{"id":35830,"date":"2025-01-01T15:00:13","date_gmt":"2025-01-01T07:00:13","guid":{"rendered":"https:\/\/17aitech.com\/?p=35830"},"modified":"2025-01-01T15:00:13","modified_gmt":"2025-01-01T07:00:13","slug":"%e3%80%8apython%e6%9c%ba%e5%99%a8%e5%ad%a6%e4%b9%a0%e3%80%8b%e4%bd%9c%e8%80%85%e7%a7%91%e6%99%ae%e9%95%bf%e6%96%87%ef%bc%9a%e4%bb%8e%e5%a4%b4%e6%9e%84%e5%bb%ba%e7%b1%bbgpt%e6%96%87%e6%9c%ac%e5%88%86","status":"publish","type":"post","link":"https:\/\/17aitech.com\/?p=35830","title":{"rendered":"\u300aPython\u673a\u5668\u5b66\u4e60\u300b\u4f5c\u8005\u79d1\u666e\u957f\u6587\uff1a\u4ece\u5934\u6784\u5efa\u7c7bGPT\u6587\u672c\u5206\u7c7b\u5668\uff0c\u4ee3\u7801\u5f00\u6e90"},"content":{"rendered":"

\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:\u300aPython\u673a\u5668\u5b66\u4e60\u300b\u4f5c\u8005\u79d1\u666e\u957f\u6587\uff1a\u4ece\u5934\u6784\u5efa\u7c7bGPT\u6587\u672c\u5206\u7c7b\u5668\uff0c\u4ee3\u7801\u5f00\u6e90<\/a><\/p>\n

\n
\n
\n
\n

\u5b66\u8d77\u6765\u5427\uff01<\/p>\n<\/section>\n<\/blockquote>\n<\/section>\n<\/section>\n

\u8fd1\u65e5\uff0c\u673a\u5668\u5b66\u4e60\u7814\u7a76\u5458\u3001\u7545\u9500\u4e66\u300aPython \u673a\u5668\u5b66\u4e60\u300b\u4f5c\u8005 Sebastian Raschka \u53c8\u5206\u4eab\u4e86\u4e00\u7bc7\u957f\u6587\uff0c\u4e3b\u9898\u4e3a\u300a\u4ece\u5934\u5f00\u59cb\u6784\u5efa\u4e00\u4e2a GPT \u98ce\u683c\u7684 LLM \u5206\u7c7b\u5668\u300b\u3002<\/p>\n

<\/a><\/p>\n

\u6587\u7ae0\u5c55\u793a\u4e86\u5982\u4f55\u5c06\u9884\u8bad\u7ec3\u7684\u5927\u578b\u8bed\u8a00\u6a21\u578b\uff08LLM\uff09\u8f6c\u5316\u4e3a\u5f3a\u5927\u7684\u6587\u672c\u5206\u7c7b\u5668\u3002\u673a\u5668\u4e4b\u5fc3\u5bf9\u6587\u7ae0\u5185\u5bb9\u8fdb\u884c\u4e86\u4e0d\u6539\u53d8\u539f\u610f\u7684\u7f16\u8bd1\u3001\u6574\u7406\uff1a<\/p>\n

\u4e3a\u4ec0\u4e48\u8981\u5173\u6ce8\u5206\u7c7b\u5462\uff1f\u9996\u5148\uff0c\u9488\u5bf9\u5206\u7c7b\u4efb\u52a1\uff0c\u5bf9\u9884\u8bad\u7ec3\u6a21\u578b\u8fdb\u884c\u5fae\u8c03\u662f\u4e00\u4e2a\u7b80\u5355\u6709\u6548\u7684 LLM \u77e5\u8bc6\u5165\u95e8\u65b9\u5f0f\u3002\u5176\u6b21\uff0c\u6587\u672c\u5206\u7c7b\u6709\u8bb8\u591a\u5546\u4e1a\u5e94\u7528\u573a\u666f\uff0c\u6bd4\u5982\uff1a\u5783\u573e\u90ae\u4ef6\u68c0\u6d4b\u3001\u60c5\u611f\u5206\u6790\u3001\u5ba2\u6237\u53cd\u9988\u5206\u7c7b\u3001\u4e3b\u9898\u5206\u7c7b\u7b49\u7b49\u3002<\/p>\n

<\/a><\/p>\n

\u9605\u8bfb\u5b8c\u672c\u6587\uff0c\u4f60\u5c06\u627e\u5230\u4ee5\u4e0b 7 \u4e2a\u95ee\u9898\u7684\u7b54\u6848\uff1a<\/p>\n

1. \u9700\u8981\u8bad\u7ec3\u6240\u6709\u5c42\u5417\uff1f<\/p>\n

2. \u4e3a\u4ec0\u4e48\u5fae\u8c03\u6700\u540e\u4e00\u4e2a token\uff0c\u800c\u4e0d\u662f\u7b2c\u4e00\u4e2a token\uff1f<\/p>\n

3. BERT \u4e0e GPT \u5728\u6027\u80fd\u4e0a\u6709\u4f55\u6bd4\u8f83\uff1f<\/p>\n

4. \u5e94\u8be5\u7981\u7528\u56e0\u679c\u63a9\u7801\u5417\uff1f<\/p>\n

5. \u6269\u5927\u6a21\u578b\u89c4\u6a21\u4f1a\u6709\u4ec0\u4e48\u5f71\u54cd\uff1f<\/p>\n

6. LoRA \u53ef\u4ee5\u5e26\u6765\u4ec0\u4e48\u6539\u8fdb\uff1f<\/p>\n

7. Padding \u8fd8\u662f\u4e0d Padding\uff1f<\/p>\n

\u5b8c\u6574\u4ee3\u7801\u53ef\u4ee5\u4ece GitHub \u627e\u5230\uff1ahttps:\/\/github.com\/rasbt\/LLMs-from-scratch\/blob\/main\/ch06\/01_main-chapter-code\/ch06.ipynb<\/p>\n

Different categories of finetuning<\/p>\n

\u5fae\u8c03\u7684\u4e0d\u540c\u79cd\u7c7b<\/strong><\/p>\n

\u6307\u4ee4\u5fae\u8c03\u548c\u5206\u7c7b\u5fae\u8c03\u662f\u6700\u5e38\u89c1\u7684\u8bed\u8a00\u6a21\u578b\u5fae\u8c03\u65b9\u6cd5\u3002\u6307\u4ee4\u5fae\u8c03\u662f\u7528\u7279\u5b9a\u4efb\u52a1\u8bad\u7ec3\u6a21\u578b\uff0c\u63d0\u9ad8\u5b83\u7406\u89e3\u548c\u6267\u884c\u81ea\u7136\u8bed\u8a00\u63d0\u793a\u4e2d\u6240\u63cf\u8ff0\u4efb\u52a1\u7684\u80fd\u529b\uff0c\u5982\u4e0b\u56fe 1 \u6240\u793a\u3002<\/p>\n

