{"id":2384,"date":"2024-03-10T21:40:25","date_gmt":"2024-03-10T13:40:25","guid":{"rendered":"https:\/\/www.aqwu.net\/wp\/?p=2384"},"modified":"2024-04-28T20:03:24","modified_gmt":"2024-04-28T12:03:24","slug":"%e4%ba%86%e8%a7%a3-openai-%e7%9a%84-gpt2-%e6%a8%a1%e5%9e%8b%e7%bb%93%e6%9e%843","status":"publish","type":"post","link":"https:\/\/www.aqwu.net\/wp\/?p=2384","title":{"rendered":"\u4e86\u89e3 OpenAI \u7684 gpt2 \u6a21\u578b\u7ed3\u6784(3)"},"content":{"rendered":"\n<h2 class=\"wp-block-heading\"><strong>10. \u4f4d\u7f6e\u7f16\u7801<\/strong><\/h2>\n\n\n\n<p><code>transformer.wpe.weight<\/code>\u4ee3\u8868\u7684\u662f\u4f4d\u7f6e\u7f16\u7801\uff08Positional Encoding\uff09\u7684\u6743\u91cd\u77e9\u9635\uff0c\u5728Transformer\u6a21\u578b\u4e2d\u975e\u5e38\u91cd\u8981\u3002\u8fd9\u4e2a\u77e9\u9635\u7684\u5927\u5c0f\u4e3a<code>torch.Size([1024, 768])<\/code>\uff0c\u8fd9\u610f\u5473\u7740\u5b83\u53ef\u4ee5\u8868\u793a\u6700\u591a1024\u4e2a\u4f4d\u7f6e\uff08\u5373\u5e8f\u5217\u7684\u6700\u5927\u957f\u5ea6\u4e3a1024\uff09\uff0c\u6bcf\u4e2a\u4f4d\u7f6e\u88ab\u7f16\u7801\u4e3a\u4e00\u4e2a768\u7ef4\u7684\u5411\u91cf\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\"><strong>10.1 \u4e3a\u4ec0\u4e48\u9700\u8981\u4f4d\u7f6e\u7f16\u7801\uff1f<\/strong><\/h3>\n\n\n\n<p>Transformer\u6a21\u578b\u662f\u57fa\u4e8e\u81ea\u6ce8\u610f\u529b\u673a\u5236\u7684\uff0c\u8fd9\u610f\u5473\u7740\u5b83\u4eec\u5728\u5904\u7406\u5e8f\u5217\u6570\u636e\u65f6\uff0c\u6ca1\u6709\u50cf\u5faa\u73af\u795e\u7ecf\u7f51\u7edc\uff08RNN\uff09\u90a3\u6837\u7684\u9010\u6b65\u8ba1\u7b97\u8fc7\u7a0b\u6765\u81ea\u7136\u5730\u6355\u6349\u987a\u5e8f\u4fe1\u606f\u3002\u4e3a\u4e86\u8ba9\u6a21\u578b\u80fd\u591f\u7406\u89e3\u5355\u8bcd\u5728\u53e5\u5b50\u4e2d\u7684\u987a\u5e8f\uff0cTransformer\u5f15\u5165\u4e86\u4f4d\u7f6e\u7f16\u7801\u3002\u901a\u8fc7\u5c06\u4f4d\u7f6e\u7f16\u7801\u6dfb\u52a0\u5230\u6bcf\u4e2a\u5355\u8bcd\u7684\u5d4c\u5165\u5411\u91cf\u4e2d\uff0c\u6a21\u578b\u5c31\u80fd\u591f\u5229\u7528\u8fd9\u4e9b\u4fe1\u606f\u6765\u7406\u89e3\u8bcd\u5e8f\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\"><strong>10.2 \u4f4d\u7f6e\u7f16\u7801\u7684\u5de5\u4f5c\u539f\u7406<\/strong><\/h3>\n\n\n\n<p>\u4f4d\u7f6e\u7f16\u7801\u901a\u5e38\u662f\u901a\u8fc7\u4e00\u7ec4\u56fa\u5b9a\u7684\u51fd\u6570\u6765\u751f\u6210\u7684\uff0c\u4f8b\u5982\u4f7f\u7528\u6b63\u5f26\u548c\u4f59\u5f26\u51fd\u6570\u7684\u7ec4\u5408\uff0c\u6bcf\u4e2a\u7ef4\u5ea6\u5bf9\u5e94\u4e00\u4e2a\u51fd\u6570\u3002\u8fd9\u6837\u505a\u7684\u76ee\u7684\u662f\u4e3a\u4e86\u8ba9\u6bcf\u4e2a\u4f4d\u7f6e\u7684\u7f16\u7801\u5177\u6709\u72ec\u4e00\u65e0\u4e8c\u7