{"id":3969,"date":"2022-05-03T15:38:02","date_gmt":"2022-05-03T07:38:02","guid":{"rendered":"http:\/\/139.9.1.231\/?p=3969"},"modified":"2022-05-03T15:38:03","modified_gmt":"2022-05-03T07:38:03","slug":"deeplearning","status":"publish","type":"post","link":"http:\/\/139.9.1.231\/index.php\/2022\/05\/03\/deeplearning\/","title":{"rendered":"\u6df1\u5ea6\u5b66\u4e60\u8c03\u53c2\u6280\u5de7"},"content":{"rendered":"\n<p>\u4ee5\u4e0b\u6587\u7ae0\u6765\u6e90\u4e8eAI\u7b97\u6cd5\u4e0e\u56fe\u50cf\u5904\u7406 \uff0c\u4f5c\u8005AI_study<\/p>\n\n\n\n<p>\u6700\u8fd1\u5728\u8dd1\u6a21\u578b\uff0c\u53d1\u73b0\u81ea\u5df1\u8dd1\u51fa\u6765\u7684\u6027\u80fd\u603b\u662f\u8ddf\u8bba\u6587\u91cc\u7684\u6709\u4e9b\u5dee\u522b\uff0c\u5bf9\u4e8e\u5f88\u591a\u53c2\u6570\uff0c\u5b66\u4e60\u7387\u3001\u6279\u6b21\u5927\u5c0f\u7b49\u6ca1\u5565\u6982\u5ff5\u3002\u3002\u3002<\/p>\n\n\n\n<p>\u8bad\u7ec3\u6df1\u5ea6\u795e\u7ecf\u7f51\u7edc\u662f\u56f0\u96be\u7684\u3002\u5b83\u9700\u8981\u77e5\u8bc6\u548c\u7ecf\u9a8c\uff0c\u4ee5\u9002\u5f53\u7684\u8bad\u7ec3\u548c\u83b7\u5f97\u4e00\u4e2a\u6700\u4f18\u6a21\u578b\u3002\u5728\u8fd9\u7bc7\u6587\u7ae0\u4e2d\uff0c\u6211\u60f3\u5206\u4eab\u6211\u5728\u8bad\u7ec3\u6df1\u5ea6\u795e\u7ecf\u7f51\u7edc\u65f6\u5b66\u5230\u7684\u4e1c\u897f\u3002\u4ee5\u4e0b\u63d0\u793a\u548c\u6280\u5de7\u53ef\u80fd\u5bf9\u4f60\u7684\u7814\u7a76\u6709\u76ca\uff0c\u5e76\u53ef\u4ee5\u5e2e\u52a9\u4f60\u52a0\u901f\u7f51\u7edc\u67b6\u6784\u6216\u53c2\u6570\u641c\u7d22\u3002<\/p>\n\n\n\n<ul><li>1\u3001<\/li><\/ul>\n\n\n\n<p>\u5728\u4f60\u5f00\u59cb\u5efa\u7acb\u4f60\u7684\u7f51\u7edc\u4f53\u7cfb\u7ed3\u6784\uff0c\u4f60\u9700\u8981\u505a\u7684\u7b2c\u4e00\u4ef6\u4e8b\u662f\u9a8c\u8bc1\u8f93\u5165\u5230\u7f51\u7edc\u7684\u6570\u636e\uff0c\u786e\u4fdd\u8f93\u5165(x)\u5bf9\u5e94\u4e8e\u4e00\u4e2a\u6807\u7b7e(y)\u3002\u5728\u9884\u6d4b\u7684\u60c5\u51b5\u4e0b\uff0c\u786e\u4fdd\u771f\u5b9e\u6807\u7b7e(y)\u6b63\u786e\u7f16\u7801\u6807\u7b7e\u7d22\u5f15(\u6216\u8005one-hot-encoding)\u3002\u5426\u5219\uff0c\u8bad\u7ec3\u5c31\u4e0d\u8d77\u4f5c\u7528\u3002<\/p>\n\n\n\n<ul><li>2\u3001<\/li><\/ul>\n\n\n\n<p>\u51b3\u5b9a\u662f\u9009\u62e9\u4f7f\u7528\u9884\u6a21\u578b\u8fd8\u662f\u4ece\u5934\u5f00\u59cb\u8bad\u7ec3\u4f60\u7684\u7f51\u7edc?