<\/a><\/p>\n

\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0\u56fe 1\uff1a\u6307\u4ee4\u5fae\u8c03\u7684\u4e24\u79cd\u573a\u666f\u3002\u4e0a\u65b9\uff1a\u6a21\u578b\u7684\u4efb\u52a1\u662f\u5224\u65ad\u6587\u672c\u662f\u5426\u4e3a\u5783\u573e\u90ae\u4ef6\uff1b\u4e0b\u65b9\uff1a\u6a21\u578b\u7684\u4efb\u52a1\u662f\u5c06\u82f1\u6587\u53e5\u5b50\u7ffb\u8bd1\u6210\u5fb7\u8bed\u3002<\/sup><\/em><\/p>\n

\u5728\u5206\u7c7b\u5fae\u8c03\u4e2d\uff0c\u6a21\u578b\u88ab\u8bad\u7ec3\u7528\u4e8e\u8bc6\u522b\u7279\u5b9a\u7684\u7c7b\u522b\u6807\u7b7e\uff0c\u6bd4\u5982\u300c\u5783\u573e\u90ae\u4ef6\u300d\u548c\u300c\u975e\u5783\u573e\u90ae\u4ef6\u300d\u3002\u5206\u7c7b\u4efb\u52a1\u8fd8\u5305\u62ec\u4ece\u56fe\u50cf\u4e2d\u8bc6\u522b\u4e0d\u540c\u7684\u690d\u7269\u3001\u7ed9\u65b0\u95fb\u6309\u4f53\u80b2\u3001\u653f\u6cbb\u6216\u79d1\u6280\u7b49\u4e3b\u9898\u5206\u7c7b\uff0c\u4ece\u533b\u5b66\u5f71\u50cf\u4e2d\u533a\u5206\u826f\u6027\u548c\u6076\u6027\u80bf\u7624\u7b49\u7b49\u3002<\/p>\n

\u4e0d\u8fc7\u7ecf\u8fc7\u5206\u7c7b\u5fae\u8c03\u7684\u6a21\u578b\u53ea\u80fd\u5224\u65ad\u7c7b\u522b\uff0c\u4e0d\u80fd\u5bf9\u8f93\u5165\u7684\u6587\u672c\u4f5c\u51fa\u5176\u4ed6\u5224\u65ad\u3002<\/p>\n

<\/a><\/p>\n

\u56fe 2\uff1a\u4e00\u4e2a\u4f7f\u7528 LLM \u8fdb\u884c\u5783\u573e\u90ae\u4ef6\u5206\u7c7b\u7684\u793a\u4f8b\u3002\u9488\u5bf9\u5783\u573e\u90ae\u4ef6\u5206\u7c7b\u5fae\u8c03\u7684\u6a21\u578b\u5728\u8f93\u5165\u65f6\u4e0d\u9700\u8981\u989d\u5916\u7684\u6307\u4ee4\uff0c\u7136\u800c\uff0c\u4e0e\u6307\u4ee4\u5fae\u8c03\u6a21\u578b\u76f8\u6bd4\uff0c\u5b83\u7684\u56de\u7b54\u53ea\u80fd\u662f\u300c\u5783\u573e\u90ae\u4ef6\u300d\u548c\u300c\u975e\u5783\u573e\u90ae\u4ef6\u300d\u3002<\/sup><\/em><\/p>\n

\u6307\u4ee4\u5fae\u8c03\u7684\u6a21\u578b\u901a\u5e38\u80fd\u591f\u6267\u884c\u66f4\u5e7f\u6cdb\u7684\u4efb\u52a1\u3002\u6211\u4eec\u53ef\u4ee5\u5c06\u5206\u7c7b\u5fae\u8c03\u7684\u6a21\u578b\u89c6\u4e3a\u662f\u9ad8\u5ea6\u4e13\u4e1a\u5316\u7684\u6a21\u578b\uff0c\u4e00\u822c\u6765\u8bf4\uff0c\u5f00\u53d1\u4e00\u4e2a\u4e13\u7528\u6a21\u578b\u6bd4\u5f00\u53d1\u4e00\u4e2a\u5728\u5404\u79cd\u4efb\u52a1\u4e0a\u8868\u73b0\u826f\u597d\u7684\u901a\u7528\u6a21\u578b\u66f4\u5bb9\u6613\u3002<\/p>\n

\u4f7f\u7528\u9884\u8bad\u7ec3\u6743\u91cd\u521d\u59cb\u5316\u6a21\u578b<\/strong><\/p>\n

\u4e0b\u56fe\u4e2d\u5c55\u793a\u4e86\u5c06\u901a\u7528\u9884\u8bad\u7ec3 LLM \u8f6c\u53d8\u4e3a\u4e13\u95e8\u7528\u4e8e\u5206\u7c7b\u4efb\u52a1\u7684 LLM \u9700\u8981\u505a\u7684\u4fee\u6539\uff1a<\/p>\n

<\/a><\/p>\n

\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u56fe 3\uff1a\u5728\u6b64\u8df3\u8fc7\u6b65\u9aa4 1-5\uff0c\u76f4\u63a5\u8fdb\u5165\u6b65\u9aa4 6\uff08\u5c06\u5728\u4e0b\u4e00\u8282\u5f00\u59cb\uff09\u3002<\/sup><\/em><\/p>\n

\u5728\u505a\u4fee\u6539\u4e4b\u524d\uff0c\u8ba9\u6211\u4eec\u5148\u7b80\u5355\u4e86\u89e3\u4e00\u4e0b\u6b63\u5728\u4f7f\u7528\u7684\u9884\u8bad\u7ec3 LLM\u3002\u4e3a\u7b80\u4fbf\u8d77\u89c1\uff0c\u5047\u8bbe\u6211\u4eec\u8bbe\u7f6e\u4e86\u5982\u4e0b\u4ee3\u7801\u6765\u52a0\u8f7d\u8be5\u6a21\u578b\uff1a<\/p>\n

