{"id":2394,"date":"2024-03-10T22:15:08","date_gmt":"2024-03-10T14:15:08","guid":{"rendered":"https:\/\/www.aqwu.net\/wp\/?p=2394"},"modified":"2024-04-28T20:03:18","modified_gmt":"2024-04-28T12:03:18","slug":"%e4%ba%86%e8%a7%a3-openai-%e7%9a%84-gpt2-%e6%a8%a1%e5%9e%8b%e7%bb%93%e6%9e%844","status":"publish","type":"post","link":"https:\/\/www.aqwu.net\/wp\/?p=2394","title":{"rendered":"\u4e86\u89e3 OpenAI \u7684 gpt2 \u6a21\u578b\u7ed3\u6784(4)"},"content":{"rendered":"\n

12. Transformer\u5c42<\/strong><\/h2>\n\n\n\n

transformer.h<\/code>\u5728GPT-2\u6a21\u578b\u4e2d\u6307\u7684\u662f\u6a21\u578b\u7684\u591a\u4e2aTransformer\u5c42\uff08\u4e5f\u88ab\u79f0\u4e3aTransformer blocks\u6216layers\uff09\u3002\u8fd9\u4e9b\u5c42\u6784\u6210\u4e86\u6a21\u578b\u7684\u6838\u5fc3\uff0c\u8d1f\u8d23\u5904\u7406\u8f93\u5165\u6570\u636e\u7684\u5e8f\u5217\uff0c\u6267\u884c\u81ea\u6ce8\u610f\u529b\u673a\u5236\uff0c\u4ee5\u53ca\u8fdb\u884c\u76f8\u5173\u7684\u524d\u9988\u7f51\u7edc\u8ba1\u7b97\u3002\u6bcf\u4e00\u5c42\u90fd\u4f1a\u5bf9\u6570\u636e\u8fdb\u884c\u8fdb\u4e00\u6b65\u7684\u62bd\u8c61\u548c\u8f6c\u6362\uff0c\u4ece\u800c\u5141\u8bb8\u6a21\u578b\u6355\u6349\u590d\u6742\u7684\u8bed\u8a00\u7279\u5f81\u548c\u6a21\u5f0f\u3002<\/p>\n\n\n\n

12.1 Transformer\u5c42\u7684\u7ec4\u6210<\/strong><\/h3>\n\n\n\n

\u6bcf\u4e2aTransformer\u5c42\u901a\u5e38\u5305\u542b\u4ee5\u4e0b\u51e0\u4e2a\u5173\u952e\u90e8\u5206\uff1a<\/p>\n\n\n\n

    \n
  1. \u81ea\u6ce8\u610f\u529b\uff08Self-Attention\uff09\u673a\u5236<\/strong>\uff1a\u4f7f\u6a21\u578b\u80fd\u591f\u5728\u5904\u7406\u6bcf\u4e2a\u5355\u8bcd\uff08\u6216token\uff09\u65f6\uff0c\u8003\u8651\u5230\u6574\u4e2a\u53e5\u5b50\u4e2d\u7684\u6240\u6709\u5355\u8bcd\u3002\u8fd9\u79cd\u673a\u5236\u5e2e\u52a9\u6a21\u578b\u7406\u89e3\u5355\u8bcd\u4e4b\u95f4\u7684\u4e0a\u4e0b\u6587\u5173\u7cfb\uff0c\u4f8b\u5982\u6355\u6349\u957f\u8ddd\u79bb\u4f9d\u8d56\u3002<\/li>\n\n\n\n
  2. \u5f52\u4e00\u5316\uff08Normalization\uff09<\/strong>\uff1a\u5728\u81ea\u6ce8\u610f\u529b\u548c\u524d\u9988\u7f51\u7edc\u7684\u5404\u4e2a\u6b65\u9aa4\u4e4b\u540e\uff0c\u901a\u5e38\u4f1a\u6709\u5c42\u5f52\u4e00\u5316\uff08Layer Normalization\uff09\u6b65\u9aa4\uff0c\u4ee5\u5e2e\u52a9\u7a33\u5b9a\u8bad\u7ec3\u8fc7\u7a0b\uff0c\u52a0\u5feb\u6536\u655b\u901f\u5ea6\u3002<\/li>\n\n\n\n
  3. \u524d\u9988\u7f51\u7edc\uff08Feed-Forward Networks\uff09<\/strong>\uff1a\u5728\u6bcf\u4e2a\u81ea\u6ce8\u610f\u529b\u5c42\u4e4b\u540e\uff0c\u6709\u4e00\u4e2a\u524d\u9988\u7f51\u7edc\uff0c\u5b83\u5bf9\u81ea\u6ce8\u610f\u529b\u5c42\u7684\u8f93\u51fa\u8fdb\u884c\u8fdb\u4e00\u6b65\u7684\u5904\u7406\u3002\u8fd9\u4e2a\u7f51\u7edc\u901a\u5e38\u662f\u4e24\u4e2a\u7ebf\u6027\u53d8\u6362\uff0c\u4e2d\u95f4\u6709\u4e00\u4e2a\u6fc0\u6d3b\u51fd\u6570\u3002<\/li>\n\n\n\n
