{"id":2366,"date":"2024-03-10T20:42:12","date_gmt":"2024-03-10T12:42:12","guid":{"rendered":"https:\/\/www.aqwu.net\/wp\/?p=2366"},"modified":"2024-04-28T20:03:34","modified_gmt":"2024-04-28T12:03:34","slug":"%e4%ba%86%e8%a7%a3-openai-%e7%9a%84-gpt2-%e6%a8%a1%e5%9e%8b%e7%bb%93%e6%9e%842","status":"publish","type":"post","link":"https:\/\/www.aqwu.net\/wp\/?p=2366","title":{"rendered":"\u4e86\u89e3 OpenAI \u7684 gpt2 \u6a21\u578b\u7ed3\u6784(2)"},"content":{"rendered":"\n<h2 class=\"wp-block-heading\"><strong>5. \u5982\u4f55\u6253\u5370\u6a21\u578b\u53c2\u6570\uff1f<\/strong><\/h2>\n\n\n\n<p>\u8981\u6253\u5370\u4e00\u4e2a\u9884\u8bad\u7ec3\u6a21\u578b\u7684\u53c2\u6570\uff0c\u4f60\u53ef\u4ee5\u4f7f\u7528<code>transformers<\/code>\u5e93\u4e2d\u6a21\u578b\u7684<code>.named_parameters()<\/code>\u65b9\u6cd5\u3002\u8fd9\u4e2a\u65b9\u6cd5\u4f1a\u904d\u5386\u6a21\u578b\u7684\u6240\u6709\u53c2\u6570\uff0c\u5e76\u8fd4\u56de\u4e00\u4e2a\u8fed\u4ee3\u5668\uff0c\u5176\u4e2d\u5305\u542b\u6bcf\u4e2a\u53c2\u6570\u7684\u540d\u79f0\u548c\u53c2\u6570\u672c\u8eab\u3002\u4e0b\u9762\u662f\u4e00\u4e2a\u7b80\u5316\u7684\u793a\u4f8b\uff0c\u5c55\u793a\u4e86\u5982\u4f55\u52a0\u8f7dGPT-2\u6a21\u578b\u5e76\u6253\u5370\u5176\u53c2\u6570\u540d\u79f0\u548c\u53c2\u6570\u5927\u5c0f\u3002\u547d\u540d\u4e3a test03.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\u6a21\u578b\nmodel = GPT2LMHeadModel.from_pretrained('gpt2')\n\n# \u6253\u5370\u6a21\u578b\u7684\u53c2\u6570\u540d\u79f0\u548c\u5927\u5c0f\nfor name, param in model.named_parameters():\n    print(f\"Parameter Name: {name} \\t Size: {param.size()}\")\n<\/pre><\/div>\n\n\n\n<p>\u8fd9\u6bb5\u4ee3\u7801\u9996\u5148\u4ece<code>transformers<\/code>\u5e93\u4e2d\u5bfc\u5165\u4e86<code>GPT2LMHeadModel<\/code>\u7c7b\uff0c\u7136\u540e\u4f7f\u7528<code>from_pretrained<\/code>\u65b9\u6cd5\u52a0\u8f7d\u4e86\u9884\u8bad\u7ec3\u7684GPT-2\u6a21\u578b\u3002\u63a5\u4e0b\u6765\uff0c\u5b83\u904d\u5386\u6a21\u578b\u7684\u6240\u6709\u53c2\u6570\uff0c\u6253\u5370\u51fa\u6bcf\u4e2a\u53c2\u6570\u7684\u540d\u79f0\u548c\u5927\u5c0f\u3002<\/p>\n\n\n\n<p>\u8fd9\u79cd\u65b9\u5f0f\u5bf9\u4e8e\u7406\u89e3\u6a21\u578b\u7684\u6784\u9020\u548c\u8c03\u8bd5\u975e\u5e38\u6709\u7528\uff0c\u7279\u522b\u662f\u5f53\u4f60\u9700\u8981\u5bf9\u6a21\u578b\u8fdb\u884c\u5fae\u8c03\u6216\u8005\u7406\u89e3\u5176\u5185\u90e8\u5de5\u4f5c\u539f\u7406\u65f6\u3002\u8bf7\u6ce8\u610f\uff0c\u7531\u4e8eGPT-2\u6a21\u578b\uff08\u4ee5\u53ca\u5176\u4ed6\u5927\u578b\u6a21\u578b\uff09\u5305\u542b\u5927\u91cf\u7684\u53c2\u6570\uff0c\u8f93\u51fa\u4f1a\u975e\u5e38\u957f\u3002\u4f60\u53ef\u80fd\u53ea\u60f3\u67e5\u770b\u7279\u5b9a\u90e8\u5206\u7684\u53c2\u6570\uff0c\u6216\u8005\u6839\u636e\u53c2\u6570\u540d\u79f0\u8fdb\u884c\u7b5b\u9009\u3002<\/p>\n\n\n\n<p>\u8fd0\u884ctest03.py<\/p>\n\n\n\n<div class=\"wp-block-urvanov-syntax-highlighter-code-block\"><pre class=\"lang:python decode:true \">python newsrc\/test03.py\nParameter Name: transformer.wte.weight   Size: torch.Size([50257, 768])\nParameter Name: transformer.wpe.weight   Size: torch.Size([1024, 768])\nParameter Name: transformer.h.0.ln_1.weight      Size: torch.Size([768])\nParameter Name: transformer.h.0.ln_1.bias        Size: torch.Size([768])\nParameter Name: transformer.h.0.attn.c_attn.weight       Size: torch.Size([768, 2304])\nParameter Name: transformer.h.0.attn.c_attn.bias         Size: torch.Size([2304])\nParameter Name: transformer.h.0.attn.c_proj.weight       Size: torch.Size([768, 768])\nParameter Name: transformer.h.0.attn.c_proj.bias         Size: torch.Size([768])\nParameter Name: transformer.h.0.ln_2.weight      Size: torch.Size([768])\nParameter Name: transformer.h.0.ln_2.bias        Size: torch.Size([768])\nParameter Name: transformer.h.0.mlp.c_fc.weight          Size: torch.Size([768, 3072])\nParameter Name: transformer.h.0.mlp.c_fc.bias    Size: torch.Size([3072])\nParameter Name: transformer.h.0.mlp.c_proj.weight        Size: torch.Size([3072, 768])\nParameter Name: transformer.h.0.mlp.c_proj.bias          Size: torch.Size([768])\nParameter Name: transformer.h.1.ln_1.weight      Size: torch.Size([768])\nParameter Name: transformer.h.1.ln_1.bias        Size: torch.Size([768])\nParameter Name: transformer.h.1.attn.c_attn.weight       Size: torch.Size([768, 2304])\nParameter