684\u6a21\u5f0f\uff0c\u4e14\u4e0d\u540c\u4f4d\u7f6e\u4e4b\u95f4\u7684\u7f16\u7801\u662f\u53ef\u533a\u5206\u7684\u3002<\/p>\n\n\n\n<p>\u5728\u539f\u59cb\u7684Transformer\u6a21\u578b\u4e2d\uff0c\u4f4d\u7f6e\u7f16\u7801\u662f\u76f4\u63a5\u52a0\u5230\u4e86\u8bcd\u5d4c\u5165\u5411\u91cf\u4e0a\u7684\u3002\u800c\u5728\u4e00\u4e9b\u53d8\u79cd\u4e2d\uff0c\u5982GPT-2\uff0c\u4f4d\u7f6e\u7f16\u7801\u53ef\u80fd\u4f1a\u6709\u4e00\u4e9b\u5fae\u8c03\u6216\u8005\u4f7f\u7528\u5b66\u4e60\u5230\u7684\u4f4d\u7f6e\u5d4c\u5165\u4ee3\u66ff\u56fa\u5b9a\u7684\u6570\u5b66\u51fd\u6570\u751f\u6210\u7684\u7f16\u7801\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\"><strong>10.3 <code>transformer.wpe.weight<\/code>\u7684\u5177\u4f53\u610f\u4e49<\/strong><\/h3>\n\n\n\n<p>\u5bf9\u4e8e<code>transformer.wpe.weight<\/code>\uff0c\u8fd9\u91cc\u76841024\u4ee3\u8868\u6a21\u578b\u53ef\u4ee5\u5904\u7406\u7684<strong>\u6700\u5927\u5e8f\u5217\u957f\u5ea6<\/strong>\uff0c\u800c768\u4ee3\u8868\u6bcf\u4e2a\u4f4d\u7f6e\u7f16\u7801\u7684\u7ef4\u5ea6\uff0c\u4e0e\u8bcd\u5d4c\u5165\u5411\u91cf\u7684\u7ef4\u5ea6\u76f8\u5339\u914d\u3002\u8fd9\u4f7f\u5f97\u6a21\u578b\u80fd\u591f\u5c06\u4f4d\u7f6e\u4fe1\u606f\u76f4\u63a5\u52a0\u5230\u8bcd\u5d4c\u5165\u5411\u91cf\u4e0a\uff0c\u8fdb\u800c\u5728\u5904\u7406\u6587\u672c\u65f6\u8003\u8651\u5230\u5355\u8bcd\u7684\u4f4d\u7f6e\u4fe1\u606f\u3002<\/p>\n\n\n\n<p>\u901a\u8fc7\u5b66\u4e60\u7684\u4f4d\u7f6e\u7f16\u7801\uff08\u800c\u975e\u56fa\u5b9a\u751f\u6210\u7684\uff09\uff0c\u6a21\u578b\u80fd\u591f\u66f4\u7075\u6d3b\u5730\u9002\u5e94\u4e0d\u540c\u7684\u8bed\u8a00\u548c\u4efb\u52a1\u9700\u6c42\uff0c\u4ece\u800c\u6355\u6349\u5230\u66f4\u590d\u6742\u7684\u987a\u5e8f\u4f9d\u8d56\u6027\u3002\u8fd9\u662f\u73b0\u4ee3Transformer\u6a21\u578b\uff0c\u5982GPT\u7cfb\u5217\uff0c\u5728\u5904\u7406\u957f\u5e8f\u5217\u6570\u636e\u65f6\u53d6\u5f97\u5353\u8d8a\u6027\u80fd\u7684\u5173\u952e\u56e0\u7d20\u4e4b\u4e00\u3002<\/p>\n\n\n\n<p>\u5728Transformer\u6a21\u578b\u4e2d\uff0c<code>transformer.wpe.weight<\/code>\u4ee3\u8868\u7684\u662f\u4f4d\u7f6e\u7f16\u7801\uff08Positional Encoding\uff09\u7684\u6743\u91cd\u77e9\u9635\u3002\u8fd9\u4e2a\u77e9\u9635\u7684\u5927\u5c0f\u4e3a<code>torch.Size([1024, 768])<\/code>\uff0c\u610f\u5473\u7740\u5b83\u53ef\u4ee5\u4e3a\u6700\u591a1024\u4e2a\u4f4d\u7f6e\uff08\u5373\u5e8f\u5217\u957f\u5ea6\uff09\u7684\u6bcf\u4e00\u4e2a\u4f4d\u7f6e\u63d0\u4f9b\u4e00\u4e2a768\u7ef4\u7684\u7f16\u7801\u5411\u91cf\u3002\u8fd9\u91cc\u7684768\u7ef4\u5bf9\u5e94\u4e8e\u6a21\u578b\u7684\u9690\u85cf\u5c42\u5927\u5c0f\u3002<\/p>\n\n\n\n<p>\u4f4d\u7f6e\u7f16\u7801\u662fTransformer\u6a21\u578b\u7684\u4e