\n
model = GPTModel (BASE_CONFIG)<\/code>\r\n\r\n<\/code>load_weights_into_gpt (model, params)<\/code>\r\n\r\n<\/code>model.eval ()<\/code><\/pre>\n<\/section>\n
\u5728\u5c06\u6a21\u578b\u6743\u91cd\u52a0\u8f7d\u5230 GPT \u540e\uff0c\u4f7f\u7528\u4e0b\u5217\u6587\u672c\u751f\u6210\u7684\u51fd\u6570\u5e93\uff0c\u786e\u4fdd\u6a21\u578b\u751f\u6210\u8fde\u8d2f\u7684\u6587\u672c\uff1a<\/p>\n
\n
from chapter04 import generate_text_simple<\/code>\r\n\r\n<\/code>from chapter05 import text_to_token_ids, token_ids_to_text<\/code>\r\n\r\n<\/code>text_1 = \"Every effort moves you\"<\/code>\r\n\r\n<\/code>token_ids = generate_text_simple (<\/code>\r\n\r\n<\/code>    model=model,<\/code>\r\n\r\n<\/code>    idx=text_to_token_ids (text_1, tokenizer),<\/code>\r\n\r\n<\/code>    max_new_tokens=15,<\/code>\r\n\r\n<\/code>    context_size=BASE_CONFIG [\"context_length\"]<\/code>\r\n\r\n<\/code>)<\/code>\r\n\r\n<\/code>print (token_ids_to_text (token_ids, tokenizer))<\/code><\/pre>\n<\/section>\n
<\/p>\n
\u6839\u636e\u4ee5\u4e0b\u8f93\u51fa\uff0c\u6211\u4eec\u53ef\u4ee5\u770b\u5230\u6a21\u578b\u751f\u6210\u4e86\u8fde\u8d2f\u7684\u6587\u672c\uff0c\u8fd9\u8868\u660e\u6a21\u578b\u6743\u91cd\u5df2\u6b63\u786e\u52a0\u8f7d\uff1a<\/p>\n
\n
Every effort moves you forward.<\/code>\r\n\r\n<\/code>The first step is to understand the importance of your work<\/code><\/pre>\n<\/section>\n
\u8ba9\u6211\u4eec\u5148\u770b\u770b\u6a21\u578b\u662f\u5426\u53ef\u4ee5\u901a\u8fc7\u6307\u4ee4\u5fae\u8c03\u5b8c\u6210\u5783\u573e\u90ae\u4ef6\u7684\u5206\u7c7b\uff1a<\/p>\n
\n
text_2 = (<\/code>\r\n\r\n<\/code>    \"Is the following text'spam'? Answer with 'yes' or 'no':\"<\/code>\r\n\r\n<\/code>    \"'You are a winner you have been specially\"<\/code>\r\n\r\n<\/code>    \"selected to receive $1000 cash or a $2000 award.'\"<\/code>\r\n\r\n<\/code>)<\/code>\r\n\r\n<\/code>token_ids = generate_text_simple (<\/code>\r\n\r\n<\/code>    model=model,<\/code>\r\n\r\n<\/code>    idx=text_to_token_ids (text_2, tokenizer),<\/code>\r\n\r\n<\/code>    max_new_tokens=23,<\/code>\r\n\r\n<\/code>    context_size=BASE_CONFIG [\"context_length\"]<\/code>\r\n\r\n<\/code>)<\/code>\r\n\r\n<\/code>print (token_ids_to_text (token_ids, tokenizer))<\/code><\/pre>\n<\/section>\n
\u6a21\u578b\u7684\u8f93\u51fa\u5982\u4e0b\u6240\u793a\uff1a<\/p>\n
\n
Is the following text'spam'? Answer with 'yes' or 'no': 'You are a winner you have been specially selected to receive $1000 cash or a $2000 award.'<\/code>\r\n\r\n<\/code>The following text'spam'? Answer with 'yes' or 'no': 'You are a winner<\/code><\/pre>\n<\/section>\n
\u53ef\u4ee5\u660e\u663e\u770b\u51fa\u6a21\u578b\u5728\u51c6\u786e\u9075\u5faa\u6307\u4ee4\u65b9\u9762\u9047\u5230\u4e86\u4e00\u4e9b\u6311\u6218\u3002\u8fd9\u662f\u53ef\u4ee5\u9884\u89c1\u7684\uff0c\u56e0\u4e3a\u5b83\u4ec5\u7ecf\u8fc7\u4e86\u9884\u8bad\u7ec3\uff0c\u7f3a\u4e4f\u6307\u4ee4\u5fae\u8c03\u3002<\/p>\n
\u52a0\u5165\u5206\u7c7b\u5934<\/strong><\/p>\n
\u6211\u4eec\u5c06\u539f\u59cb\u8f93\u51fa\u5c42\uff08\u8fd9\u5c42\u7684\u529f\u80fd\u662f\u5c06\u6a21\u578b\u5185\u90e8\u751f\u6210\u7684\u9690\u85cf\u8868\u793a\u8f6c\u6362\u4e3a\u4e00\u4e2a\u5305\u542b 50,257 \u4e2a tokens \u7684\u8bcd\u8868\uff09\u66ff\u6362\u4e3a\u4e00\u4e2a\u8f83\u5c0f\u7684\u8f93\u51fa\u5c42\uff0c\u8be5\u5c42\u6620\u5c04\u5230\u4e24\u4e2a\u7c7b\u522b\uff1a0\uff08\u975e\u5783\u573e\u90ae\u4ef6\uff09\u548c 1\uff08\u5783\u573e\u90ae\u4ef6\uff09\uff0c\u5982\u4e0b\u56fe 4 \u6240\u793a\u3002<\/p>\n
<\/a><\/p>\n
\u56fe 4\uff1a\u6b64\u56fe\u5c55\u793a\u4e86\u5982\u4f55\u901a\u8fc7\u6539\u53d8\u67b6\u6784\u5c06 GPT \u6a21\u578b\u9002\u914d\u4e3a\u5783\u573e\u90ae\u4ef6\u5206\u7c7b\u3002\u6700\u521d\uff0c\u6a21\u578b\u7684\u7ebf\u6027\u8f93\u51fa\u5c42\u5c06 768 \u4e2a\u9690\u85cf\u5355\u5143\u6620\u5c04\u5230\u4e00\u4e2a\u5305\u542b 50,257 \u4e2a tokens \u7684\u8bcd\u6c47\u8868\u3002\u4e3a\u4e86\u8fdb\u884c\u5783\u573e\u90ae\u4ef6\u68c0\u6d4b\uff0c\u8fd9\u4e00\u5c42\u88ab\u66ff\u6362\u4e3a\u4e00\u4e2a\u65b0\u7684\u8f93\u51fa\u5c42\uff0c\u8be5\u5c42\u5c06\u76f8\u540c\u7684 768 \u4e2a\u9690\u85cf\u5355\u5143\u6620\u5c04\u5230\u4e24\u4e2a\u7c7b\u522b\uff0c\u5206\u522b\u8868\u793a\u300c\u5783\u573e\u90ae\u4ef6\u300d\u548c\u300c\u975e\u5783\u573e\u90ae\u4ef6\u300d\u3002<\/sup><\/em><\/p>\n
