{"id":8001,"date":"2017-01-11T01:00:00","date_gmt":"2017-01-10T17:00:00","guid":{"rendered":"http:\/\/swarma.org\/%swarma-paper%\/"},"modified":"2018-12-06T08:20:11","modified_gmt":"2018-12-06T00:20:11","slug":"%e5%85%b3%e4%ba%8e%e6%b7%b1%e5%ba%a6%e7%a5%9e%e7%bb%8f%e7%bd%91%e7%bb%9c%e5%8e%8b%e7%bc%a9%e4%b8%8b-%e5%b0%8f%e5%a3%b0%e5%98%9f%e5%9b%94%e4%b8%93%e6%a0%8f","status":"publish","type":"post","link":"https:\/\/swarma.org\/?p=8001","title":{"rendered":"\u5173\u4e8e\u6df1\u5ea6\u795e\u7ecf\u7f51\u7edc\u538b\u7f29(\u4e0b) | \u201c\u5c0f\u58f0\u561f\u56d4\u201d\u4e13\u680f"},"content":{"rendered":"
\n


<\/p>\n

\"\u96c6\u667a\"
<\/span><\/p>\n

\"\u96c6\u667a\"
<\/span><\/p>\n

\n

\u8d75\u5a01\uff0c<\/strong><\/span>\u534e\u7c73\u79d1\u6280\u7b97\u6cd5\u5de5\u7a0b\u5e08\uff0c\u4e2d\u79d1\u9662\u9065\u611f\u4e0e\u6570\u5b57\u5730\u7403\u6240\u535a\u58eb\u3002\u7814\u7a76\u5174\u8da3\uff1a\u6df1\u5ea6\u5b66\u4e60\uff0c\u89c6\u89c9\u76ee\u6807\u8bc6\u522b\uff0c\u4eba\u4f53\u751f\u7269\u4fe1\u53f7\u8bc6\u522b\u3002<\/span><\/p>\n<\/blockquote>\n


<\/span><\/strong><\/span><\/p>\n

\u524d \u671f \u56de \u987e<\/span><\/strong><\/span><\/p>\n

\u5173\u4e8e\u6df1\u5ea6\u795e\u7ecf\u7f51\u7edc\u538b\u7f29(\u4e0a) | \u201c\u5c0f\u58f0\u561f\u56d4\u201d\u4e13\u680f<\/a> <\/strong><\/span><\/p>\n

\u4e0a\u534a\u90e8\u5206\u4e3b\u8981\u4ecb\u7ecd\u4e86SqueezeNet[1]\u548c Deep Compression[2]\u4e24\u7bc7\u6587\u7ae0\uff0c\u524d\u8005\u4e3b\u8981\u601d\u8def\u662f\u901a\u8fc7\u4f7f\u7528\u66f4\u5c0f\u7684\u5377\u79ef\u6838(1*1)\u6765\u8bbe\u8ba1\u7f51\u7edc\u67b6\u6784\uff0c\u8ba9\u53c2\u6570\u6570\u91cf\u53d8\u5c11\uff1b\u540e\u8005\u4e3b\u8981\u601d\u8def\u662f\u901a\u8fc7\u7f51\u7edc\u526a\u679d\u3001\u53c2\u6570\u538b\u7f29\u8ba9\u7f51\u7edc\u6a21\u578b\u6587\u4ef6\u53d8\u5c0f\u3002<\/span><\/p>\n


<\/span><\/p>\n

\u5bfc    \u8bed<\/span><\/strong><\/span><\/p>\n

\u4eca\u5929\u8981\u5206\u4eab\u7684\u4e24\u7bc7\u6587\u7ae0\u5219\u662f\u63a2\u8ba8\u7f51\u7edc\u538b\u7f29\u7684\u53e6\u4e00\u4e2a\u91cd\u8981\u65b9\u5411\uff1a\u7f51\u7edc\u4e8c\u503c\u5316\uff0c\u4e5f\u5c31\u662f\u75281bit\u7684\u6570\u53bb\u4ee3\u66ff\u4f20\u7edf\u7f51\u7edc\u4e2d\u7684\u5355\u7cbe\u5ea6\u6d6e\u70b9\u7cfb\u6570\u3002<\/strong>\u4e8c\u503c\u5316\u7684\u7cfb\u6570\u5927\u5c0f\u662f\u5355\u7cbe\u5ea6\u6d6e\u70b9\u6570\u503c\u76841\/32\uff0c\u4e5f\u5c31\u662f\u53ef\u4ee5\u5c06\u7f51\u7edc\u538b\u7f29\u5230\u539f\u6765\u76843%\u3002\u5176\u5b9e\u6211\u89c9\u5f97\u4e8c\u503c\u5316\u7684\u795e\u7ecf\u7f51\u7edc\u4e0e\u4eba\u8111\u5e94\u8be5\u66f4\u52a0\u76f8\u4f3c\uff0c\u6211\u4eec\u5927\u8111\u4e2d\u7684\u6bcf\u4e2a\u795e\u7ecf\u5143\u53ea\u80fd\u4f20\u5bfc\u5174\u594b\u3001\u6291\u5236\u4e24\u79cd\u72b6\u6001\uff0c\u800c\u65e0\u6cd5\u4f20\u5bfc\u5f88\u7cbe\u786e\u7684\u6570\u503c\uff0c\u6240\u4ee5\u867d\u7136\u73b0\u5728\u4e8c\u503c\u7f51\u7edc\u6682\u65f6\u6bd4\u5355\u7cbe\u5ea6\u7f51\u7edc\u6027\u80fd\u7565\u4f4e\uff0c\u4f46\u6211\u4eec\u5b8c\u5168\u6709\u7406\u7531\u76f8\u4fe1\u5355\u7cbe\u5ea6\u7f51\u7edc\u80fd\u505a\u5230\u7684\uff0c\u4e8c\u503c\u7f51\u7edc\u5e94\u8be5\u90fd\u80fd\u505a\u5230\u3002<\/span><\/p>\n