00\u4e2a\u5173\u952e\u90e8\u5206\uff0c\u5b83\u5141\u8bb8\u6a21\u578b\u7406\u89e3\u5e8f\u5217\u4e2d\u5355\u8bcd\u7684\u987a\u5e8f\u3002\u7531\u4e8eTransformer\u6a21\u578b\u672c\u8eab\u5e76\u4e0d\u5904\u7406\u987a\u5e8f\u4fe1\u606f\uff08\u5b83\u4f9d\u8d56\u4e8e\u81ea\u6ce8\u610f\u529b\u673a\u5236\u6765\u5904\u7406\u8f93\u5165\u5e8f\u5217\uff09\uff0c\u4f4d\u7f6e\u7f16\u7801\u901a\u8fc7\u4e3a\u6bcf\u4e2a\u8f93\u5165\u5143\u7d20\u6dfb\u52a0\u4e00\u4e2a\u4e0e\u5176\u4f4d\u7f6e\u76f8\u5173\u7684\u5411\u91cf\u6765\u8865\u5145\u8fd9\u79cd\u987a\u5e8f\u4fe1\u606f\u3002<\/p>\n\n\n\n<p>\u4f4d\u7f6e\u7f16\u7801\u5411\u91cf\u662f\u9884\u5148\u8ba1\u7b97\u7684\uff0c\u5e76\u4e14\u5728\u6a21\u578b\u8bad\u7ec3\u548c\u63a8\u7406\u65f6\u4fdd\u6301\u4e0d\u53d8\u3002\u8fd9\u4e9b\u5411\u91cf\u901a\u5e38\u5305\u542b\u6b63\u5f26\u548c\u4f59\u5f26\u51fd\u6570\u7684\u503c\uff0c\u8fd9\u4e9b\u51fd\u6570\u7684\u9891\u7387\u968f\u4f4d\u7f6e\u800c\u53d8\u5316\uff0c\u4ece\u800c\u5141\u8bb8\u6a21\u578b\u533a\u5206\u4e0d\u540c\u4f4d\u7f6e\u7684\u8f93\u5165\u3002<\/p>\n\n\n\n<p>\u56e0\u4e3a\u4f60\u60f3\u67e5\u770b<code>transformer.wpe.weight<\/code>\u7684\u5185\u5bb9\uff0c\u4e0b\u9762\u662f\u5982\u4f55\u7528\u4ee3\u7801\u6765\u5b9e\u73b0\u7684\u7b80\u5355\u793a\u4f8b\uff0c\u5047\u8bbe\u4f60\u5df2\u7ecf\u52a0\u8f7d\u4e86\u4e00\u4e2aTransformer\u6a21\u578b\uff08\u4f8b\u5982GPT-2\uff09,\u547d\u540d\u4e3a test08.py\uff0c\u6587\u4ef6\u4fdd\u5b58\u5230 newsrc \u76ee\u5f55\u4e0b\uff1a<\/p>\n\n\n\n<div class=\"wp-block-urvanov-syntax-highlighter-code-block\"><pre class=\"lang:python decode:true \" >from transformers import GPT2LMHeadModel\n\n# \u52a0\u8f7d\u9884\u8bad\u7ec3\u7684GPT-2\u6a21\u578b\nmodel = GPT2LMHeadModel.from_pretrained('gpt2')\n\n# \u76f4\u63a5\u8bbf\u95ee\u4f4d\u7f6e\u7f16\u7801\u6743\u91cd\npositional_encoding_weight = model.transformer.wpe.weight\n\nprint(\"Positional encoding weight:\")\nprint(positional_encoding_weight)\n<\/pre><\/div>\n\n\n\n<p>\u8bf7\u6ce8\u610f\uff0c<code>positional_encoding_weight<\/code>\u7684\u5927\u5c0f\u5e94\u8be5\u662f<code>torch.Size([1024, 768])<\/code>\uff0c\u4e0e\u4f60\u4e4b\u524d\u63d0\u5230\u7684\u76f8\u7b26\u3002\u8fd9\u4e2a\u77e9\u9635\u4e3a\u6a21\u578b\u4e2d\u7684\u6bcf\u4e2a\u53ef\u80fd\u4f4d\u7f6e\u63d0\u4f9b\u4e86\u4e00\u4e2a768\u7ef4\u7684\u7f16\u7801\u5411\u91cf\uff0c\u603b\u5171\u652f\u6301\u957f\u5ea6\u4e3a1024\u7684\u5e8f\u5217\u3002\u8fd9\u4e9b\u7f16\u7801\u5e2e\u52a9\u6a21\u578b\u7406\u89e3\u5355\u8bcd\u5728\u53e5\u5b50\u4e2d\u7684\u76f8\u5bf9\u6216\u7edd\u5bf9\u4f4d\u7f6e\uff0c\u662fTransformer\u67b6\u6784\u5904\u7406\u5e8f\u5217\u6570\u636e\u7684\u5173\u952e\u673a\u5236\u4e4b\u4e00\u3002<\/p>\n\n\n\n<p>\u8fd0\u884ctest08.py<\/p>\n\n\n\n<div