\u8f93\u51fa\u5c42\u8282\u70b9<\/strong><\/p>\n
\u4ece\u6280\u672f\u4e0a\u8bb2\uff0c\u56e0\u4e3a\u8fd9\u662f\u4e00\u4e2a\u4e8c\u5143\u5206\u7c7b\u4efb\u52a1\uff0c\u53ef\u4ee5\u53ea\u7528\u4e00\u4e2a\u8f93\u51fa\u8282\u70b9\u3002\u7136\u800c\uff0c\u8fd9\u5c06\u9700\u8981\u4fee\u6539\u635f\u5931\u51fd\u6570\u3002\u56e0\u6b64\uff0c\u6211\u4eec\u9009\u62e9\u4e00\u79cd\u66f4\u901a\u7528\u7684\u65b9\u6cd5\uff0c\u5339\u914d\u8f93\u51fa\u8282\u70b9\u4e0e\u5206\u7c7b\u7684\u6570\u91cf\u3002\u4f8b\u5982\uff0c\u5bf9\u4e8e\u4e00\u4e2a\u5206\u4e09\u7c7b\u7684\u95ee\u9898\uff0c\u5982\u5c06\u65b0\u95fb\u6587\u7ae0\u5206\u7c7b\u4e3a\u300c\u79d1\u6280\u300d\u3001\u300c\u4f53\u80b2\u300d\u6216\u300c\u653f\u6cbb\u300d\uff0c\u4f7f\u7528\u4e09\u4e2a\u8f93\u51fa\u8282\u70b9\uff0c\u4f9d\u6b64\u7c7b\u63a8\u3002<\/p>\n
\u5728\u5c1d\u8bd5\u8fdb\u884c\u56fe 4 \u4e2d\u6240\u793a\u7684\u4fee\u6539\u4e4b\u524d\uff0c\u5148\u901a\u8fc7 print (model) \u8f93\u51fa\u6a21\u578b\u67b6\u6784\uff1a<\/p>\n
\n
GPTModel (<\/code>\r\n\r\n<\/code>  (tok_emb): Embedding (50257, 768)<\/code>\r\n\r\n<\/code>  (pos_emb): Embedding (1024, 768)<\/code>\r\n\r\n<\/code>  (drop_emb): Dropout (p=0.0, inplace=False)<\/code>\r\n\r\n<\/code>  (trf_blocks): Sequential (<\/code>\r\n\r\n<\/code>...<\/code>\r\n\r\n<\/code>    (11): TransformerBlock (<\/code>\r\n\r\n<\/code>      (att): MultiHeadAttention (<\/code>\r\n\r\n<\/code>        (W_query): Linear (in_features=768, out_features=768, bias=True)<\/code>\r\n\r\n<\/code>        (W_key): Linear (in_features=768, out_features=768, bias=True)<\/code>\r\n\r\n<\/code>        (W_value): Linear (in_features=768, out_features=768, bias=True)<\/code>\r\n\r\n<\/code>        (out_proj): Linear (in_features=768, out_features=768, bias=True)<\/code>\r\n\r\n<\/code>        (dropout): Dropout (p=0.0, inplace=False)<\/code>\r\n\r\n<\/code>      )<\/code>\r\n\r\n<\/code>      (ff): FeedForward (<\/code>\r\n\r\n<\/code>        (layers): Sequential (<\/code>\r\n\r\n<\/code>          (0): Linear (in_features=768, out_features=3072, bias=True)<\/code>\r\n\r\n<\/code>          (1): GELU ()<\/code>\r\n\r\n<\/code>          (2): Linear (in_features=3072, out_features=768, bias=True)<\/code>\r\n\r\n<\/code>        )<\/code>\r\n\r\n<\/code>      )<\/code>\r\n\r\n<\/code>      (norm1): LayerNorm ()<\/code>\r\n\r\n<\/code>      (norm2): LayerNorm ()<\/code>\r\n\r\n<\/code>      (drop_resid): Dropout (p=0.0, inplace=False)<\/code>\r\n\r\n<\/code>    )<\/code>\r\n\r\n<\/code>  )<\/code>\r\n\r\n<\/code>  (final_norm): LayerNorm ()<\/code>\r\n\r\n<\/code>  (out_head): Linear (in_features=768, out_features=50257, bias=False)<\/code>\r\n\r\n<\/code>)<\/code><\/pre>\n<\/section>\n
\u5982\u4e0a\u6240\u793a\uff0cGPTModel \u7531\u5d4c\u5165\u5c42\u548c 12 \u4e2a\u76f8\u540c\u7684 transformer \u5757\u7ec4\u6210\uff0c\u4e3a\u7b80\u6d01\u8d77\u89c1\uff0c\u4ec5\u663e\u793a\u6700\u540e\u4e00\u4e2a\u5757\uff0c\u7136\u540e\u662f\u6700\u7ec8\u7684 LayerNorm \u548c\u8f93\u51fa\u5c42 out_head\u3002<\/p>\n
\u63a5\u4e0b\u6765\uff0c\u6211\u4eec\u5c06 out_head \u66ff\u6362\u4e3a\u4e00\u4e2a\u65b0\u7684\u8f93\u51fa\u5c42\uff0c\u5982\u56fe 4 \u6240\u793a\uff0c\u6211\u4eec\u5c06\u5bf9\u8fd9\u4e00\u5c42\u8fdb\u884c\u5fae\u8c03\u3002<\/p>\n
\u9009\u62e9\u5fae\u8c03\u7279\u5b9a\u5c42\u4e0e\u5fae\u8c03\u6240\u6709\u5c42<\/strong><\/p>\n
\u6211\u4eec\u4e0d\u5fc5\u5bf9\u6a21\u578b\u6bcf\u4e00\u5c42\u8fdb\u884c\u5fae\u8c03\uff0c\u56e0\u4e3a\u795e\u7ecf\u7f51\u7edc\u7684\u8f83\u4f4e\u5c42\u6355\u6349\u5230\u7684\u57fa\u672c\u7684\u8bed\u8a00\u7ed3\u6784\u548c\u8bed\u4e49\u662f\u901a\u7528\u7684\uff0c\u53ef\u4ee5\u5728\u8bb8\u591a\u4e0d\u540c\u7684\u4efb\u52a1\u548c\u6570\u636e\u96c6\u4e2d\u53d1\u6325\u4f5c\u7528\u3002<\/p>\n
\u56e0\u6b64\uff0c\u6211\u4eec\u4ec5\u5fae\u8c03\u6700\u540e\u51e0\u5c42\uff08\u9760\u8fd1\u8f93\u51fa\u7684\u5c42\uff09\u5c31\u591f\u4e86\uff0c\u8fd9\u4e9b\u5c42\u66f4\u5177\u4f53\u4e8e\u7ec6\u5fae\u7684\u8bed\u8a00\u6a21\u5f0f\u548c\u4efb\u52a1\u7279\u5f81\u3002\u8fd9\u79cd\u65b9\u6cd5\u5728\u8ba1\u7b97\u4e0a\u4e5f\u5c06\u66f4\u52a0\u9ad8\u6548\u3002<\/p>\n