<\/p>\n\n\n\n<p>\u5982\u679c\u95ee\u9898\u57df\u4e2d\u7684\u6570\u636e\u96c6\u7c7b\u4f3c\u4e8eImageNet\u6570\u636e\u96c6\uff0c\u5219\u5bf9\u8be5\u6570\u636e\u96c6\u4f7f\u7528\u9884\u8bad\u7ec3\u6a21\u578b\u3002\u4f7f\u7528\u6700\u5e7f\u6cdb\u7684\u9884\u8bad\u7ec3\u6a21\u578b\u6709VGG net\u3001ResNet\u3001DenseNet\u6216Xception\u7b49\u3002\u6709\u8bb8\u591a\u5c42\u67b6\u6784\uff0c\u4f8b\u5982\uff0cVGG(19\u548c16\u5c42)\uff0cResNet(152, 101, 50\u5c42\u6216\u66f4\u5c11)\uff0cDenseNet(201, 169\u548c121\u5c42)\u3002\u6ce8\u610f:\u4e0d\u8981\u5c1d\u8bd5\u901a\u8fc7\u4f7f\u7528\u66f4\u591a\u7684\u5c42\u7f51\u6765\u641c\u7d22\u8d85\u53c2\u6570(\u4f8b\u5982VGG-19, ResNet-152\u6216densen -201\u5c42\u7f51\u7edc\uff0c\u56e0\u4e3a\u5b83\u5728\u8ba1\u7b97\u91cf\u5f88\u5927)\uff0c\u800c\u662f\u4f7f\u7528\u8f83\u5c11\u7684\u5c42\u7f51(\u4f8b\u5982VGG-16, ResNet-50\u6216densen -121\u5c42)\u3002\u9009\u62e9\u4e00\u4e2a\u9884\u5148\u8bad\u7ec3\u8fc7\u7684\u6a21\u578b\uff0c\u4f60\u8ba4\u4e3a\u5b83\u53ef\u4ee5\u7528\u4f60\u7684\u8d85\u53c2\u6570\u63d0\u4f9b\u6700\u597d\u7684\u6027\u80fd(\u6bd4\u5982ResNet-50\u5c42)\u3002\u5728\u4f60\u83b7\u5f97\u6700\u4f73\u8d85\u53c2\u6570\u540e\uff0c\u53ea\u9700\u9009\u62e9\u76f8\u540c\u4f46\u66f4\u591a\u7684\u5c42\u7f51(\u5982ResNet-101\u6216ResNet-152\u5c42)\uff0c\u4ee5\u63d0\u9ad8\u51c6\u786e\u6027\u3002ImageNet\uff1ahttp:\/\/www.image-net.org\/challenges\/LSVRC\/2012\/VGG net \uff1ahttps:\/\/arxiv.org\/abs\/1409.1556ResNet\uff1ahttps:\/\/arxiv.org\/abs\/1512.03385DenseNet\uff1ahttps:\/\/arxiv.org\/abs\/1608.06993Xception \uff1ahttps:\/\/arxiv.org\/abs\/1610.02357<\/p>\n\n\n\n<p>\u5fae\u8c03\u51e0\u5c42\uff0c\u6216\u8005\u5982\u679c\u4f60\u6709\u4e00\u4e2a\u5c0f\u7684\u6570\u636e\u96c6\uff0c\u53ea\u8bad\u7ec3\u5206\u7c7b\u5668\uff0c\u4f60\u4e5f\u53ef\u4ee5\u5c1d\u8bd5\u5728\u4f60\u8981\u5fae\u8c03\u7684\u5377\u79ef\u5c42\u4e4b\u540e\u63d2\u5165Dropout\u5c42\uff0c\u56e0\u4e3a\u5b83\u53ef\u4ee5\u5e2e\u52a9\u5bf9\u6297\u7f51\u7edc\u4e2d\u7684\u8fc7\u62df\u5408\u3002Dropout\uff1ahttp:\/\/jmlr.org\/papers\/v15\/srivastava14a.html<\/p>\n\n\n\n<p>\u5982\u679c\u4f60\u7684\u6570\u636e\u96c6\u4e0eImageNet\u6570\u636e\u96c6\u4e0d\u76f8\u4f3c\uff0c\u4f60\u53ef\u4ee5\u8003\u8651\u4ece\u5934\u6784\u5efa\u5e76\u8bad\u7ec3\u4f60\u7684\u7f51\u7edc\u3002<\/p>\n\n\n\n<ul><li>3\u3001<\/li><\/ul>\n\n\n\n<p>\u5728\u4f60\u7684\u7f51\u7edc\u4e2d\u59cb\u7ec8\u4f7f\u7528<strong>\u5f52\u4e00\u5316\u5c42\uff08normalization