT\uff09\u7684 <strong>1D \u4f4d\u7f6e\u7f16\u7801<\/strong> \u63d0\u4f9b \u3002\u800c\u7a7a\u95f4\u611f\uff08\u8c01\u5728\u524d\u8c01\u5728\u540e\uff09\u7684\u4ea7\u751f\uff0c\u66f4\u591a\u5f52\u529f\u4e8e<strong>\u591a\u6a21\u6001\u6307\u4ee4\u5fae\u8c03\u9636\u6bb5<\/strong>\u5f15\u5165\u7684\u5305\u542b Bounding Box\uff08\u8fb9\u754c\u6846\uff09\u7684\u8bad\u7ec3\u6570\u636e\uff0c\u8ba9 LLM \u5b66\u4f1a\u4e86\u5c06\u7279\u5b9a\u7684 Token ID \u4e0e\u753b\u9762\u5750\u6807\u5173\u8054\u8d77\u6765 \u3002<\/li>\n<\/ul>\n\n\n\n<ol start=\"3\" class=\"wp-block-list\">\n<li>\u4e3a\u4ec0\u4e48\u80fd\u8bfb\u51fa\u8def\u724c\u4e0a\u7684\u6587\u5b57\uff1f<\/li>\n<\/ol>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u65e9\u671f\u7684\u6a21\u5757\u5316\u6a21\u578b\uff08\u5982 LLaVA v1.0\uff09\u5728\u9884\u8bad\u7ec3\u9636\u6bb5\u901a\u5e38\u662f<strong>\u51bb\u7ed3 LLM<\/strong> \u7684\uff0c\u8fd9\u4f1a\u5bfc\u81f4\u6a21\u578b\u7f3a\u4e4f\u201c\u4e0a\u4e0b\u6587\u5b66\u4e60\u201d\u80fd\u529b \u3002\u73b0\u5728\u7684\u8d8b\u52bf\uff08\u5982 VILA \u548c LLaVA-1.5\uff09\u503e\u5411\u4e8e\u5728\u9884\u8bad\u7ec3\u9636\u6bb5\u5c31<strong>\u89e3\u51bb LLM<\/strong>\uff0c\u4ece\u800c\u5b9e\u73b0\u201c\u6df1\u5ea6\u5bf9\u9f50\u201d\uff0c\u901a\u8fc7\u5728\u56fe\u6587\u4ea4\u9519\u8bed\u6599\u4e2d\u505a\u9884\u8bad\u7ec3\uff0cLLM \u7684\u8bed\u8a00\u4e2d\u67a2\u4f1a\u76f4\u63a5\u4ecb\u5165\u89c6\u89c9\u7406\u89e3 \u3002\u5f53\u8def\u724c\u4e0a\u7684\u5b57\u6bcd\u7ebf\u6761\u88ab Token \u5316\u540e\uff0c\u5b83\u4eec\u7684\u5411\u91cf\u7279\u5f81\u76f4\u63a5\u547d\u4e2d\u4e86\u8bcd\u8868\u91cc\u201c\u9999\u69ad\u4e3d\u820d\u201d\u5bf9\u5e94\u7684\u4f4d\u7f6e\uff0c\u5b9e\u73b0\u4e86\u4e0d\u4f9d\u8d56\u989d\u5916 OCR \u63d2\u4ef6\u7684\u201c\u5929\u751f\u8bc6\u5b57\u201d\u3002<\/li>\n<\/ul>\n<\/blockquote>\n\n\n\n<h4 class=\"wp-block-heading\">2. \u539f\u751f\u7aef\u5230\u7aef\u67b6\u6784\u7684\u5de5\u4f5c\u6d41<\/h4>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u5168\u80fd\u63a5\u6536<\/strong>\uff1a\u6a21\u578b\u901a\u8fc7 Vision Tokenizer \u76f4\u63a5\u5c06\u201c\u788e\u82b1\u88d9\u5973\u5b69\u201d\u548c\u201c\u8def\u724c\u201d\u8f6c\u6362\u6210\u5b83\u719f\u6089\u7684<strong>\u539f\u751f\u89c6\u89c9 Token<\/strong>\u3002<\/li>\n\n\n\n<li><strong>\u539f\u751f\u7406\u89e3<\/strong>\uff1a\u5b83\u4e0d\u518d\u9700\u8981\u201c\u7ffb\u8bd1\u201d\uff0c\u800c\u662f\u76f4\u63a5\u5728\u540c\u4e00\u4e2a Transformer \u5e8f\u5217\u4e2d\u8ba1\u7b97\u3002\u5973\u5b69\u7684\u53d1\u8272\u3001\u62ff\u94c1\u7684\u70ed\u6c14\u3001\u751a\u81f3\u8def\u724c\u4e0a\u7684\u5b57\u7b26\uff0c\u5bf9\u6a21\u578b\u6765\u8bf4\u5c31\u662f\u5b83\u51fa\u751f\u65f6\u5c31\u5728\u5b66\u7684\u201c\u6bcd\u8bed\u201d\u3002<\/li>\n\n\n\n<li><strong>\u4f18\u52bf<\/strong>\uff1a\u56e0\u4e3a\u6ca1\u6709\u7ffb\u8bd1\u635f\u8017\uff0c\u539f\u751f\u6a21\u578b\uff08\u5982 GPT-4o\uff09\u80fd\u4ee5\u6781\u4f4e\u5ef6\u8fdf\u7406\u89e3\u56fe\u7247\u4e2d\uff0c\u5982\u201c\u62ff\u94c1\u5192\u51fa\u7684\u70ed\u6c14\u201d\u8fd9\u79cd\u7ec6\u8282\u3002<\/li>\n<\/ul>\n\n\n\n<h3 class=\"wp-block-heading\">\u7ed3\u8bed<\/h3>\n\n\n\n<p>\u771f\u6b63\u7684\u7406\u89e3\u5e76\u4e0d\u4f9d\u8d56\u4e8e\u7279\u5b9a\u7684\u611f\u5b98\uff0c\u800c\u662f\u4f9d\u8d56\u4e8e\u6a21\u578b\u80fd\u5426\u5c06\u5f02\u6784\u7684\u4fe1\u606f\u2014\u2014\u65e0\u8bba\u662f\u50cf\u7d20\u70b9\u3001\u58f0\u6ce2\u9891\u7387\u8fd8\u662f\u6587\u5b57\u5b57\u7b26\u2014\u2014\u90fd\u6620\u5c04\u5230\u540c\u4e00\u5957\u7edf\u4e00\u7684<strong>\u6570\u5b66\u8bed\u4e49\u7a7a\u95f4<\/strong>\u4e2d\u3002<\/p>\n\n\n\n<p>\u591a\u6a21\u6001\u5927\u6a21\u578b\u7684\u7ec8\u6781\u76ee\u6807\uff0c\u662f\u6d88\u9664\u6a21\u6001\u4e4b\u95f4\u7684\u2018\u7ffb\u8bd1\u611f\u2019\uff0c\u8ba9 