  4. \u6b8b\u5dee\u8fde\u63a5\uff08Residual Connection\uff09<\/strong>\uff1a\u81ea\u6ce8\u610f\u529b\u548c\u524d\u9988\u7f51\u7edc\u7684\u6bcf\u4e00\u6b65\u90fd\u4f1a\u6709\u6b8b\u5dee\u8fde\u63a5\uff0c\u5e76\u4e14\u5728\u901a\u8fc7\u5f52\u4e00\u5316\u5c42\u4e4b\u524d\uff0c\u8fd9\u4e2a\u8fde\u63a5\u4f1a\u5c06\u8f93\u5165\u76f4\u63a5\u52a0\u5230\u8f93\u51fa\u4e0a\u3002\u8fd9\u6709\u52a9\u4e8e\u907f\u514d\u5728\u6df1\u5c42\u7f51\u7edc\u4e2d\u51fa\u73b0\u7684\u68af\u5ea6\u6d88\u5931\u95ee\u9898\u3002<\/li>\n<\/ol>\n\n\n\n

    12.2 \u5de5\u4f5c\u539f\u7406<\/strong><\/h3>\n\n\n\n

    \u5f53\u8f93\u5165\u6570\u636e\uff08\u7ecf\u8fc7\u8bcd\u5d4c\u5165\u548c\u4f4d\u7f6e\u7f16\u7801\u76f8\u52a0\u540e\u7684\u5411\u91cf\uff09\u8fdb\u5165\u7b2c\u4e00\u4e2aTransformer\u5c42\u65f6\uff0c\u5b83\u9996\u5148\u7ecf\u8fc7\u81ea\u6ce8\u610f\u529b\u673a\u5236\u7684\u5904\u7406\uff0c\u7136\u540e\u662f\u6b8b\u5dee\u8fde\u63a5\u548c\u5f52\u4e00\u5316\u3002\u4e4b\u540e\uff0c\u6570\u636e\u8fdb\u5165\u524d\u9988\u7f51\u7edc\uff0c\u518d\u6b21\u7ecf\u8fc7\u6b8b\u5dee\u8fde\u63a5\u548c\u5f52\u4e00\u5316\u3002\u8fd9\u6837\u5904\u7406\u540e\u7684\u6570\u636e\u4f1a\u88ab\u4f20\u9012\u5230\u4e0b\u4e00\u4e2aTransformer\u5c42\uff0c\u4f9d\u6b64\u7c7b\u63a8\uff0c\u76f4\u5230\u901a\u8fc7\u6240\u6709\u7684Transformer\u5c42\u3002<\/p>\n\n\n\n

    \u6bcf\u7ecf\u8fc7\u4e00\u4e2a\u5c42\uff0c\u6a21\u578b\u90fd\u80fd\u591f\u6355\u6349\u5230\u66f4\u52a0\u590d\u6742\u548c\u62bd\u8c61\u7684\u4fe1\u606f\u3002\u901a\u8fc7\u8fd9\u79cd\u65b9\u5f0f\uff0c\u6a21\u578b\u53ef\u4ee5\u5b66\u4e60\u5230\u975e\u5e38\u590d\u6742\u7684\u8bed\u8a00\u7279\u5f81\uff0c\u4ece\u800c\u8fdb\u884c\u6709\u6548\u7684\u9884\u6d4b\u548c\u751f\u6210\u3002<\/p>\n\n\n\n

    12.3 \u5728\u4ee3\u7801\u4e2d\u7684\u4f53\u73b0<\/strong><\/h3>\n\n\n\n

    \u5728Hugging Face\u7684Transformers\u5e93\u4e2d\uff0c\u4f60\u53ef\u4ee5\u901a\u8fc7\u8bbf\u95ee\u6a21\u578b\u7684.transformer.h<\/code>\u5c5e\u6027\u6765\u67e5\u770b\u6216\u64cd\u4f5c\u8fd9\u4e9b\u5c42\u3002\u547d\u540d\u4e3a test10.py\uff0c\u6587\u4ef6\u4fdd\u5b58\u5230 newsrc \u76ee\u5f55\u4e0b\uff1a<\/p>\n\n\n\n

    from transformers import GPT2LMHeadModel, GPT2Tokenizer\n\n# \u52a0\u8f7d\u9884\u8bad\u7ec3\u7684GPT-2\u6a21\u578b\u548c\u5206\u8bcd\u5668\nmodel = GPT2LMHeadModel.from_pretrained('gpt2')\n\n# \u8bbf\u95ee\u7b2c\u4e00\u4e2aTransformer\u5c42\nfirst_layer = model.transformer.h[0]\n\nprint(\"first_layer:\")\nprint(first_layer)\n\n# \u904d\u5386\u6a21\u578b\u7684\u6240\u6709\u53c2\u6570\uff0c\u5e76\u7b5b\u9009\u51fa\u7b2c\u4e00\u4e2aTransformer\u5c42\u7684\u53c2\u6570\nfor name, param in model.named_parameters():\n    if \"transformer.h.0.\" in name:\n        print(f\"Parameter Name: {name}   Size: {param.size()}\")<\/pre><\/div>\n\n\n\n