\u4e3a\u4e86\u51c6\u5907\u8fdb\u884c\u5206\u7c7b\u5fae\u8c03\uff0c\u9996\u5148\u6211\u4eec\u51bb\u7ed3\u6a21\u578b\uff0c\u5373\u5c06\u6240\u6709\u5c42\u8bbe\u7f6e\u4e3a\u4e0d\u53ef\u8bad\u7ec3\uff1a<\/p>\n
\n
for param in model.parameters ():<\/code>\r\n\r\n<\/code>\u00a0 \u00a0 param.requires_grad = False<\/code><\/pre>\n<\/section>\n
\u7136\u540e\uff0c\u5982\u56fe 4 \u6240\u793a\uff0c\u6211\u4eec\u4fee\u6539\u8f93\u51fa\u5c42 model.out_head \uff1a<\/p>\n
\n
torch.manual_seed (123)<\/code>\r\n\r\n<\/code>num_classes = 2<\/code>\r\n\r\n<\/code>model.out_head = torch.nn.Linear (<\/code>\r\n\r\n<\/code>    in_features=BASE_CONFIG [\"emb_dim\"],<\/code>\r\n\r\n<\/code>    out_features=num_classes<\/code>\r\n\r\n<\/code>)<\/code><\/pre>\n<\/section>\n
\u6ce8\u610f\uff0c\u5728\u4e0a\u8ff0\u4ee3\u7801\u4e2d\uff0c\u6211\u4eec\u4f7f\u7528\u4e86 BASE_CONFIG [“emb_dim”]\uff0c\u5b83\u7684\u503c\u5728 \u201cgpt2-small\uff08124M\uff09\u201d \u6a21\u578b\u4e2d\u4e3a 768\u3002\u8fd9\u6837\u505a\u7684\u76ee\u7684\u662f\u4e3a\u4e86\u8ba9\u540e\u7eed\u7684\u4ee3\u7801\u66f4\u52a0\u901a\u7528\uff0c\u76f8\u540c\u7684\u4ee3\u7801\u4e5f\u80fd\u5904\u7406\u5176\u4ed6\u578b\u53f7\u7684 GPT-2 \u6a21\u578b\u3002<\/p>\n
\u65b0\u7684 model.out_head \u8f93\u51fa\u5c42\u7684 requires_grad \u5c5e\u6027\u9ed8\u8ba4\u8bbe\u7f6e\u4e3a True\uff0c\u8fd9\u610f\u5473\u7740\u8fd9\u662f\u6a21\u578b\u4e2d\u552f\u4e00\u4f1a\u5728\u8bad\u7ec3\u671f\u95f4\u66f4\u65b0\u7684\u5c42\u3002<\/p>\n
\u4ece\u6280\u672f\u4e0a\u8bb2\uff0c\u53ea\u8bad\u7ec3\u521a\u521a\u6dfb\u52a0\u7684\u8f93\u51fa\u5c42\u5c31\u8db3\u591f\u4e86\u3002\u7136\u800c\uff0c\u6211\u5728\u5b9e\u9a8c\u4e2d\u53d1\u73b0\uff0c\u5fae\u8c03\u989d\u5916\u7684\u5c42\uff0c\u53ef\u4ee5\u663e\u8457\u63d0\u9ad8\u5fae\u8c03\u6a21\u578b\u7684\u9884\u6d4b\u6027\u80fd\u3002<\/p>\n
\u6b64\u5916\uff0c\u6211\u4eec\u5c06\u6700\u540e\u4e00\u4e2a transformer \u5757\u4ee5\u53ca\u8fde\u63a5\u8be5\u5757\u4e0e\u8f93\u51fa\u5c42\u7684 LayerNorm \u6a21\u5757\u8bbe\u7f6e\u4e3a\u53ef\u8bad\u7ec3\uff0c\u5982\u56fe 5 \u6240\u793a\u3002<\/p>\n
<\/a><\/p>\n
\u56fe 5\uff1a\u7528\u6211\u7684\u6b65\u9aa4\u5f00\u53d1\u7684 GPT \u6a21\u578b\u5305\u542b 12 \u4e2a\u91cd\u590d\u7684 transformer \u5757\u3002\u9664\u4e86\u8f93\u51fa\u5c42\uff0c\u6211\u4eec\u5c06\u6700\u540e\u7684 LayerNorm \u548c\u6700\u540e\u4e00\u4e2a transformer \u5757\u8bbe\u7f6e\u4e3a\u53ef\u8bad\u7ec3\uff0c\u800c\u5176\u4f59 11 \u4e2a transformer \u5757\u548c\u5d4c\u5165\u5c42\u4fdd\u6301\u4e3a\u4e0d\u53ef\u8bad\u7ec3\u3002<\/sup><\/em><\/p>\n
\u4e3a\u4e86\u505a\u5230\u8fd9\u70b9\uff0c\u6211\u4eec\u5c06\u5b83\u4eec\u5404\u81ea\u7684 requires_grad \u8bbe\u7f6e\u4e3a True\uff1a<\/p>\n
\n
for param in model.trf_blocks [-1].parameters ():<\/code>\r\n\r\n<\/code>    param.requires_grad = True<\/code>\r\n\r\n<\/code>for param in model.final_norm.parameters ():<\/code>\r\n\r\n<\/code>\u00a0 \u00a0 param.requires_grad = True<\/code><\/pre>\n<\/section>\n
\u5c3d\u7ba1\u6211\u4eec\u6dfb\u52a0\u4e86\u4e00\u4e2a\u65b0\u7684\u8f93\u51fa\u5c42\uff0c\u5e76\u5c06\u67d0\u4e9b\u5c42\u8bbe\u7f6e\u4e3a\u4e0d\u53ef\u8bad\u7ec3\uff0c\u6211\u4eec\u4ecd\u7136\u53ef\u4ee5\u4f7f\u7528\u8fd9\u4e2a\u6a21\u578b\u3002\u4f8b\u5982\uff0c\u6211\u4eec\u53ef\u4ee5\u50cf\u4e4b\u524d\u90a3\u6837\u8f93\u5165\u4e00\u6bb5\u793a\u4f8b\u6587\u672c\uff1a<\/p>\n
\n
inputs = tokenizer.encode (\"Do you have time\")<\/code>\r\n\r\n<\/code>inputs = torch.tensor (inputs).unsqueeze (0)<\/code>\r\n\r\n<\/code>print (\"Inputs:\", inputs)<\/code>\r\n\r\n<\/code>print (\"Inputs dimensions:\", inputs.shape)<\/code><\/pre>\n<\/section>\n
\u5982\u8f93\u51fa\u6240\u793a\uff0c\u4e0a\u8ff0\u4ee3\u7801\u5c06\u8f93\u5165\u7f16\u7801\u4e3a\u4e00\u4e2a\u5305\u542b 4 \u4e2a\u8f93\u5165 tokens \u7684\u5f20\u91cf\uff1a<\/p>\n
\n
Inputs: tensor ([[5211,  345,  423,  640]])<\/code>\r\n\r\n<\/code>Inputs dimensions: torch.Size ([1, 4])<\/code><\/pre>\n<\/section>\n
\u7136\u540e\uff0c\u6211\u4eec\u5c06\u7f16\u7801\u540e\u7684 token IDs \u8f93\u5165\u6a21\u578b\uff1a<\/p>\n
\n