\u4e8c\u503c\u7f51\u7edc\u6700\u65e9\u7684\u7a81\u7834\u662fMatthieu\u4e8e2015\u5e7411\u6708\u53d1\u8868\u7684BinaryConnect[3]\uff0c\u5b83\u5b9e\u73b0\u4e86\u7f51\u7edc\u6743\u91cd\u7cfb\u6570\u7684\u4e8c\u503c\u5316\uff0c\u4f46\u53ea\u80fd\u8fbe\u52302\u500d\u7684\u52a0\u901f\u5ea6\u3002\u4e0d\u8fc7\u4e09\u4e2a\u6708\u540e\uff0cMatthieu\u53d1\u8868\u4e86BinaryNet [4]\uff0c\u5b9e\u73b0\u4e86\u6743\u91cd\u7cfb\u6570\u548c\u4e2d\u95f4\u7ed3\u679c\u7684\u4e8c\u503c\u5316\uff0c\u7531\u4e8e\u4e2d\u95f4\u7ed3\u679c\u7684\u4e8c\u503c\u5316\uff0cBinaryNet\u80fd\u591f\u8fbe\u52307\u500d\u7684\u52a0\u901f\u3002\u4e4b\u540e\uff0cMohammad\u6539\u8fdbBinaryNet\u7684\u5f97\u5230XNOR-Net[4]\uff0c\u5b9e\u73b0\u4e86\u8f93\u5165\u6570\u636e\u3001\u6743\u91cd\u7cfb\u6570\u5168\u90e8\u4e8c\u503c\u5316\uff0c\u76f8\u6bd4Alexnet\uff0c\u5b83\u80fd\u8fbe\u523058\u500d\u7684\u52a0\u901f\u3002<\/span><\/p>\n


<\/span><\/p>\n

<\/span><\/p>\n

\n
\n
\n
BinaryNet<\/span><\/strong><\/section>\n<\/section>\n<\/section>\n<\/section>\n


<\/p>\n


<\/p>\n

<\/span><\/p>\n

1. \u4e8c\u503c\u5316<\/strong><\/span><\/p>\n

<\/span><\/p>\n

\u6d6e\u70b9\u578b\u7cfb\u6570\u7684\u4e8c\u503c\u5316\u6709\u4e24\u79cd\u601d\u8def\uff0c\u4e00\u79cd\u662f\u6839\u636e\u7cfb\u6570\u7684\u7b26\u53f7\u8f6c\u6362\u4e3a+1\/-1<\/strong>\uff0c\u5982\u4e0b\uff1a<\/span><\/p>\n

\"\u96c6\u667a\"
 <\/span><\/p>\n

\u7b2c\u4e8c\u79cd\u65b9\u6cd5\u662f\u6309\u7167\u6839\u636e\u6d6e\u70b9\u6570\u7b97\u51fa\u4e00\u4e2a\u6982\u7387p\uff0c\u6309\u7167\u8fd9\u4e2a\u6982\u7387\u5c06\u5b83\u53d8\u6210+1\/-1\u3002<\/span><\/strong><\/p>\n

 \"\u96c6\u667a\"
<\/span><\/p>\n

\u4ece\u6570\u5b66\u4e0a\u8bb2\uff0c\u6982\u7387\u8f6c\u6362\u66f4\u597d\uff0c\u4f46\u662f\u8981\u53bb\u4e0d\u65ad\u4ea7\u751f\u968f\u673a\u6570\uff0c\u662f\u5f88\u96be\u5b9e\u73b0\u7684\uff0c\u6240\u4ee5\u6587\u7ae0\u8fd8\u662f\u7528\u7b2c\u4e00\u79cd\u65b9\u6cd5\u8fdb\u884c\u4e8c\u503c\u5316\u3002<\/span><\/p>\n