    \u8fd9\u6bb5\u4ee3\u7801\u9996\u5148\u52a0\u8f7d\u4e86\u9884\u8bad\u7ec3\u7684GPT-2\u6a21\u578b\u3002\u7136\u540e\uff0c\u5b83\u4f7f\u7528named_parameters()<\/code>\u65b9\u6cd5\u6765\u904d\u5386\u6a21\u578b\u4e2d\u7684\u6240\u6709\u53c2\u6570\u3002\u8fd9\u4e2a\u65b9\u6cd5\u8fd4\u56de\u6bcf\u4e2a\u53c2\u6570\u7684\u540d\u79f0\u548c\u53c2\u6570\u5bf9\u8c61\u672c\u8eab\u3002\u901a\u8fc7\u68c0\u67e5\u53c2\u6570\u540d\u79f0\u662f\u5426\u5305\u542b\"transformer.h.0.\"<\/code>\uff0c\u4ee3\u7801\u53ef\u4ee5\u7b5b\u9009\u51fa\u7b2c\u4e00\u4e2aTransformer\u5c42\u76f8\u5173\u7684\u53c2\u6570\u3002\u5bf9\u4e8e\u6bcf\u4e2a\u7b26\u5408\u6761\u4ef6\u7684\u53c2\u6570\uff0c\u5b83\u6253\u5370\u51fa\u53c2\u6570\u7684\u540d\u79f0\u548c\u5c3a\u5bf8\u3002<\/p>\n\n\n\n
    \u8fd9\u6837\uff0c\u4f60\u5c31\u80fd\u591f\u5f97\u5230\u4e0e\u4f60\u63d0\u4f9b\u7684\u53c2\u6570\u4fe1\u606f\u76f8\u5339\u914d\u7684\u8f93\u51fa\uff0c\u5c55\u793a\u4e86\u7b2c\u4e00\u4e2aTransformer\u5c42\u4e2d\u5404\u4e2a\u7ec4\u4ef6\u7684\u53c2\u6570\u540d\u79f0\u548c\u5c3a\u5bf8\u3002\u8fd9\u5305\u62ec\u5c42\u5f52\u4e00\u5316\uff08Layer Normalization\uff09\u3001\u81ea\u6ce8\u610f\u529b\uff08Self-Attention\uff09\u673a\u5236\u7684\u53c2\u6570\uff0c\u4ee5\u53ca\u524d\u9988\u7f51\u7edc\uff08Feed-Forward Network\uff09\u7684\u6743\u91cd\u548c\u504f\u7f6e\u3002<\/p>\n\n\n\n
    \u8fd0\u884c setup10.py<\/p>\n\n\n\n
    python newsrc\/test10.py\nfirst_layer:\nGPT2Block(\n  (ln_1): LayerNorm((768,), eps=1e-05, elementwise_affine=True)\n  (attn): GPT2Attention(\n    (c_attn): Conv1D()\n    (c_proj): Conv1D()\n    (attn_dropout): Dropout(p=0.1, inplace=False)\n    (resid_dropout): Dropout(p=0.1, inplace=False)\n  )\n  (ln_2): LayerNorm((768,), eps=1e-05, elementwise_affine=True)\n  (mlp): GPT2MLP(\n    (c_fc): Conv1D()\n    (c_proj): Conv1D()\n    (act): NewGELUActivation()\n    (dropout): Dropout(p=0.1, inplace=False)\n  )\n)\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])<\/pre><\/div>\n\n\n\n
    \u6839\u636e\u4f60\u63d0\u4f9b\u7684\u4fe1\u606f\uff0c\u8fd9\u4e9b\u53c2\u6570\u8be6\u7ec6\u63cf\u8ff0\u4e86GPT-2\u6a21\u578b\u4e2d\u7b2c\u4e00\u4e2aTransformer\u5c42\u7684\u5404\u4e2a\u7ec4\u4ef6\u3002\u6bcf\u4e00\u9879\u90fd\u662f\u6a21\u578b\u5b66\u4e60\u8fc7\u7a0b\u4e2d\u7684\u5173\u952e\u90e8\u5206\uff0c\u5b83\u4eec\u5171\u540c\u5de5\u4f5c\u5904\u7406\u8f93\u5165\u6570\u636e\uff0c\u6355\u6349\u6587\u672c\u7684\u590d\u6742\u6a21\u5f0f\u3002\u8fd9\u91cc\u7b80\u5355\u56de\u987e\u4e00\u4e0b\u6bcf\u4e2a\u53c2\u6570\u7684\u4f5c\u7528\uff1a<\/p>\n\n\n\n
      \n
    • \u5c42\u5f52\u4e00\u5316\uff08Layer Normalization\uff09\u53c2\u6570<\/strong>\uff08ln_1<\/code> \u548c ln_2<\/code>\uff09:\n
        \n
      • ln_1.weight<\/code> \u548c ln_1.bias<\/code>\uff0c\u4ee5\u53ca ln_2.weight<\/code> \u548c ln_2.bias<\/code>\uff0c\u662f\u5c42\u5f52\u4e00\u5316\u64cd\u4f5c\u7684\u6743\u91cd\u548c\u504f\u7f6e\u53c2\u6570\uff0c\u5b83\u4eec\u5e2e\u52a9\u7a33\u5b9a\u6a21\u578b\u7684\u8bad\u7ec3\u8fc7\u7a0b\uff0c\u63d0\u9ad8\u6536\u655b\u901f\u5ea6\u3002\u5c42\u5f52\u4e00\u5316\u5728\u81ea\u6ce8\u610f\u529b\u673a\u5236\u548c\u524d\u9988\u7f51\u7edc\u4e4b\u540e\u5e94\u7528\u3002<\/li>\n\n\n\n
      • \u5c3a\u5bf8\u4e3a[768]<\/code>\uff0c\u5bf9\u5e94\u4e8e\u6a21\u578b\u7684\u9690\u85cf\u5c42\u5927\u5c0f\u3002<\/li>\n<\/ul>\n<\/li>\n\n\n\n