class=\"wp-block-urvanov-syntax-highlighter-code-block\"><pre class=\"lang:python decode:true \" >python newsrc\/test08.py\nPositional encoding weight:\nParameter containing:\ntensor([[-1.8821e-02, -1.9742e-01,  4.0267e-03,  ..., -4.3044e-02,\n          2.8267e-02,  5.4490e-02],\n        [ 2.3959e-02, -5.3792e-02, -9.4879e-02,  ...,  3.4170e-02,\n          1.0172e-02, -1.5573e-04],\n        [ 4.2161e-03, -8.4764e-02,  5.4515e-02,  ...,  1.9745e-02,\n          1.9325e-02, -2.1424e-02],\n        ...,\n        [-1.7987e-03,  1.6052e-03, -5.5103e-02,  ...,  1.3617e-02,\n         -7.1805e-03,  3.7552e-03],\n        [ 3.2105e-03,  1.5501e-03, -4.8944e-02,  ...,  2.0725e-02,\n         -1.1838e-02, -5.5683e-04],\n        [ 2.6610e-04,  3.0272e-03, -1.7086e-03,  ..., -4.6506e-03,\n         -2.3541e-03, -5.7855e-03]], requires_grad=True)<\/pre><\/div>\n\n\n\n<h2 class=\"wp-block-heading\"><strong>11. \u4f4d\u7f6e\u7f16\u7801(wpe)\u548c\u8bcd\u5d4c\u5165(wte)\u7684\u5173\u7cfb<\/strong><\/h2>\n\n\n\n<p>\u5728Transformer\u6a21\u578b\u4e2d\uff0c\u4f4d\u7f6e\u7f16\u7801\uff08Positional Encoding\uff09\u548c\u8bcd\u5d4c\u5165\uff08Word Embeddings\uff09\u8d77\u7740\u4e0d\u540c\u4f46\u4e92\u8865\u7684\u4f5c\u7528\uff0c\u5171\u540c\u4e3a\u6a21\u578b\u63d0\u4f9b\u8f93\u5165\u5e8f\u5217\u7684\u8868\u793a\u3002\u5b83\u4eec\u662f\u72ec\u7acb\u7684\uff0c\u4f46\u5728\u6a21\u578b\u5904\u7406\u8f93\u5165\u65f6\u4f1a\u88ab\u76f8\u52a0\u5408\u5e76\uff0c\u8fd9\u6837\u6bcf\u4e2a\u8bcd\u4e0d\u4ec5\u643a\u5e26\u4e86\u81ea\u8eab\u7684\u8bed\u4e49\u4fe1\u606f\uff0c\u8fd8\u5305\u542b\u4e86\u5176\u5728\u5e8f\u5217\u4e2d\u7684\u4f4d\u7f6e\u4fe1\u606f\u3002\u4e0b\u9762\u662f\u5b83\u4eec\u5982\u4f55\u5173\u8054\u548c\u4f5c\u7528\u7684\u8be6\u7ec6\u89e3\u91ca\uff1a<\/p>\n\n\n\n<h3 class=\"wp-block-heading\"><strong>11.1 \u8bcd\u5d4c\u5165\uff08Word Embeddings\uff09<\/strong><\/h3>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u4f5c\u7528<\/strong>\uff1a\u8bcd\u5d4c\u5165\u662f\u9ad8\u7ef4\u7a7a\u95f4\u4e2d\u7684\u5411\u91cf\uff0c\u7528\u4e8e\u8868\u793a\u8bcd\u6c47\u8868\u4e2d\u7684\u6bcf\u4e2a\u5355\u8bcd\u6216token\u3002\u8fd9\u4e9b\u5411\u91cf\u6355\u6349\u4e86\u8bcd\u6c47\u7684\u8bed\u4e49\u7279\u5f81\uff0c\u6bd4\u5982\u610f\u601d\u76f8\u8fd1\u7684\u8bcd\u4f1a\u5728\u8fd9\u4e2a\u7a7a\u95f4\u4e2d\u5f7c\u6b64\u63a5\u8fd1\u3002<\/li>\n\n\n\n<li><strong>\u6765\u6e90<\/strong>\uff1a\u5728\u6a21\u578b\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u5b66\u4e60\u5f97\u5230\uff0c\u901a\u8fc7\u4f18\u5316\u8fc7\u7a0b\