with torch.no_grad ():<\/code>\r\n\r\n<\/code>    outputs = model (inputs)<\/code>\r\n\r\n<\/code>print (\"Outputs:n\", outputs)<\/code>\r\n\r\n<\/code>print (\"Outputs dimensions:\", outputs.shape)<\/code><\/pre>\n<\/section>\n
\u8f93\u51fa\u5f20\u91cf\u5982\u4e0b\u6240\u793a\uff1a<\/p>\n
\n
Outputs:<\/code>\r\n\r\n<\/code> tensor ([[[-1.5854,  0.9904],<\/code>\r\n\r\n<\/code>          [-3.7235,  7.4548],<\/code>\r\n\r\n<\/code>          [-2.2661,  6.6049],<\/code>\r\n\r\n<\/code>          [-3.5983,  3.9902]]])<\/code>\r\n\r\n<\/code>Outputs dimensions: torch.Size ([1, 4, 2])<\/code><\/pre>\n<\/section>\n
\u6a21\u578b\u5c06\u8f93\u51fa\u4e00\u4e2a [1, 4, 50257] \u7684\u8f93\u51fa\u5f20\u91cf\uff0c\u5176\u4e2d 50,257 \u4ee3\u8868\u8bcd\u6c47\u8868\u7684\u5927\u5c0f\u3002\u8f93\u51fa\u884c\u6570\u5bf9\u5e94\u4e8e\u8f93\u5165\u6807\u8bb0\u7684\u6570\u91cf\uff08\u5728\u672c\u4f8b\u4e2d\u662f 4\uff09\u3002\u6bcf\u4e2a\u8f93\u51fa\u7684\u5d4c\u5165\u7ef4\u5ea6\uff08\u5217\u6570\uff09\u73b0\u5728\u51cf\u5c11\u5230 2\uff0c\u800c\u4e0d\u662f 50,257\uff0c\u56e0\u4e3a\u6211\u4eec\u66ff\u6362\u4e86\u6a21\u578b\u7684\u8f93\u51fa\u5c42\u3002<\/p>\n
\u7531\u4e8e\u6211\u4eec\u7684\u4e3b\u8981\u76ee\u6807\u662f\u5fae\u8c03\u51fa\u66f4\u64c5\u957f\u5bf9\u5783\u573e\u90ae\u4ef6\u8fdb\u884c\u5206\u7c7b\u7684\u6a21\u578b\u3002\u4e3a\u4e86\u5b9e\u73b0\u8fd9\u4e00\u70b9\uff0c\u6211\u4eec\u4e0d\u9700\u8981\u5bf9\u6240\u6709\u884c\u8fdb\u884c\u5fae\u8c03\uff0c\u53ef\u4ee5\u4e13\u6ce8\u4e8e\u4e00\u4e2a\u5355\u4e00\u7684\u8f93\u51fa token\u3002\u5177\u4f53\u6765\u8bf4\uff0c\u6211\u4eec\u5c06\u4e13\u6ce8\u4e8e\u6700\u540e\u4e00\u884c\uff0c\u5bf9\u5e94\u7684\u6700\u540e\u4e00\u4e2a\u8f93\u51fa token\uff0c\u5982\u56fe 6 \u6240\u793a\u3002<\/p>\n
<\/a><\/p>\n
\u56fe 6: \u672c\u56fe\u5c55\u793a\u4e86 GPT \u6a21\u578b\u5904\u7406\u4e00\u4e2a\u5305\u542b 4 \u4e2a token \u7684\u8f93\u5165\u793a\u4f8b\uff0c\u5e76\u751f\u6210\u76f8\u5e94\u8f93\u51fa\u7684\u8be6\u7ec6\u8fc7\u7a0b\u3002\u6a21\u578b\u7684\u8f93\u51fa\u5c42\u7ecf\u8fc7\u8c03\u6574\uff0c\u8f93\u51fa\u5f20\u91cf\u4ec5\u5305\u542b 2 \u5217\uff0c\u4e3a\u4e86\u5b8c\u6210\u5206\u7c7b\u5fae\u8c03\uff0c\u6211\u4eec\u4e13\u6ce8\u4e8e\u8f93\u51fa\u7684\u6700\u540e\u4e00\u884c\uff0c\u5bf9\u5e94\u7684\u6700\u540e\u4e00\u4e2a token\u3002<\/sup><\/em><\/p>\n
\u53ef\u4ee5\u4f7f\u7528\u4ee5\u4e0b\u4ee3\u7801\u4ece\u8f93\u51fa\u5f20\u91cf\u4e2d\u63d0\u53d6\u6700\u540e\u4e00\u4e2a\u8f93\u51fa token\uff1a<\/p>\n
\n
print (\"Last output token:\", outputs [:, -1, :])<\/code><\/pre>\n<\/section>\n
Print \u51fa\u6765\u7ed3\u679c\u5982\u4e0b\uff1a<\/p>\n
\n
Last output token: tensor([[-3.5983, \u00a03.9902]])<\/code><\/pre>\n<\/section>\n
\u90a3\u4e48\uff0c\u6211\u4eec\u4e3a\u4ec0\u4e48\u8981\u9009\u62e9\u6700\u540e\u4e00\u4e2a token\uff0c\u800c\u4e0d\u662f\u5176\u4ed6\u4f4d\u7f6e\u4e0a\u7684 token \u5462\uff1f<\/p>\n
\u6ce8\u610f\u529b\u673a\u5236\u5efa\u7acb\u4e86\u6bcf\u4e2a\u8f93\u5165 token \u4e0e\u5176\u4ed6 token \u4e4b\u95f4\u7684\u5173\u7cfb\uff0c\u4e3a\u4e86\u8ba9\u300c\u6ce8\u610f\u529b\u300d\u96c6\u4e2d\uff0c\u9700\u8981\u7528\u5230\u56e0\u679c\u6ce8\u610f\u529b\u63a9\u7801\u3002\u5b83\u7684\u539f\u7406\u662f\u9650\u5236\u6bcf\u4e2a token \u53ea\u5173\u6ce8\u81ea\u5df1\u548c\u524d\u9762\u7684 token\uff0c\u5982\u4e0b\u56fe 7 \u6240\u793a\uff1a<\/p>\n
<\/a><\/p>\n
\u56fe 7\uff1a\u56e0\u679c\u6ce8\u610f\u529b\u673a\u5236\uff0c\u77e9\u9635\u663e\u793a\u4e86\u6bcf\u4e2a\u8f93\u5165 token \u4e4b\u95f4\u7684\u6ce8\u610f\u529b\u5f97\u5206\u3002\u7a7a\u767d\u5355\u5143\u683c\u8868\u793a\u88ab\u63a9\u7801\u5c4f\u853d\u7684\u4f4d\u7f6e\uff0c\u9632\u6b62 token \u5173\u6ce8\u540e\u6765\u7684 token\u3002\u6700\u540e\u4e00\u4e2a token\u300ctime\u300d\u662f\u552f\u4e00\u9700\u8981\u4e3a\u6240\u6709\u4e4b\u524d\u7684 token \u8ba1\u7b97\u6ce8\u610f\u529b\u5f97\u5206\u7684 token\u3002<\/sup><\/em><\/p>\n
\u5982\u56fe\u6240\u793a\uff0c\u5e8f\u5217\u4e2d\u7684\u6700\u540e\u4e00\u4e2a token \u79ef\u7d2f\u4e86\u6700\u591a\u7684\u4fe1\u606f\uff0c\u56e0\u6b64\uff0c\u5728\u5fae\u8c03\u8fc7\u7a0b\u4e2d\uff0c\u6211\u4eec\u91cd\u70b9\u5173\u6ce8\u8fd9\u4e2a\u6700\u540e\u7684 token\u3002<\/p>\n