Name: transformer.h.1.attn.c_attn.bias         Size: torch.Size([2304])\nParameter Name: transformer.h.1.attn.c_proj.weight       Size: torch.Size([768, 768])\nParameter Name: transformer.h.1.attn.c_proj.bias         Size: torch.Size([768])\nParameter Name: transformer.h.1.ln_2.weight      Size: torch.Size([768])\nParameter Name: transformer.h.1.ln_2.bias        Size: torch.Size([768])\nParameter Name: transformer.h.1.mlp.c_fc.weight          Size: torch.Size([768, 3072])\nParameter Name: transformer.h.1.mlp.c_fc.bias    Size: torch.Size([3072])\nParameter Name: transformer.h.1.mlp.c_proj.weight        Size: torch.Size([3072, 768])\nParameter Name: transformer.h.1.mlp.c_proj.bias          Size: torch.Size([768])\n...\nParameter Name: transformer.h.11.ln_1.weight     Size: torch.Size([768])\nParameter Name: transformer.h.11.ln_1.bias       Size: torch.Size([768])\nParameter Name: transformer.h.11.attn.c_attn.weight      Size: torch.Size([768, 2304])\nParameter Name: transformer.h.11.attn.c_attn.bias        Size: torch.Size([2304])\nParameter Name: transformer.h.11.attn.c_proj.weight      Size: torch.Size([768, 768])\nParameter Name: transformer.h.11.attn.c_proj.bias        Size: torch.Size([768])\nParameter Name: transformer.h.11.ln_2.weight     Size: torch.Size([768])\nParameter Name: transformer.h.11.ln_2.bias       Size: torch.Size([768])\nParameter Name: transformer.h.11.mlp.c_fc.weight         Size: torch.Size([768, 3072])\nParameter Name: transformer.h.11.mlp.c_fc.bias   Size: torch.Size([3072])\nParameter Name: transformer.h.11.mlp.c_proj.weight       Size: torch.Size([3072, 768])\nParameter Name: transformer.h.11.mlp.c_proj.bias         Size: torch.Size([768])\nParameter Name: transformer.ln_f.weight          Size: torch.Size([768])\nParameter Name: transformer.ln_f.bias    Size: torch.Size([768])\n<\/pre><\/div>\n\n\n\n<h2 class=\"wp-block-heading\"><strong>6. \u8bcd\u5d4c\u5165\u6743\u91cd<\/strong><\/h2>\n\n\n\n<p>\u4e3a\u4e86\u8f93\u51fa\u7279\u5b9a\u6a21\u578b\u53c2\u6570\u7684\u5185\u5bb9\uff0c\u4f8b\u5982<code>transformer.wte.weight<\/code>\uff08\u8fd9\u662fGPT-2\u6a21\u578b\u4e2d\u7684\u8bcd\u5d4c\u5165\u6743\u91cd\uff09\uff0c\u4f60\u53ef\u4ee5\u4f7f\u7528\u4ee5\u4e0b\u4ee3\u7801\u7247\u6bb5\u3002\u8bf7\u6ce8\u610f\uff0c\u7531\u4e8e\u6a21\u578b\u53c2\u6570\u53ef\u80fd\u975e\u5e38\u5927\uff0c\u76f4\u63a5\u6253\u5370\u51fa\u6574\u4e2a\u53c2\u6570\u7684\u5185\u5bb9\u53ef\u80fd\u4e0d\u662f\u4e00\u4e2a\u597d\u4e3b\u610f\uff0c\u56e0\u4e3a\u8fd9\u53ef\u80fd\u5bfc\u81f4\u5de8\u5927\u7684\u8f93\u51fa\uff0c\u96be\u4ee5\u5728\u7ec8\u7aef\u6216\u7b14\u8bb0\u672c\u73af\u5883\u4e2d\u67e5\u770b\u548c\u7ba1\u7406\u3002\u56e0\u6b64\uff0c\u6211\u4f1a\u5c55\u793a\u5982\u4f55\u83b7\u53d6\u5e76\u6253\u5370\u8fd9\u4e9b\u6743\u91cd\u7684\u4e00\u5c0f\u90e8\u5206\uff08\u4f8b\u5982\uff0c\u524d\u51e0\u4e2a\u6743\u91cd\uff09\u6765\u793a\u8303\u5982\u4f55\u8bbf\u95ee\u8fd9\u4e9b\u53c2\u6570\u3002<\/p>\n\n\n\n<p>\u547d\u540d\u4e3a test04.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\n\n# \u52a0\u8f7d\u9884\u8bad\u7ec3\u6a21\u578b\nmodel = GPT2LMHeadModel.from_pretrained('gpt2')\n\n# \u76f4\u63a5\u83b7\u53d6\u8bcd\u5d4c\u5165\u77e9\u9635\uff1atransformer.wte.weight\nwte_weights = model.transformer.wte.weight\n\nprint(\"\u76f4\u63a5\u6253\u5370\u524d5\u4e2a\u5d4c\u5165\u5411\u91cf:\")\nprint(wte_weights[:5])\n\n# \u53e6\u4e00\u79cd\u65b9\u6cd5\u83b7\u53d6\u8bcd\u5d4c\u5165\u77e9\u9635\u7684\u6743\u91cd\uff0c\u6700\u540e\u8f93\u51fa\u7ed3\u679c\u4e00\u6837\nembedding_weight = model.get_input_embeddings().weight.data\n\n# \u6253\u5370\u524d5\u4e2a\u5d4c\u5165\u5411\u91cf\nprint(\"\u524d5\u4e2a\u5d4c\u5165\u5411\u91cf:\")\nprint(embedding_weight[:5])\n<\/pre><\/div>\n\n\n\n<p>\u8fd9\u6bb5\u4ee3\u7801\u9996\u5148\u52a0\u8f7d\u4e86\u9884\u8bad\u7ec3\u7684GPT-2\u6a21\u578b\u3002\u7136\u540e\uff0c\u5b83\u8bbf\u95ee<code>transformer.wte.weight<\/code>\u53c2\u6570\uff0c\u8fd9\u662f\u6a21\u578b\u4e2d\u8bcd\u5d4c\u5165\u5c42\u7684\u6743\u91cd\u3002\u6700\u540e\uff0c\u4ee3\u7801\u53ea\u6253\u5370\u51fa\u8fd9\u4e9b\u6743\u91cd\u7684\u4e00\u5c0f\u90e8\u5206\uff0c\u4ee5\u4fbf\u4e8e\u67e5\u770b\u548c\u5206\u6790\u3002<\/p>\n\n\n\n<p>\u8bf7\u786e\u4fdd\u5728\u8fd0\u884c\u8fd9\u6bb5\u4ee3\u7801\u524d\u5df2\u7ecf\u5b89\u88c5\u4e86<code>transformers<\/code>\u5e93\u3002\u5982\u679c\u4f60\u60f3\u8981\u67e5\u770b\u66f4\u591a\u6216\u4e0d\u540c\u7684\u6743\u91cd\u90e8\u5206\uff0c\u53ef\u4ee5\u901a\u8fc7\u8c03\u6574\u5207\u7247\u64cd\u4f5c\uff08\u4f8b\u5982\uff0c<code>[:5]<\/code>\uff09\u6765\u5b9e\u73b0\u3002\u8fd9\u4e2a\u4f8b\u5b50\u4f7f\u7528\u4e86\u5207\u7247\u6765\u83b7\u53d6\u524d5\u4e2a\u6743\u91cd\uff0c\u4f46\u4f60\u53ef\u4ee5\u6839\u636e\u9700\u8981\u8c03\u6574\u8fd9\u4e2a\u8303\u56f4\u3002<\/p>\n\n\n\n<p>\u8fd0\u884ctest04.py<\/p>\n\n\n\n<div