<\/p>\n

<\/span><\/p>\n

2. \u8bad\u7ec3\u65f6\u68af\u5ea6\u8ba1\u7b97\u4e0e\u66f4\u65b0<\/strong><\/span><\/p>\n


<\/span><\/p>\n

BinaryNet\u5176\u5b9e\u5e76\u6ca1\u6709\u5b8c\u5168\u4e8c\u503c\u5316\uff0c\u800c\u662f\u5728\u8bad\u7ec3\u65f6\u4fdd\u6301\u4e86\u4e00\u4efd\u6d6e\u70b9\u578b\u7684\u6743\u91cd\uff0c\u5728\u524d\u5411\u4f20\u64ad\u65f6\uff0c\u8981\u5148\u628a\u6d6e\u70b9\u578b\u6743\u91cd\u4e8c\u503c\u5316\uff0c\u7136\u540e\u518d\u505a\u5377\u79ef\u3002<\/span><\/p>\n

\u53e6\u4e00\u4e2a\u95ee\u9898\u662f\uff0cSign(x)\u51fd\u6570\u7684\u68af\u5ea6\u90fd\u662f\u96f6\uff0c\u90a3\u5c31\u65e0\u6cd5\u7528\u68af\u5ea6\u4e0b\u964d\u6cd5\u6765\u8bad\u7ec3\u7f51\u7edc\u4e86\uff0c\u6240\u4ee5\u6587\u7ae0\u7528\u4e00\u4e2aHard tanh\u51fd\u6570\uff08\u56fe 1\uff09\u6765\u8fd1\u4f3cSign(x)\u51fd\u6570\uff0c\u8fd9\u6837\u53cd\u5411\u4f20\u64ad\u65f6\uff0c\u5c31\u53ef\u4ee5\u8ba1\u7b97\u68af\u5ea6\u4e86\u3002<\/span><\/p>\n

\u503c\u5f97\u6ce8\u610f\u7684\u662f\uff0c\u65e2\u7136\u524d\u5411\u4f20\u64ad\u65f6\u6211\u4eec\u4f7f\u7528\u7684\u662f\u4e8c\u503c\u5316\u6743\u503c\uff0c\u90a3\u4e48\u8fd9\u91cc\u8ba1\u7b97\u5f97\u5230\u7684\u68af\u5ea6\u4e5f\u662f\u76f8\u5bf9\u4e8e\u7684\uff0c\u4f46\u5728\u505a\u6743\u503c\u66f4\u65b0\u65f6\uff0c\u6211\u4eec\u662f\u5bf9\u6d6e\u70b9\u578b\u6743\u503c\u8fdb\u884c\u66f4\u65b0\uff0c\u8fd9\u4e3b\u8981\u8fd8\u662f\u56e0\u4e3a\u4e8c\u503c\u5316\u6743\u503c\u5f88\u96be\u505a\u5c0f\u5e45\u7684\u66f4\u65b0\u8c03\u6574\u3002<\/span><\/p>\n

 \"\u96c6\u667a\"
<\/span><\/p>\n

\u56fe 1 Sign\u51fd\u6570\u4e0eHtanh\u51fd\u6570<\/span><\/p>\n

<\/span><\/p>\n

3. \u5377\u79ef\u8ba1\u7b97\u7684\u4f18\u5316<\/span><\/strong><\/p>\n


<\/span><\/strong><\/p>\n

\u5728CNN\u7f51\u7edc\u4e2d\u6700\u4e3b\u8981\u7684\u8ba1\u7b97\u5c31\u662f\u5377\u79ef \uff0c\u5176\u672c\u8d28\u6765\u8bf4\u662f\u4e24\u4e2a\u5411\u91cf\u6309\u4f4d\u76f8\u4e58\u540e\u76f8\u52a0\uff0c\u800c\u5728\u4e8c\u503c\u7f51\u7edc\u4e2d\uff0c\u6743\u503c\u77e9\u9635\u548c\u6fc0\u6d3b\u503c\u90fd\u4e8c\u503c\u5316\u540e\uff0c\u5377\u79ef\u64cd\u4f5c\u53ef\u4ee5\u7528\u7b80\u5355\u7684\u4f4d\u64cd\u4f5c\u6765\u5b9e\u73b0\u3002<\/span>
<\/p>\n

\u5047\u8bbe\u4e00\u4e2a\u4e8c\u503c\u6fc0\u6d3b\u5411\u91cf\u548c\u4e00\u4e2a\u4e8c\u503c\u6743\u91cd\u5411\u91cf\u505a\u5377\u79ef\uff0c\u7ed3\u679c\u7d2f\u52a0\u5230\uff0c\u6211\u4eec\u53ef\u4ee5\u901a\u8fc7\u4ee5\u4e0b\u4f4d\u64cd\u4f5c\u5b9e\u73b0\uff1a<\/span><\/p>\n