\u5982\u4f55\u5c06\u6700\u540e\u4e00\u4e2a token \u8f6c\u6362\u4e3a\u5206\u7c7b\u6807\u7b7e\u9884\u6d4b\uff0c\u5e76\u8ba1\u7b97\u6a21\u578b\u7684\u521d\u59cb\u9884\u6d4b\u51c6\u786e\u7387\u3002\u63a5\u4e0b\u6765\uff0c\u6211\u4eec\u5c06\u5728\u540e\u7eed\u90e8\u5206\u5fae\u8c03\u6a21\u578b\u4ee5\u5b8c\u6210\u5783\u573e\u90ae\u4ef6\u5206\u7c7b\u4efb\u52a1\u3002<\/p>\n
\u8bc4\u4f30\u6a21\u578b\u6027\u80fd<\/strong><\/p>\n
\u7531\u4e8e\u8fd9\u90e8\u5206\u5185\u5bb9\u5df2\u7ecf\u5f88\u957f\uff0c\u6211\u5c31\u4e0d\u8be6\u7ec6\u8ba8\u8bba\u6a21\u578b\u8bc4\u4f30\u7684\u7ec6\u8282\u4e86\u3002\u4e0d\u8fc7\uff0c\u6211\u60f3\u81f3\u5c11\u5206\u4eab\u4e00\u5f20\u56fe\uff0c\u5c55\u793a\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\uff0c\u6a21\u578b\u8bad\u7ec3\u96c6\u548c\u9a8c\u8bc1\u96c6\u7684\u5206\u7c7b\u51c6\u786e\u7387\uff0c\u4ee5\u5c55\u793a\u6a21\u578b\u786e\u5b9e\u5b66\u5f97\u5f88\u597d\u3002<\/p>\n
<\/a><\/p>\n
\u56fe 8\uff1a\u8bad\u7ec3\u51c6\u786e\u7387\uff08\u5b9e\u7ebf\uff09\u548c\u9a8c\u8bc1\u51c6\u786e\u7387\uff08\u865a\u7ebf\uff09\u5728\u65e9\u671f\u7684\u8bad\u7ec3\u5468\u671f\u4e2d\u5927\u5e45\u4e0a\u5347\uff0c\u7136\u540e\u8d8b\u4e8e\u5e73\u7a33\uff0c\u8fbe\u5230\u4e86\u51e0\u4e4e\u5b8c\u7f8e\u7684\u51c6\u786e\u7387 1.0\uff0c\u5bf9\u5e94 100%\u3002\u4e24\u6761\u7ebf\u5728\u6574\u4e2a\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u76f8\u8ddd\u8f83\u8fd1\uff0c\u8868\u660e\u6a21\u578b\u5bf9\u8bad\u7ec3\u6570\u636e\u5e76\u6ca1\u6709\u8fc7\u5ea6\u62df\u5408\u3002<\/sup><\/em><\/p>\n
\u6a21\u578b\u7684\u9a8c\u8bc1\u51c6\u786e\u7387\u7ea6\u4e3a 97%\u3002\u6d4b\u8bd5\u51c6\u786e\u7387\u7ea6\u4e3a 96%\u3002\u6b64\u5916\uff0c\u6211\u4eec\u53ef\u4ee5\u770b\u5230\u6a21\u578b\u7565\u5fae\u6709\u4e00\u70b9\u70b9\u8fc7\u62df\u5408\uff0c\u56e0\u4e3a\u8bad\u7ec3\u96c6\u7684\u51c6\u786e\u7387\u7a0d\u9ad8\u3002<\/p>\n
\u4ece\u8865\u5145\u5b9e\u9a8c\u5f97\u51fa\u7684\u6d1e\u89c1<\/strong><\/p>\n
\u5230\u8fd9\u91cc\uff0c\u4f60\u53ef\u80fd\u5bf9\u67d0\u4e9b\u8bbe\u8ba1\u9009\u62e9\u6709\u5f88\u591a\u7591\u95ee\uff0c\u6240\u4ee5\u6211\u8fdb\u884c\u4e86\u4e00\u4e9b\u8865\u5145\u5b9e\u9a8c\u5e76\u628a\u7ed3\u679c\u5206\u4eab\u4e86\u51fa\u6765\u3002\u91cd\u65b0\u8fd0\u884c\u8fd9\u4e9b\u5b9e\u9a8c\u7684\u4ee3\u7801\u5df2\u7ecf\u653e\u5728\u4e86\u4ee5\u4e0b GitHub \u9879\u76ee\u4e2d\u3002<\/p>\n
GitHub \u5730\u5740\uff1ahttps:\/\/github.com\/rasbt\/LLMs-from-scratch\/tree\/main\/ch06\/02_bonus_additional-experiments<\/p>\n
\u9700\u8981\u8bad\u7ec3\u6240\u6709\u5c42\u5417\uff1f<\/strong><\/p>\n
\u51fa\u4e8e\u6548\u7387\u539f\u56e0\uff0c\u6211\u4eec\u4ec5\u8bad\u7ec3\u8f93\u51fa\u5c42\u548c\u6700\u540e\u4e00\u4e2a transformer \u5757\u3002\u5982\u524d\u6240\u8ff0\uff0c\u5bf9\u4e8e\u5206\u7c7b\u5fae\u8c03\uff0c\u65e0\u9700\u66f4\u65b0 LLM \u4e2d\u7684\u6240\u6709\u5c42\u3002\u6211\u4eec\u66f4\u65b0\u7684\u6743\u91cd\u8d8a\u5c11\uff0c\u8bad\u7ec3\u901f\u5ea6\u5c31\u8d8a\u5feb\uff0c\u56e0\u4e3a\u6211\u4eec\u4e0d\u9700\u8981\u5728\u53cd\u5411\u4f20\u64ad\u671f\u95f4\u8ba1\u7b97\u6743\u91cd\u7684\u68af\u5ea6\u3002<\/p>\n
\u4f46\u662f\uff0c\u4f60\u53ef\u80fd\u60f3\u77e5\u9053\u5982\u679c\u4e0d\u66f4\u65b0\u6240\u6709\u5c42\uff0c\u6211\u4eec\u4f1a\u7559\u4e0b\u591a\u5c11\u9884\u6d4b\u6027\u80fd\u3002\u56e0\u6b64\uff0c\u5728\u4e0b\u8868\u4e2d\uff0c\u6211\u5bf9\u6240\u6709\u5c42\u3001\u4ec5\u6700\u540e\u4e00\u4e2a transformer \u5757\uff08\u5305\u62ec\u6700\u540e\u4e00\u5c42\uff09\u3001\u4ec5\u6700\u540e\u4e00\u5c42\u8fdb\u884c\u4e86\u5fae\u8c03\u3002<\/p>\n
<\/a><\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u8868 1\uff1a\u8bad\u7ec3\u6240\u6709\u5c42 vs \u4ec5\u8bad\u7ec3\u6700\u540e\u4e00\u4e2a Transformer \u5757\uff08\u5305\u62ec\u6700\u540e\u4e00\u5c42\uff09vs \u4ec5\u8bad\u7ec3\u6700\u540e\u4e00\u5c42<\/sup><\/em><\/p>\n
\u5982\u4e0a\u8868 1 \u6240\u793a\uff0c\u8bad\u7ec3\u6240\u6709\u5c42\u7684\u6027\u80fd\u7a0d\u597d\u4e00\u4e9b\uff1a96.67% vs 95.00%\u3002\u4e0d\u8fc7\uff0c\u8fd9\u4f7f\u8fd0\u884c\u65f6\u95f4\u589e\u52a0\u4e86\u7ea6 2.5 \u500d\u3002<\/p>\n
\u4e3a\u4ec0\u4e48\u8981\u5fae\u8c03\u6700\u540e\u4e00\u4e2a token\uff0c\u800c\u4e0d\u662f\u7b2c\u4e00\u4e2a token\uff1f<\/strong><\/p>\n