      • \u81ea\u6ce8\u610f\u529b\uff08Self-Attention\uff09\u53c2\u6570<\/strong>\uff08attn<\/code>\uff09:\n
          \n
        • attn.c_attn.weight<\/code> \u548c attn.c_attn.bias<\/code> \u662f\u81ea\u6ce8\u610f\u529b\u5c42\u4e2d\u7528\u4e8e\u751f\u6210\u67e5\u8be2\uff08Query\uff09\u3001\u952e\uff08Key\uff09\u548c\u503c\uff08Value\uff09\u7684\u5408\u5e76\u77e9\u9635\u7684\u6743\u91cd\u548c\u504f\u7f6e\u3002<\/li>\n\n\n\n
        • \u8fd9\u4e9b\u53c2\u6570\u5c5e\u4e8e\u81ea\u6ce8\u610f\u529b\u673a\u5236\u4e2d\u7684\u591a\u5934\u6ce8\u610f\u529b\uff08Multi-head Attention\uff09\u7684\u5408\u5e76\u6ce8\u610f\u529b\uff08Concatenated Attention\uff09\u77e9\u9635\u7684\u6743\u91cd\u548c\u504f\u7f6e\u3002<\/li>\n\n\n\n
        • c_attn.weight<\/code>\u7684\u5c3a\u5bf8\u4e3a[768, 2304]<\/code>\uff0c\u8fd9\u91cc768\u662f\u9690\u85cf\u5c42\u5927\u5c0f\uff0c2304\u6765\u81ea\u4e8e\u5c06768\u4e58\u4ee53\uff08\u56e0\u4e3a\u5bf9\u4e8e\u6bcf\u4e2a\u5934\u90e8\uff0c\u6709\u4e09\u4e2a\u77e9\u9635\uff1a\u67e5\u8be2\uff08Query\uff09\u3001\u952e\uff08Key\uff09\u3001\u503c\uff08Value\uff09\uff09\uff0cGPT-2\u9ed8\u8ba4\u4f7f\u752812\u4e2a\u6ce8\u610f\u529b\u5934\uff0c\u6240\u4ee5\u6bcf\u4e2a\u5934\u90e8\u6709768\/12=64\u4e2a\u7279\u5f81\u3002<\/li>\n\n\n\n
        • c_attn.bias<\/code>\u7684\u5c3a\u5bf8\u4e3a[2304]<\/code>\uff0c\u5bf9\u5e94\u4e8e\u5408\u5e76\u6ce8\u610f\u529b\u77e9\u9635\u7684\u504f\u7f6e\u3002<\/li>\n\n\n\n
        • attn.c_proj.weight<\/code> \u548c attn.c_proj.bias<\/code> \u662f\u81ea\u6ce8\u610f\u529b\u5c42\u8f93\u51fa\u7684\u6295\u5f71\u5c42\u7684\u6743\u91cd\u548c\u504f\u7f6e\uff0c\u7528\u4e8e\u5c06\u6ce8\u610f\u529b\u5c42\u7684\u8f93\u51fa\u8f6c\u6362\u56de\u6a21\u578b\u7684\u9690\u85cf\u7ef4\u5ea6\u3002<\/li>\n\n\n\n
        • \u8fd9\u4e9b\u53c2\u6570\u662f\u81ea\u6ce8\u610f\u529b\u673a\u5236\u7684\u6295\u5f71\u5c42\uff08Projection Layer\uff09\u7684\u6743\u91cd\u548c\u504f\u7f6e\uff0c\u7528\u4e8e\u5c06\u591a\u5934\u6ce8\u610f\u529b\u7684\u8f93\u51fa\u5408\u5e76\u6210\u4e0b\u4e00\u4e2a\u5c42\u6216\u524d\u9988\u7f51\u7edc\u80fd\u63a5\u53d7\u7684\u683c\u5f0f\u3002<\/li>\n\n\n\n
        • \u5c3a\u5bf8[768, 768]<\/code>\u8868\u793a\u8f93\u5165\u548c\u8f93\u51fa\u90fd\u4e0e\u9690\u85cf\u5c42\u5927\u5c0f\u4e00\u81f4\u3002<\/li>\n<\/ul>\n<\/li>\n\n\n\n
        • \u524d\u9988\u7f51\u7edc\uff08Feed-Forward Network\uff09\u53c2\u6570<\/strong>\uff08mlp<\/code>\uff09:\n
            \n
          • mlp.c_fc.weight<\/code> \u548c mlp.c_fc.bias<\/code> \u662f\u524d\u9988\u7f51\u7edc\uff08MLP\uff09\u4e2d\u7b2c\u4e00\u5c42\uff08\u901a\u5e38\u79f0\u4e3a\u5168\u8fde\u63a5\u5c42\uff09\u7684\u6743\u91cd\u548c\u504f\u7f6e\uff0c\u5b83\u5c06\u8f93\u5165\u7684\u7ef4\u5ea6\u6269\u5c55\u5230\u4e00\u4e2a\u66f4\u5927\u7684\u7ef4\u5ea6\uff08\u5728\u8fd9\u4e2a\u4f8b\u5b50\u4e2d\u662f768\u52303072\uff09\u3002<\/li>\n\n\n\n