 \"\u96c6\u667a\"
<\/span><\/p>\n

\u5176\u4e2dpopcount\u7684\u662f\u8ba1\u7b97\u4e00\u4e32\u4e8c\u8fdb\u5236\u4e2d\u6709\u591a\u5c11\u4e2a1\uff0cxnor\u5c31\u662f\u5bf9\u4e24\u4e2a\u4e8c\u8fdb\u5236\u4e32\u6309\u4f4d\u53d6\u5f02\u6216\uff0c\u8fd9\u6837\u6bd4\u9010\u4e2a\u505a\u4e58\u6cd5\uff0c\u518d\u7d2f\u52a0\u51cf\u5c11\u4e86\u5927\u91cf\u8ba1\u7b97\u3002<\/span><\/p>\n

BinaryNet\u7684\u4f18\u52bf\u4e3b\u8981\u5728\u4e8e\u5176\u9884\u6d4b\u65f6\u7684\u901f\u5ea6\u5feb\uff08\u5728GPU\u4e0a\u5feb\u4e867\u500d\uff09\uff0c\u5360\u7528\u5185\u5b58\u5c0f(\u4e0e32bit DNN\u76f8\u6bd4\uff0c\u5b83\u7684\u5185\u5b58\u5360\u7528\u51cf\u5c11\u4e8632\u500d)\uff0c\u5185\u5b58\u8bbf\u95ee\u662f\u6bd4\u8ba1\u7b97\u8017\u80fd\u66f4\u5927\u7684\u64cd\u4f5c\uff0c\u6240\u4ee5\u5185\u5b58\u51cf\u5c0f\u5bf9\u80fd\u8017\u7684\u964d\u4f4e\u975e\u5e38\u663e\u8457\u3002<\/span><\/strong><\/p>\n


<\/p>\n

\n
\n
\n
XNOR-Net<\/strong><\/span><\/section>\n<\/section>\n<\/section>\n<\/section>\n


<\/span><\/p>\n

Xnor-net\u5176\u5b9e\u548cBinaryNet\u6574\u4f53\u601d\u8def\u4e5f\u662f\u5dee\u4e0d\u591a\u7684\uff0c\u53ea\u662f\u5728\u4e8c\u503c\u5316\u4e0a\u505a\u5f97\u66f4\u52a0\u5f7b\u5e95\uff0c\u5b83\u5c06\u8f93\u5165\u3001\u6743\u91cd\u3001\u4e2d\u95f4\u7ed3\u679c\u3001\u53cd\u5411\u4f20\u64ad\u68af\u5ea6\u7b49\u5168\u90e8\u4e8c\u503c\u5316\u4e86\uff0c\u4ece\u800c\u8fbe\u5230\u4e86\u66f4\u9ad8\u7684\u52a0\u901f\u6bd4\uff0c\u5f53\u7136\u5b8c\u5168\u4e8c\u503c\u5316\u540e\uff0c\u6027\u80fd\u964d\u4f4e\u4e86\u3002<\/span><\/p>\n


<\/span><\/p>\n

<\/span><\/p>\n

1. \u4e8c\u503c\u5316<\/strong><\/span><\/p>\n


<\/strong><\/span><\/p>\n

\u4e0eBinaryNet\u76f4\u63a5\u7528\u7b26\u53f7\u5c06\u6743\u503c\u4e8c\u503c\u5316\u4e3a+1\/-1\uff0cXnor-net\u7684\u505a\u6cd5\u7565\u6709\u4e0d\u540c\uff0c\u5047\u8bbe\u6743\u91cd\u77e9\u9635\u4e3a\uff0c\u6587\u7ae0\u63d0\u51fa\u7528\u4e00\u4e2a\u6807\u91cf\u548c\u4e8c\u503c\u77e9\u9635\u7684\u4e58\u79ef\u6765\u8fd1\u4f3c\uff0c\u5373\u901a\u8fc7\u7b80\u5355\u63a8\u5bfc\uff08\u89c1\u8bba\u6587\uff09\uff0c\u53ef\u4ee5\u53d1\u73b0\uff0c\u4e5f\u5c31\u662f\u8bf4\u4e8c\u503c\u77e9\u9635\u53ef\u4ee5\u7531\u7b26\u53f7\u51fd\u6570Sign\u5f97\u5230\uff0c\u800c\u6807\u91cf\u7cfb\u6570\u53ef\u4ee5\u7531\u5b9e\u503c\u6743\u91cd\u7edd\u5bf9\u503c\u7684\u5e73\u5747\u503c\u6765\u8fd1\u4f3c<\/strong>\u3002\u8fd9\u662f\u5bf9\u6743\u91cd\u77e9\u9635\u7684\u4e8c\u503c\u5316\uff0c\u5bf9\u8f93\u5165\u6570\u636e\u4e8c\u503c\u5316\u4e5f\u662f\u4e00\u6837\uff0c\u5047\u8bbe\u8f93\u5165\u6570\u636e\uff0c\u4e5f\u53ef\u4ee5\u5c06\u5176\u8fd1\u4f3c\u4e3a\u4e00\u4e2a\u6807\u91cf\u548c\u4e8c\u503c\u77e9\u9635\u7684\u4e58\u79ef\uff0c\u5373\uff0c\u5176\u4e2d, \u4e0eBinaryNet\u76f8\u6bd4\uff0c\u5728\u4e8c\u503c\u5316\u65f6\u52a0\u5165\u4e86\u4e00\u4e2a\u6807\u91cf\u7cfb\u6570\uff0c\u8868\u8fbe\u80fd\u529b\u662f\u80fd\u591f\u5f97\u5230\u4e00\u4e9b\u589e\u52a0\u7684\u3002<\/span>
<\/p>\n