layers\uff09<\/strong>\u3002\u5982\u679c\u4f60\u4f7f\u7528\u8f83\u5927\u7684\u6279\u5904\u7406\u5927\u5c0f(\u6bd4\u598210\u4e2a\u6216\u66f4\u591a)\u6765\u8bad\u7ec3\u7f51\u7edc\uff0c\u8bf7\u4f7f\u7528<strong>\u6279\u6807\u51c6\u5316\u5c42\uff08BatchNormalization\uff09<\/strong>\u3002\u5426\u5219\uff0c\u5982\u679c\u4f60\u4f7f\u7528\u8f83\u5c0f\u7684\u6279\u5927\u5c0f(\u6bd4\u59821)\u8fdb\u884c\u8bad\u7ec3\uff0c\u5219\u4f7f\u7528InstanceNormalization\u5c42\u3002\u8bf7\u6ce8\u610f\uff0c\u5927\u90e8\u5206\u4f5c\u8005\u53d1\u73b0\uff0c\u5982\u679c\u589e\u52a0\u6279\u5904\u7406\u5927\u5c0f\uff0c\u90a3\u4e48\u6279\u5904\u7406\u89c4\u8303\u5316\u4f1a\u63d0\u9ad8\u6027\u80fd\uff0c\u800c\u5f53\u6279\u5904\u7406\u5927\u5c0f\u8f83\u5c0f\u65f6\uff0c\u5219\u4f1a\u964d\u4f4e\u6027\u80fd\u3002\u4f46\u662f\uff0c\u5982\u679c\u4f7f\u7528\u8f83\u5c0f\u7684\u6279\u5904\u7406\u5927\u5c0f\uff0c<strong>InstanceNormalization<\/strong>\u4f1a\u7565\u5fae\u63d0\u9ad8\u6027\u80fd\u3002\u6216\u8005\u4f60\u4e5f\u53ef\u4ee5\u5c1d\u8bd5<strong>\u7ec4\u89c4\u8303\u5316\uff08GroupNormalization<\/strong>\uff09\u3002BatchNormalization\uff1ahttps:\/\/arxiv.org\/abs\/1502.03167InstanceNormalization\uff1ahttps:\/\/arxiv.org\/abs\/1607.08022GroupNormalization\uff1ahttps:\/\/arxiv.org\/abs\/1803.08494<\/p>\n\n\n\n<ul><li>4\u3001SpatialDropout<\/li><\/ul>\n\n\n\n<p>\u5982\u679c\u4f60\u6709\u4e24\u4e2a\u6216\u66f4\u591a\u7684\u5377\u79ef\u5c42(\u6bd4\u5982Li)\u5bf9\u76f8\u540c\u7684\u8f93\u5165(\u6bd4\u5982F)\u8fdb\u884c\u64cd\u4f5c\uff08\u53c2\u8003\u4e0b\u9762\u7684\u793a\u610f\u56fe\u7406\u89e3\uff09\uff0c\u90a3\u4e48\u5728\u7279\u5f81\u8fde\u63a5\u540e\u4f7f\u7528SpatialDropout\u3002\u7531\u4e8e\u8fd9\u4e9b\u5377\u79ef\u5c42\u662f\u5728\u76f8\u540c\u7684\u8f93\u5165\u4e0a\u64cd\u4f5c\u7684\uff0c\u56e0\u6b64\u8f93\u51fa\u7279\u5f81\u5f88\u53ef\u80fd\u662f\u76f8\u5173\u7684\u3002\u56e0\u6b64\uff0cSpatialDropout\u5220\u9664\u4e86\u90a3\u4e9b\u76f8\u5173\u7684\u7279\u5f81\uff0c\u5e76\u9632\u6b62\u7f51\u7edc\u4e2d\u7684\u8fc7\u62df\u5408\u3002<strong>