AI \u80fd\u591f\u50cf\u4eba\u7c7b\u4e00\u6837\uff0c\u4ea7\u751f\u5bf9\u7269\u7406\u4e16\u754c\u7684\u2018\u76f4\u89c9\u6027\u7406\u89e3\u2019\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u591a\u6a21\u6001\u5927\u6a21\u578b\uff08MLLM\uff09\u5df2\u7ecf\u6210\u4e3a\u5fc5\u7136\u7684\u53d1\u5c55\u8d8b\u52bf\uff0c\u6587\u5b57\u91cc\u7684\u4fe1\u606f\u548c\u77e5\u8bc6\u53ea\u662f\u51b0\u5c71\u4e00\u89d2\uff0c\u6a21\u578b\u4e0d\u4ec5\u8981\u80fd\u7406\u89e3\u6587\u5b57\uff0c\u8fd8\u8981\u80fd\u770b [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":225,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"site-sidebar-layout":"default","site-content-layout":"","ast-site-content-layout":"default","site-content-style":"default","site-sidebar-style":"default","ast-global-header-display":"","ast-banner-title-visibility":"","ast-main-header-display":"","ast-hfb-above-header-display":"","ast-hfb-below-header-display":"","ast-hfb-mobile-header-display":"","site-post-title":"","ast-breadcrumbs-content":"","ast-featured-img":"","footer-sml-layout":"","theme-transparent-header-meta":"","adv-header-id-meta":"","stick-header-meta":"","header-above-stick-meta":"","header-main-stick-meta":"","header-below-stick-meta":"","astra-migrate-meta-layouts":"default","ast-page-background-enabled":"default","ast-page-background-meta":{"desktop":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"tablet":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"mobile":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""}},"ast-content-background-meta":{"desktop":{"background-color":"var(--ast-global-color-4)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"tablet":{"background-color":"var(--ast-global-color-4)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"mobile":{"background-color":"var(--ast-global-color-4)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""}},"footnotes":""},"categories":[1],"tags":[],"class_list":["post-223","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-uncategorized"],"_links":{"self":[{"href":"https:\/\/www.xiewuyi.com\/index.php\/wp-json\/wp\/v2\/posts\/223","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.xiewuyi.com\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.xiewuyi.com\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.xiewuyi.com\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.xiewuyi.com\/index.php\/wp-json\/wp\/v2\/comments?post=223"}],"version-history":[{"count":9,"href":"https:\/\/www.xiewuyi.com\/index.php\/wp-json\/wp\/v2\/posts\/223\/revisions"}],"predecessor-version":[{"id":233,"href":"https:\/\/www.xiewuyi.com\/index.php\/wp-json\/wp\/v2\/posts\/223\/revisions\/233"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.xiewuyi.com\/index.php\/wp-json\/wp\/v2\/media\/225"}],"wp:attachment":[{"href":"https:\/\/www.xiewuyi.com\/index.php\/wp-json\/wp\/v2\/media?parent=223"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.xiewuyi.com\/index.php\/wp-json\/wp\/v2\/categories?post=223"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.xiewuyi.com\/index.php\/wp-json\/wp\/v2\/tags?post=223"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}