\u56fe 2 \u662f\u4f5c\u8005\u7ed9\u51fa\u7684\u4e0a\u8ff0\u4e8c\u503c\u5316\u601d\u8def\u793a\u610f\u56fe\uff0c\u5176\u4e2d\u8f93\u5165X\u7684\u4e8c\u503c\u5316\u6709\u4e00\u4e9b\u6280\u5de7\uff0c\u56fe\u4e2d(2)\u662f\u76f4\u63a5\u6309\u7167\u4e00\u4e2a\u4e2a\u5377\u79ef\u6838\u7684\u5927\u5c0f\u53bb\u5bf9\u8f93\u5165\u7684\u6bcf\u4e2a\u6570\u636e\u5757\u53bb\u8ba1\u7b97\u6807\u91cf\u7cfb\u6570\uff0c\u8fd9\u6837\u5176\u5b9e\u6709\u5f88\u591a\u7684\u5197\u4f59\u8ba1\u7b97\uff0c\u56fe\u4e2d(3)\u662f\u4f5c\u8005\u5b9e\u9645\u4f7f\u7528\u7684\u5feb\u901f\u505a\u6cd5\uff0c\u4e5f\u5c31\u662f\u5148\u5c06\u8f93\u5165\u7684\u6bcf\u4e2a\u50cf\u7d20\u70b9\u6240\u6709\u901a\u9053(channel)\u503c\u6c42\u5e73\u5747\uff0c\u5c06\u4e09\u7ef4\u8f93\u5165\u8f6c\u6210\u4e8c\u7ef4\u77e9\u9635A\uff0c\u7136\u540e\u5728\u4e8c\u7ef4\u77e9\u9635A\u4e0a\u505a\u5c40\u57df\u5e73\u5747\u6765\u8ba1\u7b97\u6bcf\u4e2a\u5377\u79ef\u533a\u57df\u7684\u7cfb\u6570\u3002\u6700\u7ec8\u6548\u679c\u662f\u5c06\u5377\u79ef\u64cd\u4f5c\u8fd1\u4f3c\u6210\u8ba1\u7b97\u6210\u672c\u4f4e\u5f97\u591a\u7684\u4e8c\u503c\u64cd\u4f5c\u4e0e\u6807\u91cf\u7684\u4e58\u6cd5\u64cd\u4f5c\uff0c\u5982\u4e0b\u5f0f\u6240\u793a\uff1a<\/span><\/p>\n

 \"\u96c6\u667a\"
<\/span><\/p>\n

\u56fe 2 \u5377\u79ef\u5c42\u8f93\u5165\u3001\u6743\u91cd\u4e8c\u503c\u5316\u7684\u601d\u8def\u56fe\u793a<\/span><\/p>\n