          • c_fc.weight<\/code>\u7684\u5c3a\u5bf8[768, 3072]<\/code>\u548cc_fc.bias<\/code>\u7684\u5c3a\u5bf8[3072]<\/code>\u8868\u793a\u8fd9\u4e00\u5c42\u4f1a\u6269\u5c55\u9690\u85cf\u5c42\u7684\u7ef4\u5ea6\uff0c3072\u662f\u9690\u85cf\u5c42\u5927\u5c0f\u76844\u500d\uff0c\u8fd9\u662fGPT-2\u6a21\u578b\u4e2d\u7684\u4e00\u4e2a\u7279\u5b9a\u9009\u62e9\u3002<\/li>\n\n\n\n
          • mlp.c_proj.weight<\/code> \u548c mlp.c_proj.bias<\/code> \u662f\u524d\u9988\u7f51\u7edc\uff08MLP\uff09\u7b2c\u4e8c\u5c42\uff08\u6295\u5f71\u5c42\uff09\u7684\u6743\u91cd\u548c\u504f\u7f6e\uff0c\u5b83\u5c06\u6570\u636e\u4ece\u6269\u5c55\u7684\u7ef4\u5ea6\u538b\u7f29\u56de\u6a21\u578b\u7684\u539f\u59cb\u9690\u85cf\u7ef4\u5ea6\u3002<\/li>\n\n\n\n
          • \u5c3a\u5bf8[3072, 768]<\/code>\u548c[768]<\/code>\u5206\u522b\u5bf9\u5e94\u4e8e\u6743\u91cd\u548c\u504f\u7f6e\u3002<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n\n\n\n
            \u6bcf\u4e2aTransformer\u5c42\u901a\u8fc7\u8fd9\u4e9b\u7ec4\u4ef6\u5b9e\u73b0\u4e86\u590d\u6742\u7684\u6570\u636e\u5904\u7406\u6d41\u7a0b\uff0c\u5305\u62ec\u5904\u7406\u8f93\u5165\u6570\u636e\u7684\u81ea\u6ce8\u610f\u529b\u8ba1\u7b97\u3001\u6570\u636e\u7684\u5f52\u4e00\u5316\u5904\u7406\uff0c\u4ee5\u53ca\u901a\u8fc7\u524d\u9988\u7f51\u7edc\u7684\u975e\u7ebf\u6027\u53d8\u6362\u3002\u8fd9\u79cd\u7ed3\u6784\u7684\u91cd\u590d\u4f7f\u7528\u4f7f\u5f97GPT-2\u80fd\u591f\u5b66\u4e60\u5e76\u751f\u6210\u9ad8\u5ea6\u590d\u6742\u7684\u8bed\u8a00\u7ed3\u6784\u3002<\/p>\n\n\n\n
            13. \u6700\u7ec8\u5c42\u5f52\u4e00\u5316<\/strong><\/h2>\n\n\n\n
            \u5728\u6a21\u578b\u7684\u6700\u540e\u9636\u6bb5\u4f7f\u7528\u5c42\u5f52\u4e00\u5316\uff08Layer Normalization\uff09\u662fTransformer\u6a21\u578b\u8bbe\u8ba1\u4e2d\u7684\u5e38\u89c1\u505a\u6cd5\u3002\u5b83\u7d27\u8ddf\u5728\u6700\u540e\u4e00\u4e2aTransformer\u5c42\u7684\u8f93\u51fa\u540e\uff0c\u786e\u4fdd\u5728\u8fdb\u884c\u6700\u7ec8\u7684\u8f93\u51fa\u9884\u6d4b\u4e4b\u524d\uff0c\u6570\u636e\u5904\u4e8e\u4e00\u4e2a\u7a33\u5b9a\u7684\u5206\u5e03\u72b6\u6001\u3002\u5373\u6a21\u578b\u7684\u6700\u540e\u4e00\u4e2a\u5904\u7406\u6b65\u9aa4\u4e4b\u4e00\u3002\u5728GPT-2\u548c\u8bb8\u591aTransformer\u6a21\u578b\u4e2d\uff0c\u5c42\u5f52\u4e00\u5316\u7528\u4e8e\u7a33\u5b9a\u9690\u85cf\u5c42\u7684\u8f93\u51fa\uff0c\u6709\u52a9\u4e8e\u6a21\u578b\u7684\u8bad\u7ec3\u548c\u6cdb\u5316\u80fd\u529b\u3002<\/p>\n\n\n\n
            13.1 \u5c42\u5f52\u4e00\u5316\uff08Layer Normalization\uff09<\/strong><\/h3>\n\n\n\n
              \n
            • \u4f5c\u7528<\/strong>\uff1a\u5c42\u5f52\u4e00\u5316\u901a\u8fc7\u5bf9\u6bcf\u4e2a\u6837\u672c\u7684\u6240\u6709\u7279\u5f81\u8fdb\u884c\u5f52\u4e00\u5316\uff0c\u5e2e\u52a9\u51cf\u5c11\u5185\u90e8\u534f\u53d8\u91cf\u504f\u79fb\uff08Internal Covariate Shift\uff09\uff0c\u5373\u51cf\u5c11\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u5c42\u8f93\u5165\u5206\u5e03\u7684\u6539\u53d8\u3002\u8fd9\u79cd\u505a\u6cd5\u53ef\u4ee5\u52a0\u901f\u8bad\u7ec3\u8fc7\u7a0b\uff0c\u540c\u65f6\u63d0\u9ad8\u6a21\u578b\u7684\u7a33\u5b9a\u6027\u548c\u6027\u80fd\u3002<\/li>\n\n\n\n