class=\"wp-block-urvanov-syntax-highlighter-code-block\"><pre class=\"lang:sh decode:true \">python newsrc\/test04.py\n\u76f4\u63a5\u6253\u5370\u524d5\u4e2a\u5d4c\u5165\u5411\u91cf:\ntensor([[-0.1101, -0.0393,  0.0331,  ..., -0.1364,  0.0151,  0.0453],\n        [ 0.0403, -0.0486,  0.0462,  ...,  0.0861,  0.0025,  0.0432],\n        [-0.1275,  0.0479,  0.1841,  ...,  0.0899, -0.1297, -0.0879],\n        [-0.0927, -0.3053,  0.2112,  ...,  0.1769,  0.0364, -0.1797],\n        [-0.0506, -0.1111,  0.1058,  ..., -0.1149,  0.0664,  0.0574]],\n       grad_fn=&lt;SliceBackward0&gt;)\n\u524d5\u4e2a\u5d4c\u5165\u5411\u91cf:\ntensor([[-0.1101, -0.0393,  0.0331,  ..., -0.1364,  0.0151,  0.0453],\n        [ 0.0403, -0.0486,  0.0462,  ...,  0.0861,  0.0025,  0.0432],\n        [-0.1275,  0.0479,  0.1841,  ...,  0.0899, -0.1297, -0.0879],\n        [-0.0927, -0.3053,  0.2112,  ...,  0.1769,  0.0364, -0.1797],\n        [-0.0506, -0.1111,  0.1058,  ..., -0.1149,  0.0664,  0.0574]])<\/pre><\/div>\n\n\n\n<p>\u8fd9\u4e9b\u5d4c\u5165\u5411\u91cf\u8868\u793a\u7684\u662f\u6a21\u578b\u8bcd\u6c47\u8868\u4e2d\u524d\u4e94\u4e2atoken\u7684\u5d4c\u5165\u3002\u5728GPT-2\u8fd9\u6837\u7684\u6a21\u578b\u4e2d\uff0c\u6bcf\u4e2a\u5d4c\u5165\u5411\u91cf\u662f\u4e00\u4e2a\u9ad8\u7ef4\u7a7a\u95f4\u4e2d\u7684\u70b9\uff0c\u4ee3\u8868\u4e86\u76f8\u5e94token\u7684\u8bed\u4e49\u548c\u8bed\u6cd5\u5c5e\u6027\u3002\u8fd9\u4e9b\u5411\u91cf\u662f\u6a21\u578b\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u5b66\u4e60\u5230\u7684\uff0c\u4f7f\u5f97\u6a21\u578b\u80fd\u591f\u57fa\u4e8e\u8fd9\u4e9b\u5d4c\u5165\u5411\u91cf\u6765\u7406\u89e3\u548c\u751f\u6210\u6587\u672c\u3002<\/p>\n\n\n\n<p>\u5d4c\u5165\u5411\u91cf\u7684\u6bcf\u4e2a\u7ef4\u5ea6\u90fd\u6355\u6349\u4e86\u4e00\u4e9b\u5173\u4e8e\u8bcd\u6c47\u7684\u4fe1\u606f\uff0c\u4f46\u8fd9\u4e9b\u4fe1\u606f\u5e76\u4e0d\u662f\u76f4\u89c2\u53ef\u89e3\u91ca\u7684\u3002\u6df1\u5ea6\u5b66\u4e60\u6a21\u578b\u901a\u8fc7\u5728\u8fd9\u4e9b\u9ad8\u7ef4\u7a7a\u95f4\u4e2d\u64cd\u4f5c\u5d4c\u5165\u5411\u91cf\u6765\u5b8c\u6210\u4efb\u52a1\uff0c\u5982\u5206\u7c7b\u3001\u751f\u6210\u6587\u672c\u7b49\u3002<\/p>\n\n\n\n<p><strong>\u95ee\uff1a\u6839\u636e transformer.wte.weight Size: torch.Size([50257, 768]\uff0c\u5e94\u8be5\u670950257\u4e2a\u5d4c\u5165\u91cf\uff0c\u6bcf\u4e00\u884c\u6709768\u4e2a\u503c\uff1f<\/strong><\/p>\n\n\n\n<p>\u662f\u7684\uff0c\u4f60\u7684\u7406\u89e3\u662f\u6b63\u786e\u7684\u3002\u5728<code>transformer.wte.weight<\/code>\u4e2d\uff0c\u5c3a\u5bf8\u4e3a<code>torch.Size([50257, 768])<\/code>\u8868\u793a\u8fd9\u4e2a\u5d4c\u5165\u77e9\u9635\u670950257\u4e2a\u5d4c\u5165\u5411\u91cf\uff0c\u6bcf\u4e2a\u5d4c\u5165\u5411\u91cf\u5305\u542b768\u4e2a\u503c\u3002\u8fd9\u91cc\u768450257\u4ee3\u8868\u6a21\u578b\u8bcd\u6c47\u8868\u7684\u5927\u5c0f\uff0c\u5373\u6a21\u578b\u53ef\u4ee5\u8bc6\u522b\u7684\u4e0d\u540ctoken\u7684\u603b\u6570\uff1b768\u5219\u662f\u5d4c\u5165\u7a7a\u95f4\u7684\u7ef4\u5ea6\uff0c\u5373\u6bcf\u4e2atoken\u88ab\u6620\u5c04\u5230\u4e00\u4e2a768\u7ef4\u7684\u5411\u91cf\u7a7a\u95f4\u4e2d\u3002<\/p>\n\n\n\n<p>\u5728\u81ea\u7136\u8bed\u8a00\u5904\u7406\u6a21\u578b\u4e2d\uff0c\u8bcd\u5d4c\u5165\u662f\u5c06\u8bcd\u6c47\u8868\u4e2d\u7684\u8bcd\uff08\u6216token\uff09\u8f6c\u6362\u4e3a\u56fa\u5b9a\u957f\u5ea6\u7684\u5411\u91cf\u7684\u6280\u672f\u3002\u8fd9\u4e9b\u5411\u91cf\u6355\u6349\u4e86\u8bcd\u7684\u8bed\u4e49\u4fe1\u606f\u548c\u8bed\u5883\u4fe1\u606f\uff0c\u4f7f\u5f97\u6a21\u578b\u80fd\u591f\u5904\u7406\u81ea\u7136\u8bed\u8a00\u3002\u5bf9\u4e8e\u7ed9\u5b9a\u7684\u8bcd\uff08\u6216token\uff09\uff0c\u5176\u5d4c\u5165\u5411\u91cf\u901a\u8fc7\u5728\u5d4c\u5165\u77e9\u9635\u4e2d\u67e5\u627e\u76f8\u5e94\u7684\u884c\u6765\u83b7\u53d6\u3002\u4f8b\u5982\uff0c\u5982\u679c\u4f60\u60f3\u8