<\/span><\/p>\n

2. \u5176\u4ed6\u4f18\u5316<\/strong><\/span><\/p>\n

\u4e00\u822c\u7684CNN\u7f51\u7edc\u4e2d\uff0c\u4e00\u4e2a\u5b8c\u6574\u7684\u5377\u79ef\u5c42\u662f\u6309\u71671)\u5377\u79ef\uff0c2)Batch Normalize\uff0c3)\u6fc0\u6d3b\uff0c4)\u6c60\u5316\u8fd9\u6837\u7684\u987a\u5e8f\u8fdb\u884c\u7684\uff0c\u7136\u800c\u5bf9\u4e8c\u503c\u5316\u540e\uff0c\u6bcf\u5c42\u7684\u6fc0\u6d3b\u503c\u90fd\u662f\u4e8c\u503c\u7684\uff0c\u5bf9\u4e8c\u503c\u6570\u7ec4\u505aMaxPool\u540e\uff0c\u5927\u90e8\u5206\u503c\u90fd\u4f1a\u53d8\u6210+1\uff0c\u8fd9\u6837\u6570\u636e\u4fe1\u606f\u5c31\u4e22\u5931\u4e86\uff0c\u6240\u4ee5XNOR-net\u505a\u4e86\u4e2a\u8c03\u6574\uff0c\u5c06\u6c60\u5316\u653e\u5728\u5377\u79ef\u4e4b\u540e\uff0c\u800c\u628aBatch Normalize\u7528\u5728\u8f93\u5165\u6570\u636e\u4e0a\u3002\u4e0d\u8fc7\u4e2a\u4eba\u611f\u89c9\u8fd9\u6837\u4e00\u4e2a\u5c0f\u7684trick\uff0c\u4f3c\u4e4e\u4e0d\u592a\u6709\u4ec0\u4e48\u597d\u7684\u7406\u8bba\u652f\u6491\uff0c\u901a\u7528\u6027\u4e5f\u6709\u5f85\u9a8c\u8bc1\u3002<\/span>
<\/p>\n

 \"\u96c6\u667a\"
<\/span><\/p>\n

\u56fe 3 XNOR-net\u4e0e\u5178\u578bCNN\u7f51\u7edc\u4e2d\u5377\u79ef\u5c42\u7684\u987a\u5e8f\u533a\u522b<\/span> <\/span><\/p>\n

\u4e0eBinaryNet\u4e00\u6837\uff0c\u56e0\u4e3a\u8bad\u7ec3\u65f6\uff0c\u8981\u91c7\u7528\u968f\u673a\u68af\u5ea6\u4e0b\u964d\u6cd5\uff0c\u800c\u7b26\u53f7\u51fd\u6570\u4f1a\u4f7f\u5f97\u68af\u5ea6\u6d88\u5931\uff0c\u6240\u4ee5\u5728\u8bad\u7ec3\u65f6\u8981\u4f7f\u7528\u4e00\u4e2a\u5b9e\u503c\u7684\u6743\u91cd\u6765\u505a\u8bad\u7ec3\u3002<\/strong> <\/span><\/p>\n

\u901a\u8fc7\u4e8c\u503c\u5316\uff0c\u539f\u5148\u7684\u5377\u79ef\u64cd\u4f5c\u4e2d\u7684\u4e58\u6cd5\uff0c\u90fd\u53ef\u4ee5\u8f6c\u6362\u6210\u4e8c\u503c\u4f4d\u8fd0\u7b97\u7684\u5f02\u6216\u64cd\u4f5c\u548c\u64cd\u4f5c\uff0c\u8fd9\u4e0eBinaryNet\u601d\u8def\u4e5f\u662f\u4e00\u6837\u7684\uff0c\u53ea\u4e0d\u8fc7\u73b0\u5728\u6240\u6709\u5730\u65b9\u90fd\u4e8c\u503c\u5316\u4e86\uff0c\u90fd\u53ef\u4ee5\u8fd9\u6837\u4f18\u5316\u8ba1\u7b97\u4e86\u3002<\/span><\/p>\n

XNOR-net\u5c06Alexnet\u4e8c\u503c\u5316\u540e\u518dImagenet\u4e0a\u505a\u4e86\u5206\u7c7b\u5b9e\u9a8c\uff0c\u9884\u6d4b\u901f\u5ea6\u63d0\u5347\u4e8658\u500d\uff0c\u4f46\u5206\u7c7b\u7cbe\u5ea6\u4e0b\u964d\u4e86\u8d85\u8fc710\u4e2a\u70b9\u3002<\/span><\/p>\n

<\/span><\/p>\n

\n
\n
\n
\u603b\u7ed3<\/strong><\/span><\/section>\n<\/section>\n<\/section>\n<\/section>\n