            • \u53c2\u6570<\/strong>\uff1a\n
                \n
              • transformer.ln_f.weight<\/code><\/strong>\uff1a\u8fd9\u662f\u5c42\u5f52\u4e00\u5316\u7684\u7f29\u653e\u53c2\u6570\uff08scale\uff09\uff0c\u5176\u5c3a\u5bf8\u4e3atorch.Size([768])<\/code>\uff0c\u5bf9\u5e94\u4e8e\u6a21\u578b\u7684\u9690\u85cf\u5c42\u5927\u5c0f\u3002\u8fd9\u4e2a\u53c2\u6570\u7528\u4e8e\u6309\u5143\u7d20\u7f29\u653e\u5f52\u4e00\u5316\u540e\u7684\u503c\u3002<\/li>\n\n\n\n
              • transformer.ln_f.bias<\/code><\/strong>\uff1a\u8fd9\u662f\u5c42\u5f52\u4e00\u5316\u7684\u504f\u79fb\u53c2\u6570\uff08shift\uff09\uff0c\u5176\u5c3a\u5bf8\u4e5f\u4e3atorch.Size([768])<\/code>\u3002\u8fd9\u4e2a\u53c2\u6570\u7528\u4e8e\u6309\u5143\u7d20\u504f\u79fb\u5f52\u4e00\u5316\u540e\u7684\u503c\uff0c\u4ee5\u4fdd\u6301\u7f51\u7edc\u7684\u8868\u8fbe\u80fd\u529b\u3002<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n\n\n\n
                \u5728\u6a21\u578b\u7684\u6700\u540e\u9636\u6bb5\u4f7f\u7528\u5c42\u5f52\u4e00\u5316\u662fTransformer\u6a21\u578b\u8bbe\u8ba1\u4e2d\u7684\u5e38\u89c1\u505a\u6cd5\u3002\u5b83\u7d27\u8ddf\u5728\u6700\u540e\u4e00\u4e2aTransformer\u5c42\u7684\u8f93\u51fa\u540e\uff0c\u786e\u4fdd\u5728\u8fdb\u884c\u6700\u7ec8\u7684\u8f93\u51fa\u9884\u6d4b\u4e4b\u524d\uff0c\u6570\u636e\u5904\u4e8e\u4e00\u4e2a\u7a33\u5b9a\u7684\u5206\u5e03\u72b6\u6001\u3002<\/p>\n\n\n\n
                13.2 \u4ee3\u7801\u5b9e\u73b0\u793a\u4f8b<\/strong><\/h3>\n\n\n\n
                \u5982\u679c\u4f60\u60f3\u67e5\u770b\u6216\u8f93\u51faGPT-2\u6a21\u578b\u4e2d\u8fd9\u4e2a\u6700\u7ec8\u5c42\u5f52\u4e00\u5316\u7684\u53c2\u6570\uff0c\u53ef\u4ee5\u4f7f\u7528\u4ee5\u4e0b\u4ee3\u7801\u547d\u540d\u4e3a test11.py\uff0c\u6587\u4ef6\u4fdd\u5b58\u5230 newsrc \u76ee\u5f55\u4e0b\uff1a<\/p>\n\n\n\n
                from transformers import GPT2LMHeadModel, GPT2Tokenizer\n\n# \u52a0\u8f7d\u9884\u8bad\u7ec3\u7684GPT-2\u6a21\u578b\u548c\u5206\u8bcd\u5668\nmodel = GPT2LMHeadModel.from_pretrained('gpt2')\n\n# \u8f93\u51fa\u6700\u7ec8\u5c42\u5f52\u4e00\u5316\u7684\u53c2\u6570\nprint(\"Parameter Name: transformer.ln_f.weight   Size:\", model.transformer.ln_f.weight.size())\nprint(\"Parameter Name: transformer.ln_f.bias     Size:\", model.transformer.ln_f.bias.size())\n<\/pre><\/div>\n\n\n\n
                \u8fd9\u6bb5\u4ee3\u7801\u52a0\u8f7d\u4e86\u9884\u8bad\u7ec3\u7684GPT-2\u6a21\u578b\uff0c\u5e76\u6253\u5370\u51fa\u4e86\u6700\u7ec8\u5c42\u5f52\u4e00\u5316\u7684\u6743\u91cd\uff08weight<\/code>\uff09\u548c\u504f\u7f6e\uff08bias<\/code>\uff09\u7684\u53c2\u6570\u540d\u548c\u5c3a\u5bf8\u3002\u8fd9\u4e9b\u53c2\u6570\u5bf9\u4e8e\u6a21\u578b\u7684\u6700\u540e\u8f93\u51fa\u5904\u7406\u81f3\u5173\u91cd\u8981\uff0c\u786e\u4fdd\u4e86\u8f93\u51fa\u7684\u7a33\u5b9a\u6027\u548c\u4e00\u81f4\u6027\u3002<\/p>\n\n\n\n