\u5982\u679c\u4f60\u719f\u6089 BERT\uff08Devlin et al. 2018\uff09\u7b49\u7f16\u7801\u5668\u5f0f\u8bed\u8a00\u6a21\u578b\uff0c\u4f60\u53ef\u80fd\u77e5\u9053\u8fd9\u4e9b\u6a21\u578b\u6709\u4e00\u4e2a\u6307\u5b9a\u7684\u5206\u7c7b token \u4f5c\u4e3a\u5176\u7b2c\u4e00\u4e2a token\uff0c\u5982\u4e0b\u56fe\u6240\u793a\uff1a<\/p>\n
<\/a><\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0\u56fe\u6765\u81ea BERT \u539f\u59cb\u8bba\u6587\uff1ahttps:\/\/arxiv.org\/abs\/1810.04805<\/sup><\/em><\/p>\n
\u4e0e BERT \u76f8\u6bd4\uff0cGPT \u662f\u4e00\u79cd\u5177\u6709\u56e0\u679c\u6ce8\u610f\u529b\u63a9\u7801\u7684\u89e3\u7801\u5668\u5f0f\u6a21\u578b\uff08\u5982\u56fe 7 \u6240\u793a\uff09\u3002\u8fd9\u610f\u5473\u7740\u7b2c\u4e00\u4e2a token \u6ca1\u6709\u8f93\u5165\u4e2d\u4efb\u4f55\u5176\u4ed6 token \u7684\u4e0a\u4e0b\u6587\u4fe1\u606f\u3002\u53ea\u6709\u6700\u540e\u4e00\u4e2a token \u5177\u6709\u6709\u5173\u6240\u6709\u5176\u4ed6 token \u7684\u4fe1\u606f\u3002<\/p>\n
\u56e0\u6b64\uff0c\u5982\u679c\u6211\u4eec\u60f3\u4f7f\u7528\u50cf GPT \u8fd9\u6837\u7684\u6a21\u578b\u8fdb\u884c\u5206\u7c7b\u5fae\u8c03\uff0c\u6211\u4eec\u5e94\u8be5\u5173\u6ce8\u6700\u540e\u4e00\u4e2a token \u6807\u8bb0\u4ee5\u6355\u83b7\u6240\u6709\u5176\u4ed6\u8f93\u5165 token \u7684\u4e0a\u4e0b\u6587\u4fe1\u606f\u3002<\/p>\n
\u5982\u4e0b\u8868\u6240\u793a\uff0c\u6211\u4eec\u53ef\u4ee5\u770b\u5230\u4f7f\u7528\u7b2c\u4e00\u4e2a token \u6765\u5fae\u8c03 GPT \u6a21\u578b\u8fdb\u884c\u5206\u7c7b\u4f1a\u5bfc\u81f4\u6027\u80fd\u66f4\u5dee\u3002<\/p>\n
<\/a><\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0\u8868 2\uff1a\u5fae\u8c03 GPT \u6a21\u578b\u4e2d\u7684\u6700\u540e\u4e00\u4e2a token \u4e0e\u7b2c\u4e00\u4e2a token\u3002<\/sup><\/em><\/p>\n
BERT \u4e0e GPT \u7684\u6027\u80fd\u6bd4\u8f83\u5982\u4f55\uff1f<\/strong><\/p>\n
\u8bf4\u5230 BERT\uff0c\u4f60\u53ef\u80fd\u60f3\u77e5\u9053\u5b83\u5728\u5206\u7c7b\u4efb\u52a1\u4e0a\u4e0e\u7c7b GPT \u6a21\u578b\u7684\u6027\u80fd\u6bd4\u8f83\u5982\u4f55\uff1f\u7b80\u5355\u6765\u8bf4\uff0c\u5728\u5783\u573e\u90ae\u4ef6\u5206\u7c7b\u4efb\u52a1\u4e0a\uff0c\u66f4\u5c0f\u7684 GPT-2\uff08124M\uff09\u4e0e\u66f4\u5927 BERT\uff08340M\uff09\u7684\u6027\u80fd\u7c7b\u4f3c\uff0c\u5177\u4f53\u5982\u4e0b\u8868 3 \u6240\u793a\u3002<\/p>\n
<\/a><\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u8868 3\uff1aGPT-2 \u4e0e BERT \u7684\u7ed3\u679c\u6bd4\u8f83\u3002<\/sup><\/em><\/p>\n
\u53ef\u4ee5\u770b\u5230\uff0cBERT \u6a21\u578b\u7684\u8868\u73b0\u6bd4 GPT-2 \u7a0d\u5fae\u597d\u4e00\u70b9\uff08\u6d4b\u8bd5\u51c6\u786e\u7387\u9ad8 1%\uff09\uff0c\u4f46 BERT \u7684\u53c2\u6570\u89c4\u6a21\u51e0\u4e4e\u662f GPT-2 \u7684 3 \u500d\u3002\u6b64\u5916\uff0c\u6570\u636e\u96c6\u53ef\u80fd\u592a\u5c0f\u4e14\u592a\u7b80\u5355\u4e86\uff0c\u56e0\u6b64\u6211\u53c8\u5728 IMDB Movie Review \u6570\u636e\u96c6\u4e0a\u5c1d\u8bd5\u6bd4\u8f83\u4e86\u60c5\u611f\u5206\u7c7b\u8868\u73b0\uff08\u5373\u9884\u6d4b\u89c2\u770b\u8005\u662f\u5426\u559c\u6b22\u4e00\u90e8\u7535\u5f71\uff09\u3002<\/p>\n
<\/a><\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u8868 4\uff1aGPT-2 \u4e0e BERT \u5728\u5f71\u8bc4\u5206\u7c7b\u4efb\u52a1\u4e0a\u7684\u6bd4\u8f83\u3002<\/sup><\/em><\/p>\n
\u53ef\u4ee5\u770b\u5230\uff0c\u5728\u8fd9\u4e2a\u66f4\u5927\u7684\u6570\u636e\u96c6\u4e0a\uff08\u5305\u542b 25k \u8bad\u7ec3\u548c 25k \u6d4b\u8bd5\u96c6\u8bb0\u5f55\uff09\uff0cGPT-2 \u4e0e BERT \u4e24\u4e2a\u6a21\u578b\u7684\u9884\u6d4b\u6027\u80fd\u540c\u6837\u7c7b\u4f3c\u3002<\/p>\n
\u603b\u7684\u6765\u8bf4\uff0c\u5728\u5206\u7c7b\u4efb\u52a1\u4e0a\uff0cBERT \u548c\u5176\u4ed6\u7f16\u7801\u5668\u98ce\u683c\u7684\u6a21\u578b\u88ab\u8ba4\u4e3a\u4f18\u4e8e\u89e3\u7801\u5668\u98ce\u683c\u7684\u6a21\u578b\u3002\u4f46\u662f\uff0c\u5b9e\u9a8c\u7ed3\u679c\u4e5f\u8868\u660e\uff0c\u7f16\u7801\u5668\u98ce\u683c\u7684 BERT \u548c\u89e3\u7801\u5668\u98ce\u683c\u7684 GPT \u6a21\u578b\u4e4b\u95f4\u6ca1\u6709\u592a\u5927\u7684\u5dee\u5f02\u3002<\/p>\n
\u6b64\u5916\uff0c\u5982\u679c\u4f60\u5bf9\u66f4\u591a\u57fa\u51c6\u6bd4\u8f83\u4ee5\u53ca\u5982\u4f55\u8fdb\u4e00\u6b65\u63d0\u5347\u89e3\u7801\u5668\u98ce\u683c\u6a21\u578b\u7684\u5206\u7c7b\u6027\u80fd\u611f\u5174\u8da3\uff0c\u53ef\u4ee5\u53c2\u9605\u4ee5\u4e0b\u4e24\u7bc7\u6700\u8fd1\u7684\u8bba\u6587\uff1a<\/p>\n