981\u83b7\u53d6\u8bcd\u6c47\u8868\u4e2d\u7b2c\u4e00\u4e2atoken\u7684\u5d4c\u5165\u5411\u91cf\uff0c\u4f60\u53ef\u4ee5\u7b80\u5355\u5730\u67e5\u770b\u5d4c\u5165\u77e9\u9635\u7684\u7b2c\u4e00\u884c\u3002<\/p>\n\n\n\n<p>\u5728GPT-2\u8fd9\u6837\u7684\u6a21\u578b\u4e2d\uff0c\u8fd9\u4e9b\u5d4c\u5165\u5411\u91cf\u662f\u6a21\u578b\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u5b66\u4e60\u5230\u7684\uff0c\u5e76\u4e14\u5b83\u4eec\u662f\u6a21\u578b\u7406\u89e3\u548c\u751f\u6210\u6587\u672c\u7684\u57fa\u7840\u3002\u6bcf\u4e2a\u5d4c\u5165\u5411\u91cf\u901a\u8fc7\u5728\u9ad8\u7ef4\u7a7a\u95f4\u4e2d\u7684\u4f4d\u7f6e\u6765\u8868\u793a\u5176\u5bf9\u5e94\u7684token\u7684\u7279\u5f81\uff0c\u5305\u62ec\u8bed\u4e49\u548c\u8bed\u6cd5\u5c5e\u6027\u3002\u8fd9\u79cd\u8868\u793a\u65b9\u6cd5\u5141\u8bb8\u6a21\u578b\u6355\u6349\u8bcd\u4e0e\u8bcd\u4e4b\u95f4\u7684\u590d\u6742\u5173\u7cfb\uff0c\u5982\u540c\u4e49\u8bcd\u3001\u53cd\u4e49\u8bcd\u4ee5\u53ca\u8bcd\u5728\u4e0d\u540c\u8bed\u5883\u4e2d\u7684\u4f7f\u7528\u7b49\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\"><strong>7. \u524d5\u4e2a\u8bcd\u5982\u4f55\u6253\u5370\uff1f<\/strong><\/h2>\n\n\n\n<p><br>\u5728Transformer\u6a21\u578b\uff08\u5982GPT-2\uff09\u4e2d\uff0c<code>transformer.wte.weight<\/code>\u662f\u8bcd\u5d4c\u5165\u77e9\u9635\u7684\u6743\u91cd\uff0c\u5b83\u5c06\u8bcd\u6c47\u8868\u4e2d\u7684\u6bcf\u4e2a\u5355\u8bcd\uff08\u6216\u66f4\u51c6\u786e\u5730\u8bf4\uff0c\u6bcf\u4e2atoken\uff09\u6620\u5c04\u5230\u4e00\u4e2a\u9ad8\u7ef4\u7a7a\u95f4\u3002\u6bcf\u4e00\u884c\u5bf9\u5e94\u4e8e\u8bcd\u6c47\u8868\u4e2d\u7684\u4e00\u4e2atoken\u7684\u5d4c\u5165\u5411\u91cf\u3002\u56e0\u6b64\uff0c\u5982\u679c\u4f60\u60f3\u67e5\u770b\u524d5\u4e2a\u8bcd\u7684\u5d4c\u5165\u5411\u91cf\uff0c\u4f60\u53ef\u4ee5\u76f4\u63a5\u7d22\u5f15\u8fd9\u4e2a\u77e9\u9635\u7684\u524d5\u884c\u3002\u540c\u6837\uff0c\u5982\u679c\u4f60\u60f3\u77e5\u9053\u8fd9\u4e9b\u5d4c\u5165\u5411\u91cf\u5bf9\u5e94\u7684\u8bcd\u662f\u4ec0\u4e48\uff0c\u4f60\u53ef\u4ee5\u4f7f\u7528\u5206\u8bcd\u5668\u7684<code>convert_ids_to_tokens<\/code>\u65b9\u6cd5\u3002<\/p>\n\n\n\n<p>\u4e0b\u9762\u662f\u4e00\u4e2a\u5982\u4f55\u5b9e\u73b0\u8fd9\u4e00\u70b9\u7684\u4f8b\u5b50,\u547d\u540d\u4e3a test05.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 GPT2Tokenizer, GPT2Model\n\n# \u52a0\u8f7d\u9884\u8bad\u7ec3\u7684GPT-2\u6a21\u578b\u548c\u5206\u8bcd\u5668\nmodel = GPT2Model.from_pretrained('gpt2')\ntokenizer = GPT2Tokenizer.from_pretrained('gpt2')\n\n# \u83b7\u53d6\u8bcd\u5d4c\u5165\u77e9\u9635\u7684\u6743\u91cd\nembedding_weight = model.get_input_embeddings().weight.data\n\n# \u6253\u5370\u524d5\u4e2a\u5d4c\u5165\u5411\u91cf\nprint(\"\u524d5\u4e2a\u5d4c\u5165\u5411\u91cf:\")\nprint(embedding_weight[:5])\n\n# \u83b7\u53d6\u524d5\u4e2atoken\u7684ID\nfirst_five_tokens_ids = list(range(5))\n\n# \u5c06\u8fd9\u4e9bID\u8f6c\u6362\u4e3atoken\nfirst_five_tokens = tokenizer.convert_ids_to_tokens(first_five_tokens_ids)\n\n# \u6253\u5370\u524d5\u4e2a\u8bcd\nprint(\"\\n\u5bf9\u5e94\u7684\u524d5\u4e2a\u8bcd:\")\nprint(first_five_tokens)\n<\/pre><\/div>\n\n\n\n<p>\u8fd9\u6bb5\u4ee3\u7801\u9996\u5148\u52a0\u8f7d\u4e86GPT-2\u6a21\u578b\u548c\u5bf9\u5e94\u7684\u5206\u8bcd\u5668\u3002\u7136\u540e\uff0c\u5b83\u83b7\u53d6\u8bcd\u5d4c\u5165\u77e9\u9635\u7684\u6743\u91cd\uff0c\u5e76\u6253\u5370\u51fa\u77e9\u9635\u7684\u524d5\u884c\uff0c\u8fd9\u4e9b\u884c\u5bf9\u5e94\u4e8e\u8bcd\u6c47\u8868\u4e2d\u7684\u524d5\u4e2atoken\u7684\u5d4c\u5165\u5411\u91cf\u3002\u4e4b\u540e\uff0c\u5b83\u521b\u5efa\u4e86\u4e00\u4e2a\u5305\u542b\u524d5\u4e2atoken ID\u7684\u5217\u8868\uff0c\u5e76\u4f7f\u7528\u5206\u8bcd\u5668\u5c06\u8fd9\u4e9bID\u8f6c\u6362\u6210\u5b9e\u9645\u7684token\uff08\u8bcd\uff09\u3002\u6700\u540e\uff0c\u5b83\u6253\u5370\u51fa\u8fd9\u4e9btoken\u3002<\/p>\n\n\n\n<p>\u8bf7\u6ce8\u610f\uff0c\u8bcd\u6c47\u8868\u7684\u5f00\u59cb\u901a\u5e38\u5305\u542b\u7279\u6b8atoken\uff08\u5982[CLS]\u3001[SEP]\u3001[PAD]\u7b49\uff09\uff0c\u56e0\u6b64\u524d5\u4e2atoken\u53ef\u80fd\u4e0d\u662f\u5e38\u89c4\u7684\u201c\u8bcd\u201d\u3002\u8981\u67e5\u770b\u5b9e\u9645\u7684\u5355\u8bcd\u6216\u77ed\u8bed\uff0c\u4f60\u53ef\u80fd\u9700\u8981\u9009\u62e9\u4e00\u4e2a\u4e0d\u540c\u7684\u7d22\u5f15\u8303\u56f4\uff0c\u5177\u4f53\u53d6\u51b3\u4e8e\u5206\u8bcd\u5668\u7684\u8bcd\u6c47\u8868\u7ed3\u6784\u3002<\/p>\n\n\n\n<p>\u8fd0\u884ctest05.py<\/p>\n\n\n\n<div