u8c03\u6574\u4ee5\u66f4\u597d\u5730\u5b8c\u6210\u7279\u5b9a\u4efb\u52a1\uff08\u6bd4\u5982\u6587\u672c\u5206\u7c7b\u3001\u95ee\u9898\u56de\u7b54\u7b49\uff09\u3002<\/li>\n\n\n\n<li><strong>\u5173\u8054<\/strong>\uff1a\u6bcf\u4e2a\u8f93\u5165\u7684\u5355\u8bcd\u6216token\u90fd\u901a\u8fc7\u67e5\u627e\u8bcd\u5d4c\u5165\u77e9\u9635\uff08\u5982<code>transformer.wte.weight<\/code>\uff09\u6765\u8f6c\u6362\u4e3a\u4e00\u4e2a\u56fa\u5b9a\u957f\u5ea6\u7684\u5411\u91cf\u3002<\/li>\n<\/ul>\n\n\n\n<h3 class=\"wp-block-heading\"><strong>11.2 \u4f4d\u7f6e\u7f16\u7801\uff08Positional Encoding\uff09<\/strong><\/h3>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u4f5c\u7528<\/strong>\uff1a\u4f4d\u7f6e\u7f16\u7801\u63d0\u4f9b\u4e86\u6a21\u578b\u5904\u7406\u5e8f\u5217\u6570\u636e\u65f6\u5fc5\u9700\u7684\u4f4d\u7f6e\u4fe1\u606f\u3002\u7531\u4e8eTransformer\u6a21\u578b\u4f7f\u7528\u7684\u662f\u81ea\u6ce8\u610f\u529b\u673a\u5236\uff0c\u5b83\u672c\u8eab\u4e0d\u7406\u89e3\u8f93\u5165\u4e2d\u5404\u4e2a\u5143\u7d20\u7684\u987a\u5e8f\u3002\u4f4d\u7f6e\u7f16\u7801\u901a\u8fc7\u4e3a\u6a21\u578b\u7684\u6bcf\u4e2a\u8f93\u5165\u4f4d\u7f6e\u6dfb\u52a0\u4e00\u4e2a\u552f\u4e00\u7684\u7f16\u7801\u6765\u89e3\u51b3\u8fd9\u4e2a\u95ee\u9898\u3002<\/li>\n\n\n\n<li><strong>\u6765\u6e90<\/strong>\uff1a\u4f4d\u7f6e\u7f16\u7801\u53ef\u4ee5\u662f\u9884\u5148\u8ba1\u7b97\u597d\u7684\u56fa\u5b9a\u503c\uff08\u5982\u4f7f\u7528\u6b63\u5f26\u548c\u4f59\u5f26\u51fd\u6570\u7684\u56fa\u5b9a\u6a21\u5f0f\uff09\uff0c\u4e5f\u53ef\u4ee5\u662f\u6a21\u578b\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u5b66\u4e60\u5f97\u5230\u7684\u3002<\/li>\n\n\n\n<li><strong>\u5173\u8054<\/strong>\uff1a\u4f4d\u7f6e\u7f16\u7801\u5411\u91cf\u76f4\u63a5\u52a0\u5230\u76f8\u5e94\u4f4d\u7f6e\u7684\u8bcd\u5d4c\u5165\u5411\u91cf\u4e0a\u3002\u8fd9\u6837\uff0c\u6a21\u578b\u7684\u6bcf\u4e2a\u8f93\u5165\u4f4d\u7f6e\u7684\u8868\u793a\u5c31\u540c\u65f6\u5305\u542b\u4e86\u5173\u4e8e\u8fd9\u4e2a\u4f4d\u7f6e\u4e0a\u8bcd\u6c47\u7684\u8bed\u4e49\u4fe1\u606f\u548c\u5176\u5728\u5e8f\u5217\u4e2d\u7684\u4f4d\u7f6e\u4fe1\u606f\u3002<\/li>\n<\/ul>\n\n\n\n<h3 class=\"wp-block-heading\"><strong>11.3 \u5408\u5e76\u8fc7\u7a0b<\/strong><\/h3>\n\n\n\n<p>\u5f53\u8f93\u5165\u5e8f\u5217\u88ab\u9001\u5165\u6a21\u578b\u65f6\uff0c\u6bcf\u4e2atoken\u7684\u8bcd\u5d4c\u5165\u548c\u76f8\u5e94\u4f4d\u7f6e\u7684\u4f4d\u7f6e\u7f16\u7801\u88ab\u76f8\u52a0\uff0c\u5408\u5e76\u6210\u4e00\u4e2a\u589e\u5f3a\u7684\u8868\u793a\uff0c\u65e2\u5305\u542b\u4e86\u8bcd\u6c47\u7684\u8bed\u4e49\u4fe1\u606f\uff0c\u4e5f\u5305\u542b\u4e86\u5176\u4f4d\u7f6e\u4fe1\u606f\u3002\u8fd9\u4e2a\u5408\u5e76\u540e