                \u8fd0\u884c setup11.py<\/p>\n\n\n\n
                python newsrc\/test11.py\nParameter Name: transformer.ln_f.weight   Size: torch.Size([768])\nParameter Name: transformer.ln_f.bias     Size: torch.Size([768])<\/pre><\/div>\n\n\n\n
                \u95ee\uff1a\u6a21\u578b\u53c2\u6570\u57fa\u672c\u4e0a\u4e86\u89e3\u5b8c\u4e86\uff0c\u4e0b\u9762\u8fd8\u9700\u8981\u4e86\u89e3\u54ea\u4e9b\u77e5\u8bc6\uff1f<\/strong><\/p>\n\n\n\n
                \u4e86\u89e3\u4e86\u6a21\u578b\u7684\u53c2\u6570\u540e\uff0c\u4f60\u53ef\u4ee5\u8fdb\u4e00\u6b65\u62d3\u5c55\u4f60\u7684\u77e5\u8bc6\u548c\u6280\u80fd\uff0c\u4f7f\u81ea\u5df1\u80fd\u591f\u66f4\u6df1\u5165\u5730\u4f7f\u7528\u548c\u4f18\u5316Transformer\u6a21\u578b\uff0c\u5982GPT-2\u3002\u4ee5\u4e0b\u662f\u4e00\u4e9b\u53ef\u80fd\u7684\u4e0b\u4e00\u6b65\u5b66\u4e60\u65b9\u5411\uff1a<\/p>\n\n\n\n
                  \n
                1. \u6a21\u578b\u8bad\u7ec3\u548c\u5fae\u8c03<\/strong>\uff1a\u4e86\u89e3\u5982\u4f55\u5728\u7279\u5b9a\u4efb\u52a1\u4e0a\u8bad\u7ec3\u6216\u5fae\u8c03\u9884\u8bad\u7ec3\u6a21\u578b\u3002\u8fd9\u5305\u62ec\u8bbe\u7f6e\u9002\u5f53\u7684\u635f\u5931\u51fd\u6570\u3001\u9009\u62e9\u4f18\u5316\u5668\u3001\u8c03\u6574\u5b66\u4e60\u7387\u7b49\u3002<\/li>\n\n\n\n
                2. \u6a21\u578b\u8bc4\u4f30<\/strong>\uff1a\u5b66\u4e60\u5982\u4f55\u8bc4\u4f30\u6a21\u578b\u7684\u6027\u80fd\u3002\u8fd9\u6d89\u53ca\u5230\u9009\u62e9\u5408\u9002\u7684\u8bc4\u4f30\u6307\u6807\uff08\u5982\u51c6\u786e\u7387\u3001\u53ec\u56de\u7387\u3001F1\u5206\u6570\u7b49\uff09\u548c\u9a8c\u8bc1\u65b9\u6cd5\uff08\u5982\u4ea4\u53c9\u9a8c\u8bc1\uff09\u3002<\/li>\n\n\n\n
                3. \u6570\u636e\u5904\u7406\u548c\u589e\u5f3a<\/strong>\uff1a\u638c\u63e1\u5982\u4f55\u9884\u5904\u7406\u548c\u589e\u5f3a\u6570\u636e\u4ee5\u63d0\u9ad8\u6a21\u578b\u6027\u80fd\u3002\u8fd9\u5305\u62ec\u6587\u672c\u6e05\u6d17\u3001\u5206\u8bcd\u3001\u4f7f\u7528\u6570\u636e\u589e\u5f3a\u6280\u672f\u7b49\u3002<\/li>\n\n\n\n
                4. \u9ad8\u7ea7\u6a21\u578b\u67b6\u6784<\/strong>\uff1a\u63a2\u7d22\u66f4\u590d\u6742\u7684Transformer\u6a21\u578b\uff0c\u5982GPT-3\u3001BERT\u3001Transformer-XL\u7b49\uff0c\u4ee5\u53ca\u5b83\u4eec\u5728\u4e0d\u540cNLP\u4efb\u52a1\u4e0a\u7684\u5e94\u7528\u3002<\/li>\n\n\n\n
                5. \u6a21\u578b\u89e3\u91ca\u6027<\/strong>\uff1a\u5b66\u4e60\u5982\u4f55\u89e3\u91ca\u6a21\u578b\u7684\u51b3\u7b56\u8fc7\u7a0b\u3002\u8fd9\u5bf9\u4e8e\u63d0\u9ad8\u6a21\u578b\u7684\u900f\u660e\u5ea6\u548c\u53ef\u4fe1\u5ea6\u975e\u5e38\u91cd\u8981\uff0c\u7279\u522b\u662f\u5728\u654f\u611f\u548c\u5173\u952e\u7684\u5e94\u7528\u9886\u57df\u3002<\/li>\n\n\n\n
                6. \u90e8\u7f72\u548c\u5e94\u7528<\/strong>\uff1a\u4e86\u89e3\u5982\u4f55\u5c06\u8bad\u7ec3\u597d\u7684\u6a21\u578b\u90e8\u7f72\u5230\u751f\u4ea7\u73af\u5883\u4e2d\u3002\u8fd9\u5305\u62ec\u6a21\u578b\u538b\u7f29\u3001\u52a0\u901f\u548c\u90e8\u7f72\u7b56\u7565\uff0c\u4ee5\u53ca\u4f7f\u7528API\u7b49\u65b9\u5f0f\u63d0\u4f9b\u6a21\u578b\u670d\u52a1\u3002<\/li>\n\n\n\n