class=\"wp-block-urvanov-syntax-highlighter-code-block\"><pre class=\"lang:sh decode:true \">python newsrc\/test05.py\n\u524d5\u4e2a\u5d4c\u5165\u5411\u91cf:\ntensor([[-0.1101, -0.0393,  0.0331,  ..., -0.1364,  0.0151,  0.0453],\n        [ 0.0403, -0.0486,  0.0462,  ...,  0.0861,  0.0025,  0.0432],\n        [-0.1275,  0.0479,  0.1841,  ...,  0.0899, -0.1297, -0.0879],\n        [-0.0927, -0.3053,  0.2112,  ...,  0.1769,  0.0364, -0.1797],\n        [-0.0506, -0.1111,  0.1058,  ..., -0.1149,  0.0664,  0.0574]])\n\n\u5bf9\u5e94\u7684\u524d5\u4e2a\u8bcd:\n['!', '\"', '#', '$', '%']<\/pre><\/div>\n\n\n\n<h2 class=\"wp-block-heading\"><strong>8. \u5bf9 prompt \u8fdb\u884c\u7f16\u7801<\/strong><\/h2>\n\n\n\n<p>\u8981\u5bf9\u7ed9\u5b9a\u7684\u63d0\u793a\u201c<code>The meaning of life is<\/code>\u201d\u4f7f\u7528GPT-2\u7684\u5206\u8bcd\u5668\u8fdb\u884c\u7f16\u7801\uff0c\u4f60\u53ef\u4ee5\u6309\u7167\u4ee5\u4e0b\u6b65\u9aa4\u8fdb\u884c,\u547d\u540d\u4e3a test06.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 GPT2Tokenizer\n\n# \u52a0\u8f7d\u9884\u8bad\u7ec3\u7684GPT-2\u5206\u8bcd\u5668\ntokenizer = GPT2Tokenizer.from_pretrained('gpt2')\n\n# \u5b9a\u4e49prompt\nprompt = \"The meaning of life is\"\n\n# \u5bf9prompt\u8fdb\u884c\u7f16\u7801\uff0c\u8fd4\u56dePython\u5217\u8868\nencoded_prompt_list = tokenizer.encode(prompt, add_special_tokens=False)\n\n# \u5bf9prompt\u8fdb\u884c\u7f16\u7801\uff0c\u8fd4\u56dePyTorch Tensor\nencoded_prompt_tensor = tokenizer.encode(prompt, return_tensors='pt', add_special_tokens=False)\n\nprint(\"List:\", encoded_prompt_list)\nprint(\"Tensor:\", encoded_prompt_tensor)\n\n# \u5c06\u8fd9\u4e9bID\u8f6c\u6362\u4e3atoken\ntokens = tokenizer.convert_ids_to_tokens(encoded_prompt_list)\nprint(\"tokens:\", tokens)\n<\/pre><\/div>\n\n\n\n<p>\u8fd9\u6bb5\u4ee3\u7801\u4f1a\u4f7f\u7528GPT-2\u5206\u8bcd\u5668\u5c06\u6587\u672c\u8f6c\u6362\u4e3a\u6a21\u578b\u80fd\u591f\u7406\u89e3\u7684\u4e00\u7cfb\u5217\u6570\u5b57\uff08token IDs\uff09\u3002<code>add_special_tokens=False<\/code>\u53c2\u6570\u544a\u8bc9\u5206\u8bcd\u5668\u4e0d\u8981\u6dfb\u52a0\u4efb\u4f55\u7279\u6b8a\u7684token\uff08\u5982\u5f00\u59cb\u6216\u7ed3\u675ftoken\uff09\uff0c\u53ea\u5bf9\u5b9e\u9645\u63d0\u4f9b\u7684\u6587\u672c\u8fdb\u884c\u7f16\u7801\u3002<\/p>\n\n\n\n<p>\u5982\u679c\u4f60\u5728\u81ea\u5df1\u7684\u73af\u5883\u4e2d\u8fd0\u884c\u8fd9\u6bb5\u4ee3\u7801\uff0c\u5b83\u5c06\u8f93\u51fa<code>prompt<\/code>\u5b57\u7b26\u4e32\u7684\u7f16\u7801\u7ed3\u679c\uff0c\u8fd9\u662f\u4e00\u4e2a\u6574\u6570\u5217\u8868\uff0c\u6bcf\u4e2a\u6574\u6570\u4ee3\u8868\u8bcd\u6c47\u8868\u4e2d\u7684\u4e00\u4e2a\u7279\u5b9atoken\u3002<\/p>\n\n\n\n<p>\u5f53\u4f60\u5728\u8c03\u7528<code>tokenizer.encode()<\/code>\u65b9\u6cd5\u65f6\u6dfb\u52a0<code>return_tensors='pt'<\/code>\u53c2\u6570\uff0c\u610f\u5473\u7740\u4f60\u5e0c\u671b\u7f16\u7801\u540e\u7684\u7ed3\u679c\u88ab\u5c01\u88c5\u6210\u4e00\u4e2aPyTorch\u7684<code>Tensor<\/code>\uff0c\u800c\u4e0d\u662f\u8fd4\u56de\u4e00\u4e2a\u7b80\u5355\u7684Python\u5217\u8868\u3002\u8fd9\u5bf9\u4e8e\u5c06\u7f16\u7801\u540e\u7684\u6570\u636e\u76f4\u63a5\u7528\u4e8ePyTorch\u6a21\u578b\u662f\u5fc5\u8981\u7684\uff0c\u56e0\u4e3aPyTorch\u6a21\u578b\u671f\u671b\u7684\u8f93\u5165\u662f<code>Tensor<\/code>\u683c\u5f0f\u3002<\/p>\n\n\n\n<p>\u4ee5\u4e0b\u662f\u4e24\u79cd\u8c03\u7528\u65b9\u5f0f\u7684\u5bf9\u6bd4\uff1a<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li><strong>\u4e0d\u4f7f\u7528<code>return_tensors='pt'<\/code><\/strong>\uff1a\n<ul class=\"wp-block-list\">\n<li>\u8fd4\u56de\u503c\u662f\u4e00\u4e2aPython\u5217\u8868\uff0c\u5305\u542b\u4e86\u7f16\u7801\u540e\u7684token ID\u3002<\/li>\n\n\n\n<li>\u8fd9\u79cd\u65b9\u5f0f\u5728\u4f60\u9700\u8981\u5bf9\u7f16\u7801\u540e\u7684\u7ed3\u679c\u8fdb\u884c\u8fdb\u4e00\u6b65\u7684Python\u7ea7\u522b\u5904\u7406\u65f6\u5f88\u6709\u7528\u3002<\/li>\n<\/ul>\n<\/li>\n\n\n\n<li><strong>\u4f7f\u7528<code>return_tensors='pt'<\/code><\/strong>\uff1a\n<ul class=\"wp-block-list\">\n<li>\u8fd4\u56de\u503c\u662f\u4e00\u4e2aPyTorch <code>Tensor<\/code>\uff0c\u8be5<code>Tensor<\/code>\u5c01\u88c5\u4e86\u7f16\u7801\u540e\u7684token