<\/p>\n

\u603b\u7684\u6765\u8bf4\uff0cSqueezeNet\u3001DeepCompression\u4e24\u7bc7\u6587\u7ae0\u8ba8\u8bba\u7684\u601d\u8def\u66f4\u52a0\u6e05\u6670\u76f4\u63a5\uff0c\u6548\u679c\u975e\u5e38\u597d\uff0c\u4e5f\u5bb9\u6613\u53bb\u91cd\u590d\u5b9e\u9a8c\uff1b\u800cBinaryNet\u548cXnorNet\u8fd9\u4e24\u7bc7\u6587\u7ae0\u8ba8\u8bba\u7684\u7f51\u7edc\u4e8c\u503c\u5316\uff0c\u5176\u5b9e\u8fd8\u662f\u6709\u5f88\u5927\u672a\u77e5\u7684\u65b9\u5411\uff0c\u7531\u4e8e\u76ee\u524d\u8fd8\u7f3a\u4e4f\u6bd4\u8f83\u597d\u7684\u4e8c\u503c\u7f51\u7edc\u5e95\u5c42\u6846\u67b6\uff0c\u4e8c\u503c\u5316\u540e\u5206\u7c7b\u6027\u80fd\u7684\u964d\u4f4e\uff0c\u5728\u4e0d\u540c\u7f51\u7edc\u67b6\u6784\u4e0a\u7684\u53ef\u6269\u5c55\u6027\u90fd\u8fd8\u6709\u5f88\u5927\u95ee\u9898\uff0c\u6240\u4ee5\u7f51\u7edc\u4e8c\u503c\u5316\u5728\u5b66\u672f\u754c\u8fd8\u6ca1\u6709\u53d7\u5230\u975e\u5e38\u5927\u7684\u91cd\u89c6\uff0c\u4e0d\u8fc7\u4ece\u5de5\u4e1a\u754c\u7684\u89d2\u5ea6\u6765\u8003\u8651\uff0c\u7f51\u7edc\u4e8c\u503c\u5316\u6240\u80fd\u5e26\u6765\u7684\u4ef7\u503c\u662f\u975e\u5e38\u91cd\u5927\u7684\uff0c\u76ee\u524d\u5f88\u591a\u516c\u53f8\u90fd\u5df2\u7ecf\u5728\u6295\u5165\u8d44\u6e90\u7814\u53d1\u6df1\u5ea6\u5b66\u4e60\u82af\u7247\uff0c\u76f8\u4fe1\u4e0d\u4e45\u7684\u5c06\u6765\u4e8c\u503c\u5316\u795e\u7ecf\u7f51\u7edc\u4f1a\u548c\u6df1\u5ea6\u5b66\u4e60\u82af\u7247\u4e00\u9053\u6210\u719f\u8d77\u6765\u3002<\/span><\/p>\n


<\/span><\/p>\n

\u53c2\u8003\u6587\u732e<\/span><\/strong><\/span><\/p>\n

[1]. Iandola F N, Moskewicz M W, Ashraf K, et al. SqueezeNet: AlexNet-level accuracy with 50x fewer parameters and< 1MB model size[J]. arXiv preprint arXiv:1602.07360, 2016.<\/span><\/p>\n

[2]. Han S, Mao H, Dally W J. Deep compression: Compressing deep neural network with pruning, trained quantization and huffman coding[J]. CoRR, abs\/1510.00149, 2015, 2.<\/span><\/p>\n

[3]. Courbariaux M, Bengio Y, David J P. Binaryconnect: Training deep neural networks with binary weights during propagations[C]\/\/Advances in Neural Information Processing Systems. 2015: 3123-3131.<\/span><\/p>\n

[4]. Courbariaux M, Bengio Y. Binarynet: Training deep neural networks with weights and activations constrained to+ 1 or-1[J]. arXiv preprint arXiv:1602.02830, 2016.<\/span><\/p>\n

[5]. Rastegari M, Ordonez V, Redmon J, et al. XNOR-Net: ImageNet Classification Using Binary Convolutional Neural Networks[J]. arXiv preprint arXiv:1603.05279, 2016.<\/span><\/p>\n

 <\/span><\/p>\n


<\/span><\/p>\n

–  END –<\/span><\/p>\n


<\/span><\/p>\n

\n


<\/span><\/p>\n

\u96c6\u667a\u4ff1\u4e50\u90e8<\/span><\/strong><\/span><\/strong><\/span><\/p>\n


<\/span><\/p>\n

\u82f1\u6587\u540d\uff1a Swarm Agents Club \uff0c<\/span>\u7b80\u79f0\u201c SA Club<\/span> \u201d\uff0c\u6210\u7acb\u4e8e 2008 \u5e74 \uff0c \u662f\u4e00\u4e2a\u4ece\u4e8b\u5b66\u672f\u7814\u7a76\u3001\u4eab\u53d7\u79d1\u5b66\u4e50\u8da3\u7684\u63a2\u7d22\u8005\u7684\u56e2\u4f53 \u3002\u5b83\u5021\u5bfc\u4ee5\u5e73\u7b49\u5f00\u653e\u7684\u6001\u5ea6\u3001\u79d1\u5b66\u5b9e\u8bc1\u7684\u7cbe\u795e\uff0c\u8fdb\u884c\u8de8\u5b66\u79d1\u7684\u7814\u7a76\u4e0e\u4ea4\u6d41\uff0c\u529b\u56fe\u642d\u5efa\u4e00\u4e2a\u4e2d\u56fd\u7684 \u201c \u6ca1\u6709\u56f4\u5899\u7684\u7814\u7a76\u6240 \u201d\u3002<\/strong><\/span><\/span><\/p>\n

\u96c6\u667a\u4ff1\u4e50\u90e8\u7684\u53e3\u53f7\uff1a<\/span><\/p>\n

\u8ba9\u82f9\u679c\u7838\u5f97\u66f4\u731b\u70c8\u4e9b\u5427\uff01\uff01\uff01<\/p>\n