                7. \u6700\u65b0\u7814\u7a76\u548c\u8fdb\u5c55<\/strong>\uff1a\u8ddf\u8e2a\u6700\u65b0\u7684\u7814\u7a76\u8bba\u6587\u548c\u6280\u672f\u535a\u5ba2\uff0c\u4e86\u89e3NLP\u548c\u6df1\u5ea6\u5b66\u4e60\u9886\u57df\u7684\u6700\u65b0\u8fdb\u5c55\u548c\u8d8b\u52bf\u3002<\/li>\n\n\n\n
                8. \u5b9e\u8df5\u9879\u76ee<\/strong>\uff1a\u901a\u8fc7\u5b9e\u9645\u9879\u76ee\u5e94\u7528\u6240\u5b66\u77e5\u8bc6\u3002\u9009\u62e9\u4e00\u4e2a\u611f\u5174\u8da3\u7684\u95ee\u9898\u6216\u6570\u636e\u96c6\uff0c\u5c1d\u8bd5\u4f7f\u7528Transformer\u6a21\u578b\u6765\u89e3\u51b3\u6216\u5206\u6790\u5b83\u3002<\/li>\n\n\n\n
                9. \u793e\u533a\u548c\u8d44\u6e90<\/strong>\uff1a\u53c2\u4e0e\u76f8\u5173\u7684\u5728\u7ebf\u793e\u533a\uff08\u5982GitHub\u3001Reddit\u7684r\/MachineLearning\u7b49\uff09\u548c\u8ba8\u8bba\u7ec4\uff0c\u4e0e\u5176\u4ed6\u7814\u7a76\u8005\u548c\u5f00\u53d1\u8005\u4ea4\u6d41\u7ecf\u9a8c\u548c\u6700\u4f73\u5b9e\u8df5\u3002<\/li>\n<\/ol>\n\n\n\n
                  \u6df1\u5165\u5b66\u4e60\u8fd9\u4e9b\u9886\u57df\u5c06\u5e2e\u52a9\u4f60\u66f4\u6709\u6548\u5730\u4f7f\u7528GPT-2\u53ca\u5176\u4ed6Transformer\u6a21\u578b\uff0c\u4e0d\u4ec5\u9650\u4e8e\u7406\u89e3\u5b83\u4eec\u7684\u5185\u90e8\u5de5\u4f5c\u673a\u5236\uff0c\u8fd8\u5305\u62ec\u80fd\u591f\u9488\u5bf9\u5177\u4f53\u95ee\u9898\u8fdb\u884c\u8c03\u4f18\u548c\u5e94\u7528\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"
                  12. Transformer\u5c42 transformer.h\u5728GPT-2\u6a21\u578b\u4e2d\u6307\u7684\u662f\u6a21\u578b\u7684\u591a\u4e2aTransfor […]<\/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":[313,289,443,442,312,43],"tags":[242,293,243],"class_list":["post-2394","post","type-post","status-publish","format-standard","hentry","category-chatgpt","category-gpt","category-llm","category-llms","category-openai","category-infoarticle","tag-chatgpt","tag-gpt","tag-openai"],"views":1942,"jetpack_sharing_enabled":true,"jetpack_featured_media_url":"","_links":{"self":[{"href":"https:\/\/www.aqwu.net\/wp\/index.php?rest_route=\/wp\/v2\/posts\/2394","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=2394"}],"version-history":[{"count":12,"href":"https:\/\/www.aqwu.net\/wp\/index.php?rest_route=\/wp\/v2\/posts\/2394\/revisions"}],"predecessor-version":[{"id":2407,"href":"https:\/\/www.aqwu.net\/wp\/index.php?rest_route=\/wp\/v2\/posts\/2394\/revisions\/2407"}],"wp:attachment":[{"href":"https:\/\/www.aqwu.net\/wp\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=2394"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.aqwu.net\/wp\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=2394"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.aqwu.net\/wp\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=2394"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}