\u6ce8\u610f:&nbsp;<\/strong>\u5b83\u4e3b\u8981\u7528\u4e8e\u8f83\u4f4e\u7684\u5c42\u800c\u4e0d\u662f\u8f83\u9ad8\u7684\u5c42\u3002SpatialDropout\uff1ahttps:\/\/arxiv.org\/abs\/1411.4280<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" width=\"800\" height=\"456\" src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/05\/image.png\" alt=\"\" class=\"wp-image-3974\" srcset=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/05\/image.png 800w, http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/05\/image-300x171.png 300w, http:\/\/139.9.1.231\/wp-content\/uploads\/2022\/05\/image-768x438.png 768w\" sizes=\"(max-width: 800px) 100vw, 800px\" \/><\/figure>\n\n\n\n<p>SpatialDropout\u662fTompson\u7b49\u4eba\u5728\u56fe\u50cf\u9886\u57df\u63d0\u51fa\u7684\u4e00\u79cddropout\u65b9\u6cd5\u3002\u666e\u901a\u7684dropout\u4f1a\u968f\u673a\u5730\u5c06\u90e8\u5206\u5143\u7d20\u7f6e\u96f6\uff0c\u800cSpatialDropout\u4f1a\u968f\u673a\u5730\u5c06\u90e8\u5206\u533a\u57df\u7f6e\u96f6\uff0c\u8be5dropout\u65b9\u6cd5\u5728\u56fe\u50cf\u8bc6\u522b\u9886\u57df\u5b9e\u8df5\u8bc1\u660e\u662f\u6709\u6548\u7684\u3002Dropout\u64cd\u4f5c\u968f\u673a\u5730\u5c06\u90e8\u5206\u5143\u7d20\u7f6e\u96f6\uff0c\u5e76\u4e14\u5bf9\u975e\u96f6\u90e8\u5206\u505a\u4e86\u4e00\u4e2a\u5c3a\u5ea6\u53d8\u6362\u3002\u5c3a\u5ea6\u53d8\u6362\u7684\u5e45\u5ea6\u8ddf\u521d\u59cb\u5316\u7684drop_rate\u6709\u5173\u3002<br>\u4f5c\u7528<br>\u4e00\u822c\uff0c\u6211\u4eec\u4f1a\u5c06dropout\u7406\u89e3\u4e3a\u201c\u4e00\u79cd\u4f4e\u6210\u672c\u7684\u96c6\u6210\u7b56\u7565\u201d\uff0c\u8fd9\u662f\u5bf9\u7684\uff0c\u5177\u4f53\u8fc7\u7a0b\u53ef\u4ee5\u5927\u6982\u8fd9\u6837\u7406\u89e3\uff1a<br>\u7ecf\u8fc7\u4e0a\u8ff0\u7f6e\u96f6\u64cd\u4f5c\u540e\uff0c\u6211\u4eec\u53ef\u4ee5\u8ba4\u4e3a\u96f6\u7684\u90e8\u5206\u662f\u88ab\u4e22\u5f03\u7684\uff0c\u4e22\u5931\u4e86\u4e00\u90e8\u5206\u4fe1\u606f\u3002\u56e0\u800c\uff0c\u903c\u7740\u6a21\u578b\u7528\u5269\u4e0b\u7684\u4fe1\u606f\u53bb\u62df\u5408\u76ee\u6807\u3002\u7136\u800c\u6bcf\u6b21dropout\u662f\u968f\u673a\u7684\u3002\u6211\u4eec\u5c31\u4e0d\u80fd\u4fa7\u91cd\u4e8e\u67d0\u4e9b\u8282\u70b9\uff0c\u6240\u4ee5\u603b\u7684\u6765\u8bf4\u5c31\u662f\u2014\u6bcf\u6b21\u903c\u7740\u6a21\u578b\u7528\u5c11\u91cf