ID\u3002<\/li>\n\n\n\n<li>\u8fd9\u79cd\u65b9\u5f0f\u76f4\u63a5\u4e3a\u6a21\u578b\u8bad\u7ec3\u6216\u63a8\u7406\u51c6\u5907\u4e86\u8f93\u5165\u6570\u636e\uff0c\u65e0\u9700\u989d\u5916\u8f6c\u6362\u6b65\u9aa4\u3002<\/li>\n\n\n\n<li><code>return_tensors='pt'<\/code>\u53c2\u6570\u786e\u4fdd\u4e86\u7f16\u7801\u540e\u7684\u6570\u636e\u53ef\u4ee5\u76f4\u63a5\u4e0ePyTorch\u6846\u67b6\u517c\u5bb9\uff0c\u4fbf\u4e8e\u8fdb\u884c\u6a21\u578b\u7684\u524d\u5411\u4f20\u64ad\u3002<\/li>\n<\/ul>\n<\/li>\n<\/ol>\n\n\n\n<p>\u8fd9\u91cc\u7684<code>'pt'<\/code>\u4ee3\u8868PyTorch\uff0c\u5982\u679c\u4f60\u5728\u4f7f\u7528TensorFlow\uff0c\u4f60\u53ef\u4ee5\u4f7f\u7528<code>return_tensors='tf'<\/code>\u6765\u83b7\u53d6\u4e00\u4e2aTensorFlow\u7684<code>Tensor<\/code>\u3002\u8fd9\u6837\uff0c\u65e0\u8bba\u662f\u7528\u4e8e\u8bad\u7ec3\u8fd8\u662f\u63a8\u7406\uff0c\u7f16\u7801\u540e\u7684\u6570\u636e\u90fd\u53ef\u4ee5\u76f4\u63a5\u88ab\u6df1\u5ea6\u5b66\u4e60\u6846\u67b6\u6240\u4f7f\u7528\u3002<\/p>\n\n\n\n<p>\u8fd0\u884ctest06.py<\/p>\n\n\n\n<div class=\"wp-block-urvanov-syntax-highlighter-code-block\"><pre class=\"lang:python decode:true \" >python newsrc\/test06.py\nList: [464, 3616, 286, 1204, 318]\nTensor: tensor([[ 464, 3616,  286, 1204,  318]])\ntokens: ['The', '\u0120meaning', '\u0120of', '\u0120life', '\u0120is']<\/pre><\/div>\n\n\n\n<p>\u6210\u529f\u5730\u6267\u884c\u4e86\u7f16\u7801\u64cd\u4f5c\uff0c\u5e76\u4ee5\u4e24\u79cd\u4e0d\u540c\u7684\u683c\u5f0f\u83b7\u53d6\u4e86\u7ed3\u679c\uff1a<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li><strong>List<\/strong>: <code>[464, 3616, 286, 1204, 318]<\/code> \u8fd9\u662f\u4e00\u4e2aPython\u5217\u8868\uff0c\u5305\u542b\u4e86\u7531\u5206\u8bcd\u5668\u8fd4\u56de\u7684token ID\u3002\u6bcf\u4e2aID\u5bf9\u5e94\u4e8e\u5206\u8bcd\u5668\u8bcd\u6c47\u8868\u4e2d\u7684\u4e00\u4e2a\u7279\u5b9atoken\u3002\u8fd9\u79cd\u683c\u5f0f\u9002\u5408\u8fdb\u884c\u8fdb\u4e00\u6b65\u7684Python\u7ea7\u522b\u7684\u5904\u7406\u6216\u5206\u6790\u3002<\/li>\n\n\n\n<li><strong>Tensor<\/strong>: <code>tensor([[ 464, 3616, 286, 1204, 318]])<\/code> \u8fd9\u662f\u4e00\u4e2aPyTorch <code>Tensor<\/code>\uff0c\u540c\u6837\u5305\u542b\u4e86token ID\uff0c\u4f46\u8fd9\u6b21\u5b83\u4eec\u88ab\u5c01\u88c5\u5728\u4e00\u4e2aTensor\u5bf9\u8c61\u4e2d\u3002\u6ce8\u610f\u8fd9\u4e2aTensor\u6709\u4e24\u4e2a\u7ef4\u5ea6\uff0c\u5176\u4e2d\u7b2c\u4e00\u4e2a\u7ef4\u5ea6\uff08\u6279\u6b21\u5927\u5c0f\uff09\u4e3a1\u3002\u8fd9\u662f\u56e0\u4e3a\u4f60\u7f16\u7801\u7684\u662f\u5355\u4e2a\u6587\u672c\u5b57\u7b26\u4e32\u3002\u8fd9\u79cd\u683c\u5f0f\u76f4\u63a5\u9002\u7528\u4e8e\u4f5c\u4e3a\u6a21\u578b\u7684\u8f93\u5165\uff0c\u4fbf\u4e8e\u8fdb\u884c\u6279\u5904\u7406\u548cGPU\u52a0\u901f\u3002<\/li>\n<\/ol>\n\n\n\n<p>\u8fd9\u4e24\u79cd\u8868\u793a\u65b9\u6cd5\u5404\u6709\u7528\u9014\uff0c\u9009\u62e9\u54ea\u4e00\u79cd\u53d6\u51b3\u4e8e\u4f60\u63a5\u4e0b\u6765\u7684\u8ba1\u5212\u3002\u5982\u679c\u4f60\u8981\u5bf9\u8fd9\u4e9b\u7f16\u7801\u540e\u7684ID\u8fdb\u884c\u67d0\u79cd\u5904\u7406\u6216\u5206\u6790\uff0c\u4f7f\u7528Python\u5217\u8868\u53ef\u80fd\u66f4\u65b9\u4fbf\u3002\u5982\u679c\u4f60\u51c6\u5907\u5c06\u8fd9\u4e9bID\u4f5c\u4e3a\u8f93\u5165\u4f20\u9012\u7ed9PyTorch\u6a21\u578b\u8fdb\u884c\u8bad\u7ec3\u6216\u63a8\u7406\uff0c\u90a3\u4e48\u4f7f\u7528Tensor\u683c\u5f0f\u4f1a\u66f4\u76f4\u63a5\u6709\u6548\u3002<\/p>\n\n\n\n<p>\u4ee5\u5c06\u7f16\u7801\u540e\u7684token ID\u5217\u8868<code>[464, 3616, 286, 1204, 318]<\/code>\u8f6c\u6362\u56de\u5bf9\u5e94\u7684token\u5b57\u7b26\u4e32\u3002\u8fd9\u91cc\u4f7f\u7528\u7684\u662f<code>tokenizer.convert_ids_to_tokens<\/code>\u65b9\u6cd5\uff0c\u5b83\u63a5\u53d7token ID\u5217\u8868\u4f5c\u4e3a\u8f93\u5165\uff0c\u5e76\u8fd4\u56de\u4e00\u4e2a\u5305\u542b\u6bcf\u4e2aID\u5bf9\u5e94token\u7684\u5b57\u7b26\u4e32\u5217\u8868\u3002<\/p>\n\n\n\n<p>tokens = tokenizer.convert_ids_to_tokens(encoded_prompt_list) \u5bf9\u4e8e\u7ed9\u5b9a\u7684ID\u5217\u8868\uff0c\u8fd9\u4e2a\u65b9\u6cd5\u5c06\u8fd4\u56de\u6bcf\u4e2aID\u5bf9\u5e94\u7684\u539f\u59cbtoken\u3002\u5728\u4f8b\u5b50\u4e2d\uff0c\u8fd9\u4e2a\u8fc7\u7a0b\u4f1a\u5c06prompt &#8220;The meaning of life