\u7684\u8868\u793a\u968f\u540e\u88ab\u7528\u4f5c\u6a21\u578b\u7684\u8f93\u5165\uff0c\u8fdb\u884c\u540e\u7eed\u7684\u5904\u7406\u548c\u5206\u6790\u3002<\/p>\n\n\n\n<p>\u8fd9\u79cd\u8bbe\u8ba1\u5141\u8bb8Transformer\u6a21\u578b\u6709\u6548\u5730\u5904\u7406\u5e8f\u5217\u6570\u636e\uff0c\u5982\u81ea\u7136\u8bed\u8a00\u6587\u672c\uff0c\u4f7f\u5176\u5728\u5404\u79cdNLP\u4efb\u52a1\u4e2d\u8868\u73b0\u51fa\u8272\u3002\u901a\u8fc7\u8fd9\u79cd\u65b9\u5f0f\uff0c\u6a21\u578b\u80fd\u591f\u57fa\u4e8e\u8bcd\u6c47\u7684\u610f\u4e49\u548c\u5b83\u4eec\u5728\u53e5\u5b50\u4e2d\u7684\u76f8\u5bf9\u4f4d\u7f6e\u6765\u505a\u51fa\u51b3\u7b56\u3002<\/p>\n\n\n\n<p>\u4e3a\u4e86\u5728\u4ee3\u7801\u4e2d\u4f53\u73b0\u8bcd\u5d4c\u5165\uff08Word Embeddings\uff09\u548c\u4f4d\u7f6e\u7f16\u7801\uff08Positional Encoding\uff09\u7684\u5408\u5e76\u8fc7\u7a0b\uff0c\u6211\u5c06\u5c55\u793a\u4e00\u4e2a\u7b80\u5316\u7684\u4f8b\u5b50\uff0c\u4f7f\u7528PyTorch\u548cHugging Face\u7684Transformers\u5e93\u3002\u8fd9\u4e2a\u4f8b\u5b50\u4f1a\u624b\u52a8\u6267\u884c\u8fd9\u4e24\u79cd\u7c7b\u578b\u7684\u7f16\u7801\u7684\u5408\u5e76\uff0c\u6a21\u62df\u5b83\u4eec\u5728\u5b9e\u9645Transformer\u6a21\u578b\u4e2d\u7684\u4ea4\u4e92\u3002<\/p>\n\n\n\n<p>,\u547d\u540d\u4e3a test09.py\uff0c\u6587\u4ef6\u4fdd\u5b58\u5230 newsrc \u76ee\u5f55\u4e0b\uff1a<\/p>\n\n\n\n<div class=\"wp-block-urvanov-syntax-highlighter-code-block\"><pre class=\"lang:python decode:true \" >from transformers import GPT2LMHeadModel, GPT2Tokenizer\nimport torch\n\n# \u52a0\u8f7d\u9884\u8bad\u7ec3\u7684GPT-2\u6a21\u578b\u548c\u5206\u8bcd\u5668\nmodel = GPT2LMHeadModel.from_pretrained('gpt2')\ntokenizer = GPT2Tokenizer.from_pretrained('gpt2')\n\n# \u5b9a\u4e49\u793a\u4f8b\u6587\u672c\ntext = \"The meaning of life is\"\n\n# \u4f7f\u7528\u5206\u8bcd\u5668\u5c06\u6587\u672c\u7f16\u7801\u4e3atoken ID\uff0c\u5e76\u5c06\u5176\u8f6c\u6362\u4e3aPyTorch tensors\ninput_ids = tokenizer.encode(text, return_tensors='pt')\n\n# \u83b7\u53d6\u8bcd\u5d4c\u5165\u77e9\u9635\nword_embeddings = model.transformer.wte.weight\n\n# \u4f7f\u7528\u7d22\u5f15\u4ece\u8bcd\u5d4c\u5165\u77e9\u9635\u4e2d\u63d0\u53d6\u5bf9\u5e94\u7684\u8bcd\u5d4c\u5165\u5411\u91cf\ntoken_embeddings = torch.index_select(word_embeddings, 0, input_ids.view(-1))\n\n# \u83b7\u53d6\u4f4d\u7f6e\u7f16\u7801\u77e9\u9635\nposition_ids = torch.arange(0, input_ids.size(1), dtype=torch.long).unsqueeze(0)\nposition_embeddings = model.transformer.wpe.weight\n\n# \u4f7f\u7528\u7d22\u5f15\u4ece\u4f4d\u7f6e\u7f16\u7801\u77e9\u9635\u4e2d\u63d0\u53d6\u5bf9\u5e94\u7684\u4f4d\u7f6e\u7f16\u7801\u5411\u91cf\npositional_encodings = torch.index_select(position_embeddings, 0, position_ids.view(-1))\n\n# \u5c06\u8bcd\u5d4c\u5165\u5411\u91cf\u548c\u4f4d\u7f6e\u7f16\u7801\u5411\u91cf\u76f8\u52a0\uff0c\u5f97\u5230\u6700\u7ec8\u7684\u8f93\u5165\u8868\u793a\nfinal_input_embeddings = token_embeddings + positional_encodings\n\nprint(\"Final input embeddings shape:\", final_input_embeddings.shape)\n<\/pre><\/div>\n\n\n\n<p>\u8fd9\u6bb5\u4ee3\u7801\u4e2d\u7684\u5173\u952e\u4fee\u6539\u662f\u4f7f\u7528<code>torch.index_select<\/code>\u51fd\u6570\u6765\u7d22\u5f15<code>word_embeddings<\/code>\u548c<code>position_embeddings<\/code>\u77e9\u9635\uff0c\u83b7\u53d6\u5bf9\u5e94\u7684\u8bcd\u5d4c\u5165\u5411\u91cf\u548c\u4f4d\u7f6e\u7f16\u7801\u5411\u91cf\u3002\u8bf7\u6ce8\u610f\uff0c\u8fd9\u91cc\u6211\u4eec\u9700\u8981\u5c06<code>input_ids<\/code>\u548c<code>position_ids<\/code>\u94fa\u5e73\uff08\u4f7f\u7528<code>.view(-1)<\/code>\uff09\uff0c\u56e0\u4e3a<code>torch.index_select<\/code>\u51fd\u6570\u8981\u6c42\u7d22\u5f15\u662f\u4e00\u7ef4\u7684\u3002\u6700\u540e\uff0c\u6211\u4eec\u901a\u8fc7\u7b80\u5355\u5730\u76f8\u52a0\u8fd9\u4e24\u79cd\u5411\u91cf\u6765\u5408\u5e76\u8bcd\u5d4c\u5165\u548c\u4f4d\u7f6e\u7f16\u7801\uff0c\u5f97\u5230\u6700\u7ec8\u7684\u8f93\u5165\u8868\u793a\u3002<\/p>\n\n\n\n<p>\u8fd0\u884ctest09.py<\/p>\n\n\n\n<div class=\"wp-block-urvanov-syntax-highlighter-code-block\"><pre class=\"lang:sh decode:true \" >python newsrc\/test09.py\nFinal input embeddings shape: torch.Size([5, 768])<\/pre><\/div>\n\n\n\n<p>\u6210\u529f\u8fd0\u884c\u4fee\u6b63\u540e\u7684\u4ee3\u7801\u5e76\u5f97\u5230\u4e86\u6700\u7ec8\u8f93\u5165\u8868\u793a\u7684\u5f62\u72b6\u4e3a<code>torch.Size([5, 768])<\/code>\uff0c\u8fd9\u8868\u660e\u4f60\u7684\u8f93\u5165\u6587\u672c<code>\"The meaning of life is\"<\/code>\u88ab\u5206\u8bcd\u6210\u4e865\u4e2atokens\uff0c\u6bcf\u4e2atoken\u90fd\u88ab\u6620\u5c04\u5230\u4e86\u4e00\u4e2a768\u7ef4\u7684\u5411\u91cf\u3002\u8fd9\u4e2a\u5411\u91cf\u65e2\u5305\u542b\u4e86\u8bcd\u5d4c\u5165\uff08\u5373\u8bcd\u7684\u8bed\u4e49\u4fe1\u606f\uff09\u4e5f\u5305\u542b\u4e86\u4f4d\u7f6e\u7f16\u7801\uff08\u5373\u8bcd\u5728\u53e5\u5b50\u4e2d\u7684\u4f4d\u7f6e\u4fe1\u606f\uff09\uff0c\u6b63\u662f\u8fd9\u4e24\u8005\u7684\u7ed3\u5408\u4e3aTransformer\u6a21\u578b\u63d0\u4f9b\u4e86\u4e30\u5bcc\u7684\u4fe1\u606f\u6765\u5904\u7406\u81ea\u7136\u8bed\u8a00\u4efb\u52a1\u3002<\/p>\n\n\n\n<p>\u8fd9\u91cc\u7684768\u7ef4\u5bf9\u5e94\u4e8eGPT-2\u6a21\u578b\u7684\u9690\u85cf\u5c42\u5927\u5c0f\uff0c\u800c5\u5219\u662f\u4f60\u7684\u793a\u4f8b\u6587\u672c\u5728\u7ecf\u8fc7\u5206\u8bcd\u5668\u5904\u7406\u540e\u5f97\u5230\u7684token\u6570\u91cf\u