\u7684\u7279\u5f81\u5b66\u4e60\uff0c\u6bcf\u6b21\u88ab\u5b66\u4e60\u7684\u7279\u5f81\u53c8\u4e0d\u540c\uff0c\u90a3\u4e48\u5c31\u662f\u8bf4\uff0c\u6bcf\u4e2a\u7279\u5f81\u90fd\u5e94\u8be5\u5bf9<br>\u6a21\u578b\u7684\u9884\u6d4b\u6709\u6240\u8d21\u732e\uff08\u800c\u4e0d\u662f\u4fa7\u91cd\u4e8e\u90e8\u5206\u7279\u5f81\uff0c\u5bfc\u81f4\u8fc7\u62df\u5408\uff09\u3002<\/p>\n\n\n\n<p>\u901a\u7684dropout\u4f1a\u968f\u673a\u72ec\u7acb\u5730\u5c06\u90e8\u5206\u5143\u7d20\u7f6e\u96f6\uff0c\u800cSpatialDropout1D\u4f1a\u968f\u673a\u5730\u5bf9\u67d0\u4e2a\u7279\u5b9a\u7684\u7eac\u5ea6\u5168\u90e8\u7f6e\u96f6\u3002\u56e0\u6b64SpatialDropout1D\u9700\u8981\u6307\u5b9aDropout\u7ef4\u5ea6\uff0c\u5373\u5bf9\u5e94dropout\u51fd\u6570\u4e2d\u7684\u53c2\u6570noise_shape\u3002<br><\/p>\n\n\n\n<ul><li>5\u3001<\/li><\/ul>\n\n\n\n<p>\u4e3a\u4e86\u786e\u5b9a\u4f60\u7684\u7f51\u7edc\u5bb9\u91cf\uff0c\u5c1d\u8bd5\u7528\u4e00\u5c0f\u90e8\u5206\u8bad\u7ec3\u4f8b\u5b50\u6765\u8d85\u8f7d\u4f60\u7684\u7f51\u7edc(andrej karpathy\u7684\u63d0\u793a)\u3002\u5982\u679c\u5b83\u6ca1\u6709\u8d85\u8f7d\uff0c\u589e\u52a0\u4f60\u7684\u7f51\u7edc\u5bb9\u91cf\u3002\u5728\u8fc7\u62df\u5408\u540e\uff0c\u4f7f\u7528\u6b63\u5219\u5316\u6280\u5de7\u5982L1\u3001L2\u3001Dropout\u6216\u5176\u4ed6\u6280\u672f\u6765\u5bf9\u6297\u8fc7\u62df\u5408\u3002L1\uff1ahttps:\/\/keras.io\/regularizers\/L2\uff1ahttps:\/\/keras.io\/regularizers\/Dropout\uff1ahttp:\/\/jmlr.org\/papers\/v15\/srivastava14a.html<\/p>\n\n\n\n<ul><li>6\u3001<\/li><\/ul>\n\n\n\n<p>\u53e6\u4e00\u79cd\u6b63\u5219\u5316\u6280\u672f\u662f\u7ea6\u675f\u6216\u9650\u5236\u4f60\u7684\u7f51\u7edc\u6743\u503c\u3002\u8fd9\u4e5f\u6709\u52a9\u4e8e\u9632\u6b62\u7f51\u7edc\u4e2d\u7684\u68af\u5ea6\u7206\u70b8\u95ee\u9898\uff0c\u56e0\u4e3a\u6743\u503c\u603b\u662f\u6709\u754c\u7684\u3002\u4e0eL2\u6b63\u5219\u5316\u76f8\u53cd\uff0c\u5728\u4f60\u7684\u635f\u5931\u51fd\u6570\u4e2d\u60e9\u7f5a\u9ad8\u6743\u91cd\uff0c\u8fd9\u4e2a\u7ea6\u675f\u76f4\u63a5\u6b63\u5219\u5316\u4f60\u7684\u6743\u91cd\u3002\u4f60\u53ef\u4ee5\u5728Keras\u4e2d\u8f7b\u677e\u8bbe\u7f6e\u6743\u91cd\u7ea6\u675f<\/p>\n\n\n\n<ul><li>7\u3001<\/li><\/ul>\n\n\n\n<p>\u5bf9\u6570\u636e\u8fdb\u884c\u5747\u503c\u51cf\u6cd5\u6709\u65f6\u4f1a\u4ea7\u751f\u975e\u5e38\u7