is&#8221;\u4e2d\u6bcf\u4e2a\u8bcd\uff08\u4ee5\u53ca\u53ef\u80fd\u7684\u7279\u6b8a\u5b57\u7b26\u6216\u5b50\u8bcd\uff09\u7684ID\u8f6c\u6362\u56de\u53ef\u8bfb\u7684\u6587\u672c\u5f62\u5f0f\u3002<\/p>\n\n\n\n<p>\u8fd9\u91cc\u7684\u8f93\u51fa\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><code>'The'<\/code><\/li>\n\n\n\n<li><code>'\u0120meaning'<\/code><\/li>\n\n\n\n<li><code>'\u0120of'<\/code><\/li>\n\n\n\n<li><code>'\u0120life'<\/code><\/li>\n\n\n\n<li><code>'\u0120is'<\/code><\/li>\n<\/ul>\n\n\n\n<p>\u4e2d\u7684<code>'\u0120'<\/code>\u5b57\u7b26\uff08\u770b\u8d77\u6765\u50cf\u4e00\u4e2a\u7a7a\u683c\uff0c\u4f46\u5b9e\u9645\u4e0a\u662f\u4e00\u4e2a\u4e0d\u95f4\u65ad\u7a7a\u683c\uff09\u662fGPT-2\u5206\u8bcd\u5668\u7279\u6709\u7684\uff0c\u7528\u4e8e\u8868\u793a\u4e00\u4e2a\u5355\u8bcd\u7684\u5f00\u59cb\uff0c\u9664\u4e86\u7b2c\u4e00\u4e2a\u5355\u8bcd\u4e4b\u5916\u3002\u8fd9\u662f\u56e0\u4e3aGPT-2\u4f7f\u7528\u7684\u662fByte Pair Encoding\uff08BPE\uff09\u4f5c\u4e3a\u5176\u5206\u8bcd\u65b9\u6cd5\uff0c\u5176\u4e2d<code>'\u0120'<\/code>\u7528\u6765\u533a\u5206\u7d27\u8ddf\u5728\u5176\u4ed6\u5355\u8bcd\u540e\u9762\u7684\u5355\u8bcd\u3002\u5728\u5b9e\u9645\u6587\u672c\u4e2d\uff0c\u8fd9\u4e2a\u5b57\u7b26\u901a\u5e38\u4e0d\u4f1a\u663e\u793a\uff0c\u6216\u8005\u53ef\u4ee5\u88ab\u7406\u89e3\u4e3a\u5355\u8bcd\u4e4b\u95f4\u7684\u7a7a\u683c\u3002<\/p>\n\n\n\n<p>\u8fd9\u6837\u7684\u8f6c\u6362\u662fNLP\u4efb\u52a1\u4e2d\u5e38\u89c1\u7684\u4e00\u4e2a\u6b65\u9aa4\uff0c\u5b83\u5141\u8bb8\u6211\u4eec\u5728\u5904\u7406\u6587\u672c\u548c\u6a21\u578b\u5185\u90e8\u8868\u793a\u4e4b\u95f4\u8fdb\u884c\u8f6c\u6362\uff0c\u4ece\u800c\u53ef\u4ee5\u66f4\u5bb9\u6613\u5730\u7406\u89e3\u6a21\u578b\u7684\u8f93\u5165\u548c\u8f93\u51fa\u3002<\/p>\n\n\n\n<p><strong>\u95ee\uff1aencoded_prompt_list \u8fd4\u56de\u7684\u5b9e\u9645\u4e0a\u662f transformer.wte.weight \u6743\u91cd\u7684\u8bcd\u7d22\u5f15\u503c<\/strong>\uff1f<\/p>\n\n\n\n<p>\u662f\u7684\uff0c\u4f60\u7406\u89e3\u5f97\u975e\u5e38\u6b63\u786e\u3002\u5728\u4f7f\u7528\u5206\u8bcd\u5668\u5bf9\u6587\u672c\u8fdb\u884c\u7f16\u7801\u65f6\uff0c\u8fd4\u56de\u7684<code>encoded_prompt_list<\/code>\u5b9e\u9645\u4e0a\u662f\u8bcd\u6c47\u8868\u4e2d\u5bf9\u5e94\u5355\u8bcd\u6216token\u7684\u7d22\u5f15\u503c\u3002\u8fd9\u4e9b\u7d22\u5f15\u503c\u76f4\u63a5\u5bf9\u5e94\u4e8e<code>transformer.wte.weight<\/code>\uff08\u5373\u8bcd\u5d4c\u5165\u6743\u91cd\u77e9\u9635\uff09\u4e2d\u7684\u884c\u3002\u6362\u53e5\u8bdd\u8bf4\uff0c\u6bcf\u4e2a\u7d22\u5f15\u6307\u5411<code>transformer.wte.weight<\/code>\u4e2d\u7684\u4e00\u4e2a\u7279\u5b9a\u884c\uff0c\u800c\u8fd9\u4e00\u884c\u5c31\u662f\u5bf9\u5e94\u5355\u8bcd\u6216token\u7684\u5d4c\u5165\u5411\u91cf\u3002<\/p>\n\n\n\n<p>\u5728GPT-2\u8fd9\u6837\u7684\u6a21\u578b\u4e2d\uff0c<code>transformer.wte.weight<\/code>\u662f\u4e00\u4e2a\u5de8\u5927\u7684\u77e9\u9635\uff0c\u5176\u5f62\u72b6\u901a\u5e38\u662f<code>[vocab_size, embedding_size]<\/code>\uff0c\u5176\u4e2d<code>vocab_size<\/code>\u662f\u5206\u8bcd\u5668\u8bcd\u6c47\u8868\u7684\u5927\u5c0f\uff08\u5728GPT-2\u7684\u60c5\u51b5\u4e0b\u662f50257\uff09\uff0c\u800c<code>embedding_size<\/code>\u662f\u5d4c\u5165\u5411\u91cf\u7684\u7ef4\u5ea6\uff08\u5728GPT-2\u7684\u60c5\u51b5\u4e0b\u662f768\uff09\u3002\u8fd9\u610f\u5473\u7740\u77e9\u9635\u4e2d\u7684\u6bcf\u4e00\u884c\u4ee3\u8868\u4e00\u4e2a\u4e0d\u540c\u7684token\u7684\u5d4c\u5165\u5411\u91cf\uff0c\u8fd9\u4e9b\u5411\u91cf\u5728\u6a21\u578b\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u88ab\u5b66\u4e60\u3002<\/p>\n\n\n\n<p>\u56e0\u6b64\uff0c\u5f53\u4f60\u901a\u8fc7<code>tokenizer.encode()<\/code>\u83b7\u53d6\u5230\u4e00\u7cfb\u5217\u7d22\u5f15\u503c\u65f6\uff0c\u8fd9\u4e9b\u7d22\u5f15\u503c\u53ef\u4ee5\u88ab\u7528\u6765\u4ece<code>transformer.wte.weight<\/code>\u4e2d\u68c0\u7d22\u51fa\u76f8\u5e94\u7684\u5d4c\u5165\u5411\u91cf\u3002\u8fd9\u4e9b\u5d4c\u5165\u5411\u91cf\u968f\u540e\u88ab\u7528\u4f5c\u6a21\u578b\u7684\u8f93\u5165\uff0c\u8fdb\u884c\u524d\u5411\u4f20\u64ad\u4ee5\u6267\u884c\u5404\u79cd\u4efb\u52a1\uff0c\u5982\u6587\u672c\u751f\u6210\u3001\u5206\u7c7b\u7b49\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\"><strong>9. token ID\u5bf9\u5e94\u7684\u8bcd\u5d4c\u5165\u77e9\u9635<\/strong><\/h2>\n\n\n\n<p>\u5982\u679c\u4f60\u60f3\u67e5\u770b\u4e0e<code>encoded_prompt_list<\/code>\u4e2d\u7684token ID\u5bf9\u5e94\u7684\u8bcd\u5d4c\u5165\u77e9\u9635\uff08\u5373<code>transformer.wte.weight<\/code>\u4e2d\u7684\u5177\u4f53\u884c\uff09\uff0c\u4f60\u53ef\u4ee5\u6309\u7167\u4e0b\u9762\u7684\u65b9\u6cd5\u6765\u5b9e\u73b0\u3002\u8fd9\u6bb5\u4ee3\u7801\u5c06\u7d22\u5f15\u8bcd\u5d4c\u5165\u6743\u91cd\u77e9\u9635\uff0c\u83b7\u53d6\u7279\u5b9atoken ID\u5bf9\u5e94\u7684\u5d4c\u5165\u5411\u91cf\u3002<\/p>\n\n\n\n<p>\u9996\u5148\uff0c\u786e\u4fdd\u4f60\u5df2\u7ecf\u6709\u4e86<code>encoded_prompt_list<\/code>\u8fd9\u4e2a\u5217\u8868\uff0c\u5b83\u5305\u542b\u4e86\u4f60\u611f\u5174\u8da3\u7684token\u7684ID\u3002\u7136\u540e\uff0c\u4f7f\u7528\u8fd9\u4e9bID\u6765\u7d22\u5f15<code>transformer.wte.weight<\/code>\u77e9\u9635,\u547d\u540d\u4e3a test07.