\u96c6\u667a\u4ff1\u4e50\u90e8\u7684\u4f7f\u547d\uff08\u5373\u76ee\u6807\uff09\uff1a<\/span><\/p>\n

\u8425\u9020\u81ea\u7531\u4ea4\u6d41\u5b66\u672f\u601d\u60f3\u7684\u5c0f\u751f\u5883\uff0c<\/p>\n

 \u5b55\u80b2\u5f00\u521b\u6027\u7684\u79d1\u5b66\u53d1\u73b0\u3002<\/p>\n

\u96c6\u667a\u6bcf\u5468\u7cbe\u54c1\u63a8\u8350<\/span><\/strong>
<\/p>\n

\u96c6\u667aSuperfit AI\u5723\u8bde\u665a\u4f1a\uff0c\u8ba9\u4f60\u7684\u5723\u8bde\u5feb\u4e50\u518d\u98de\u4e00\u4f1a\u513f\uff01<\/a>
<\/strong><\/span><\/p>\n

\u6df1\u5ea6\u5b66\u4e60\u4e0e\u81ea\u7136\u8bed\u8a00\u5904\u7406<\/strong><\/a>
<\/p>\n

\u6df1\u5ea6\u5b66\u4e60\u5728\u81ea\u7136\u8bed\u8a00\u5904\u7406\u4e2d\u7684\u5e94\u7528 <\/a><\/strong>
<\/p>\n


<\/p>\n

\"\u96c6\u667a\"<\/p>\n

\u8ba9\u82f9\u679c\u7838\u5f97\u66f4\u731b\u70c8\u4e9b\u5427\uff01<\/span><\/p>\n

\n
\"\u96c6\u667a\"
<\/section>\n<\/section>\n

\u957f\u6309\u8bc6\u522b\u4e8c\u7ef4\u7801\uff0c\u5173\u6ce8\u96c6\u667aClub\uff0c<\/span><\/p>\n

\u8ba9\u6211\u4eec\u79bb\u79d1\u5b66\u63a2\u7d22\u66f4\u8fd1\u4e00\u6b65\u3002<\/span><\/p>\n

\n

\u59cb\u53d1\u4e8e\u5fae\u4fe1\u516c\u4f17\u53f7\uff1a \u96c6\u667a\u4ff1\u4e50\u90e8 <\/p>\n<\/blockquote>\n<\/div>\n","protected":false},"excerpt":{"rendered":"

\u8d75\u5a01\uff0c\u534e\u7c73\u79d1\u6280\u7b97\u6cd5\u5de5\u7a0b\u5e08\uff0c\u4e2d\u79d1\u9662\u9065\u611f\u4e0e\u6570\u5b57\u5730\u7403\u6240\u535a\u58eb\u3002\u7814\u7a76\u5174\u8da3\uff1a\u6df1\u5ea6\u5b66\u4e60\uff0c\u89c6\u89c9\u76ee\u6807\u8bc6\u522b\uff0c\u4eba\u4f53\u751f\u7269\u4fe1\u53f7\u8bc6\u522b\u3002 \u524d \u671f \u56de \u987e \u5173\u4e8e\u6df1\u5ea6\u795e\u7ecf\u7f51\u7edc\u538b\u7f29(\u4e0a) | \u201c\u5c0f\u58f0\u561f\u56d4\u201d\u4e13\u680f  \u4e0a\u534a\u90e8\u5206\u4e3b\u8981\u4ecb\u7ecd\u4e86SqueezeNet[1]\u548c Deep Compression[2]\u4e24\u7bc7\u6587\u7ae0\uff0c\u524d\u8005\u4e3b\u8981…<\/p>\n","protected":false},"author":1,"featured_media":8002,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":[],"categories":[19],"tags":[],"special":[],"_links":{"self":[{"href":"https:\/\/swarma.org\/index.php?rest_route=\/wp\/v2\/posts\/8001"}],"collection":[{"href":"https:\/\/swarma.org\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/swarma.org\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/swarma.org\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/swarma.org\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=8001"}],"version-history":[{"count":1,"href":"https:\/\/swarma.org\/index.php?rest_route=\/wp\/v2\/posts\/8001\/revisions"}],"predecessor-version":[{"id":8016,"href":"https:\/\/swarma.org\/index.php?rest_route=\/wp\/v2\/posts\/8001\/revisions\/8016"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/swarma.org\/index.php?rest_route=\/wp\/v2\/media\/8002"}],"wp:attachment":[{"href":"https:\/\/swarma.org\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=8001"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/swarma.org\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=8001"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/swarma.org\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=8001"},{"taxonomy":"special","embeddable":true,"href":"https:\/\/swarma.org\/index.php?rest_route=%2Fwp%2Fv2%2Fspecial&post=8001"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}