cdf\u7cd5\u7684\u6548\u679c\uff0c\u7279\u522b\u662f\u5bf9\u7070\u5ea6\u56fe\u50cf\u8fdb\u884c\u51cf\u6cd5(\u6211\u4e2a\u4eba\u5728\u524d\u666f\u5206\u5272\u9886\u57df\u5c31\u9047\u5230\u8fc7\u8fd9\u4e2a\u95ee\u9898)\u3002<\/p>\n\n\n\n<ul><li>8\u3001\u5728\u8bad\u7ec3\u524d\u548c\u8bad\u7ec3\u671f\u95f4\uff0c\u786e\u4fdd\u6253\u4e71\u8bad\u7ec3\u6570\u636e\uff0c\u4ee5\u9632\u4f60\u4e0d\u80fd\u4ece\u65f6\u5e8f\u6570\u636e\u4e2d\u83b7\u53d6\u6709\u7528\u4fe1\u606f\u3002\u8fd9\u53ef\u80fd\u6709\u52a9\u4e8e\u63d0\u9ad8\u60a8\u7684\u7f51\u7edc\u6027\u80fd\u3002<\/li><li>9\u3001\u5982\u679c\u4f60\u7684\u95ee\u9898\u57df\u4e0e\u7a20\u5bc6\u9884\u6d4b\uff08dense prediction\uff09\u76f8\u5173(\u5982\u8bed\u4e49\u5206\u5272)\uff0c\u6211\u5efa\u8bae\u4f60\u4f7f\u7528\u81a8\u80c0\u6b8b\u5dee\u7f51\u7edc\u4f5c\u4e3a\u9884\u8bad\u7ec3\u6a21\u578b\uff0c\u56e0\u4e3a\u5b83\u6700\u9002\u5408\u7a20\u5bc6\u9884\u6d4b\u3002Dilated Residual Networks\uff1ahttps:\/\/arxiv.org\/abs\/1705.09914<\/li><\/ul>\n\n\n\n<ul><li>10. \u8981\u6355\u83b7\u5bf9\u8c61\u5468\u56f4\u7684\u4e0a\u4e0b\u6587\u4fe1\u606f\uff0c\u53ef\u4ee5\u4f7f\u7528\u591a\u5c3a\u5ea6\u7279\u6027\u7684\u6c60\u5316\u6a21\u5757\u3002\u8be5\u601d\u60f3\u6210\u529f\u5730\u5e94\u7528\u4e8e\u8bed\u4e49\u5206\u5272\u6216\u524d\u666f\u5206\u5272\u4e2d\u3002semantic segmentation\uff1ahttps:\/\/arxiv.org\/abs\/1802.02611foreground segmentation\uff1ahttps:\/\/arxiv.org\/abs\/1808.01477<\/li><li>11 \u3001 Opt-out void labels(\u6216\u6a21\u7cca\u533a\u57df)\u4ece\u60a8\u7684\u635f\u5931\u6216\u7cbe\u5ea6\u8ba1\u7b97\uff0c\u5982\u679c\u6709\u3002\u8fd9\u53ef\u4ee5\u5e2e\u52a9\u4f60\u7684\u7f51\u7edc\u5728\u9884\u6d4b\u65f6\u66f4\u6709\u4fe1\u5fc3\u3002<\/li><li>12\u3001\u5982\u679c\u4f60\u6709\u9ad8\u5ea6\u4e0d\u5e73\u8861\u7684\u6570\u636e\u95ee\u9898\uff0c\u5728\u8bad\u7ec3\u671f\u95f4\u5e94\u7528\u7c7b\u522b\u52a0\u6743\u64cd\u4f5c\u3002\u6362\u53e5\u8bdd\u8bf4\uff0c\u7ed9\u7a00\u5c11\u7684\u7c7b\u66f4\u591a\u7684\u6743\u91cd\uff0c\u4f46\u7ed9\u4e3b\u8981\u7c7b\u66f4\u5c11\u7684\u6743\u91cd\u3002\u4f7f\u7528sklearn\u53ef\u4ee5\u5f88\u5bb9\u6613\u5730\u8ba1\u7b97\u7c7b\u6743\u91cd\u3002\u6216\u8005\u5c1d\u8bd5\u4f7f\u7528\u8fc7\u91c7\u6837\u548c\u6b20\u91c7\u6837\u6280\u672f\u91cd\u65b