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 GPT2Model\n\n# \u52a0\u8f7d\u9884\u8bad\u7ec3\u7684GPT-2\u6a21\u578b\nmodel = GPT2Model.from_pretrained('gpt2')\n\n# \u5047\u8bbe encoded_prompt_list \u5df2\u7ecf\u88ab\u5b9a\u4e49\u5e76\u5305\u542b\u4e86\u611f\u5174\u8da3\u7684token ID\n# encoded_prompt_list = [464, 3616, 286, 1204, 318]\n\n# \u83b7\u53d6\u8bcd\u5d4c\u5165\u77e9\u9635\nembedding_matrix = model.get_input_embeddings().weight.data\n\n# \u4f7f\u7528encoded_prompt_list\u4e2d\u7684ID\u7d22\u5f15\u5d4c\u5165\u77e9\u9635\ntoken_embeddings = embedding_matrix[encoded_prompt_list]\n\nprint(\"Selected token embeddings:\")\nprint(token_embeddings)\n<\/pre><\/div>\n\n\n\n<p>\u8fd9\u6bb5\u4ee3\u7801\u9996\u5148\u52a0\u8f7d\u4e86GPT-2\u6a21\u578b\uff0c\u7136\u540e\u83b7\u53d6\u4e86\u8bcd\u5d4c\u5165\u77e9\u9635\u3002<code>embedding_matrix<\/code>\u662f\u4e00\u4e2a\u5927\u77e9\u9635\uff0c\u5176\u4e2d\u6bcf\u4e00\u884c\u5bf9\u5e94\u4e8e\u8bcd\u6c47\u8868\u4e2d\u4e00\u4e2atoken\u7684\u5d4c\u5165\u5411\u91cf\u3002\u901a\u8fc7<code>embedding_matrix[encoded_prompt_list]<\/code>\u8fd9\u884c\u4ee3\u7801\uff0c\u6211\u4eec\u4f7f\u7528<code>encoded_prompt_list<\/code>\u4e2d\u7684token ID\u6765\u7d22\u5f15\u8fd9\u4e2a\u77e9\u9635\uff0c\u4ece\u800c\u5f97\u5230\u5bf9\u5e94\u7684\u5d4c\u5165\u5411\u91cf\u3002<\/p>\n\n\n\n<p><code>token_embeddings<\/code>\u53d8\u91cf\u73b0\u5728\u5305\u542b\u4e86\u4f60\u6307\u5b9a\u7684token ID\u5bf9\u5e94\u7684\u5d4c\u5165\u5411\u91cf\u3002\u8fd9\u4e9b\u5411\u91cf\u662f\u9ad8\u7ef4\u7a7a\u95f4\u4e2d\u7684\u70b9\uff0c\u5b83\u4eec\u6355\u83b7\u4e86\u4e0e\u6bcf\u4e2atoken\u76f8\u5173\u7684\u8bed\u4e49\u548c\u8bed\u6cd5\u4fe1\u606f\u3002<\/p>\n\n\n\n<p>\u8fd0\u884ctest07.py<\/p>\n\n\n\n<div class=\"wp-block-urvanov-syntax-highlighter-code-block\"><pre class=\"lang:sh decode:true \" > python newsrc\/test07.py\nList: [464, 3616, 286, 1204, 318]\nSelected token embeddings:\ntensor([[-0.0686, -0.0203,  0.0645,  ...,  0.0639, -0.0100,  0.0031],\n        [ 0.1675, -0.0834,  0.0147,  ..., -0.0559, -0.2029, -0.0829],\n        [-0.0572,  0.0183,  0.0333,  ..., -0.0689, -0.0931, -0.0714],\n        [ 0.2109,  0.1353,  0.0158,  ..., -0.0287, -0.0276, -0.1644],\n        [-0.0097,  0.0101,  0.0556,  ...,  0.1145, -0.0380, -0.0254]])<\/pre><\/div>\n","protected":false},"excerpt":{"rendered":"<p>5. \u5982\u4f55\u6253\u5370\u6a21\u578b\u53c2\u6570\uff1f \u8981\u6253\u5370\u4e00\u4e2a\u9884\u8bad\u7ec3\u6a21\u578b\u7684\u53c2\u6570\uff0c\u4f60\u53ef\u4ee5\u4f7f\u7528transformers\u5e93\u4e2d\u6a21\u578b\u7684.named [&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-2366","post","type-post","status-publish","format-standard","hentry","category-chatgpt","category-gpt","category-llm","category-llms","category-openai","tag-chatgpt","tag-openai-api"],"views":1507,"jetpack_sharing_enabled":true,"jetpack_featured_media_url":"","_links":{"self":[{"href":"https:\/\/www.aqwu.net\/wp\/index.php?rest_route=\/wp\/v2\/posts\/2366","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=2366"}],"version-history":[{"count":14,"href":"https:\/\/www.aqwu.net\/wp\/index.php?rest_route=\/wp\/v2\/posts\/2366\/revisions"}],"predecessor-version":[{"id":2383,"href":"https:\/\/www.aqwu.net\/wp\/index.php?rest_route=\/wp\/v2\/posts\/2366\/revisions\/2383"}],"wp:attachment":[{"href":"https:\/\/www.aqwu.net\/wp\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=2366"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.aqwu.net\/wp\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=2366"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.aqwu.net\/wp\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=2366"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}