3002\u8fd9\u4e2a\u7ed3\u679c\u6b63\u662f\u6211\u4eec\u9884\u671f\u7684\uff0c\u5c55\u793a\u4e86\u5982\u4f55\u624b\u52a8\u7ec4\u5408\u8bcd\u5d4c\u5165\u548c\u4f4d\u7f6e\u7f16\u7801\u6765\u51c6\u5907\u8f93\u5165\u7ed9Transformer\u6a21\u578b\u7684\u6570\u636e\u3002\u8fd9\u79cd\u8868\u793a\u4f7f\u5f97\u6a21\u578b\u80fd\u591f\u57fa\u4e8e\u6bcf\u4e2atoken\u7684\u5185\u5bb9\u53ca\u5176\u5728\u5e8f\u5217\u4e2d\u7684\u4f4d\u7f6e\u6765\u7406\u89e3\u548c\u751f\u6210\u6587\u672c\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"<p>10. \u4f4d\u7f6e\u7f16\u7801 transformer.wpe.weight\u4ee3\u8868\u7684\u662f\u4f4d\u7f6e\u7f16\u7801\uff08Positional Enco [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":0,"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":"set","ast-page-background-enabled":"default","ast-page-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":"","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-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":"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":""},"mobile":{"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":""}},"_jetpack_memberships_contains_paid_content":false,"footnotes":""},"categories":[313,289,443,442,312],"tags":[242,314],"class_list":["post-2384","post","type-post","status-publish","format-standard","hentry","category-chatgpt","category-gpt","category-llm","category-llms","category-openai","tag-chatgpt","tag-openai-api"],"views":2146,"jetpack_sharing_enabled":true,"jetpack_featured_media_url":"","_links":{"self":[{"href":"https:\/\/www.aqwu.net\/wp\/index.php?rest_route=\/wp\/v2\/posts\/2384","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.aqwu.net\/wp\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.aqwu.net\/wp\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.aqwu.net\/wp\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.aqwu.net\/wp\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=2384"}],"version-history":[{"count":7,"href":"https:\/\/www.aqwu.net\/wp\/index.php?rest_route=\/wp\/v2\/posts\/2384\/revisions"}],"predecessor-version":[{"id":2392,"href":"https:\/\/www.aqwu.net\/wp\/index.php?rest_route=\/wp\/v2\/posts\/2384\/revisions\/2392"}],"wp:attachment":[{"href":"https:\/\/www.aqwu.net\/wp\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=2384"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.aqwu.net\/wp\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=2384"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.aqwu.net\/wp\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=2384"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}