0\u91c7\u6837\u4f60\u7684\u8bad\u7ec3\u96c6\u3002\u8fd9\u4e5f\u53ef\u4ee5\u5e2e\u52a9\u63d0\u9ad8\u9884\u6d4b\u7684\u51c6\u786e\u6027\u3002<\/li><li>13\u3001\u9009\u62e9\u4e00\u4e2a\u6b63\u786e\u7684\u4f18\u5316\u5668\u3002\u6709\u8bb8\u591a\u6d41\u884c\u7684\u81ea\u9002\u5e94\u4f18\u5316\u5668\uff0c\u5982Adam, Adagrad, Adadelta\uff0c\u6216RMSprop\u7b49\u3002SGD+\u52a8\u91cf\u88ab\u5e7f\u6cdb\u5e94\u7528\u4e8e\u5404\u79cd\u95ee\u9898\u9886\u57df\u3002\u6709\u4e24\u4ef6\u4e8b\u9700\u8981\u8003\u8651\uff1a\u7b2c\u4e00\uff0c\u5982\u679c\u4f60\u5173\u5fc3\u5feb\u901f\u6536\u655b\uff0c\u4f7f\u7528\u81ea\u9002\u5e94\u4f18\u5316\u5668\uff0c\u5982Adam\uff0c\u4f46\u5b83\u53ef\u80fd\u4f1a\u9677\u5165\u5c40\u90e8\u6781\u5c0f\uff0c\u63d0\u4f9b\u4e86\u7cdf\u7cd5\u7684\u6cdb\u5316(\u4e0b\u56fe)\u3002\u7b2c\u4e8c\uff0cSGD+momentum\u53ef\u4ee5\u5b9e\u73b0\u627e\u5230\u5168\u5c40\u6700\u5c0f\u503c\uff0c\u4f46\u5b83\u4f9d\u8d56\u4e8e\u9c81\u68d2\u521d\u59cb\u5316\uff0c\u800c\u4e14\u53ef\u80fd\u6bd4\u5176\u4ed6\u81ea\u9002\u5e94\u4f18\u5316\u5668\u9700\u8981\u66f4\u957f\u7684\u65f6\u95f4\u6765\u6536\u655b(\u4e0b\u56fe)\u3002\u6211\u5efa\u8bae\u4f60\u4f7f\u7528SGD+\u52a8\u91cf\uff0c\u56e0\u4e3a\u5b83\u80fd\u8fbe\u5230\u66f4\u597d\u7684\u6700\u4f73\u6548\u679c\u3002\u6709\u4e09\u4e2a\u5b66\u4e60\u7387\u8d77\u70b9(\u53731e- 1,1e -3\u548c1e-6)\u3002\u5982\u679c\u60a8\u5bf9\u9884\u8bad\u7ec3\u6a21\u578b\u8fdb\u884c\u5fae\u8c03\uff0c\u8bf7\u8003\u8651\u5c0f\u4e8e1e-3(\u6bd4\u59821e-4)\u7684\u4f4e\u5b66\u4e60\u7387\u3002\u5982\u679c\u60a8\u4ece\u5934\u5f00\u59cb\u8bad\u7ec3\u60a8\u7684\u7f51\u7edc\uff0c\u8bf7\u8003\u8651\u4e00\u4e2a\u5927\u4e8e\u6216\u7b49\u4e8e1e-3\u7684\u5b66\u4e60\u7387\u3002\u60a8\u53ef\u4ee5\u5c1d\u8bd5\u8fd9\u4e9b\u8d77\u70b9\uff0c\u5e76\u8c03\u6574\u5b83\u4eec\uff0c\u770b\u770b\u54ea\u4e2a\u662f\u6700\u597d\u7684\uff0c\u9009\u62e9\u90a3\u4e2a\u3002\u8fd8\u6709\u4e00\u4ef6\u4e8b\uff0c\u60a8\u53ef\u4ee5\u8003\u8651\u901a\u8fc7\u4f7f\u7528 Learning Rate Schedulers\u6765\u964d\u4f4e\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u7684\u5b66\u4e60\u7387\u3002\u8fd9\u4e5f\u53ef\u4ee5\u5e2e\u52a9\u63d0\u9ad8\u7f51\u7edc\u6027\u80fd\u3002<\/li><li>14\u3001 \u9664\u4e86Learning Rate Schedule \u5916\uff0c\u5373\u5728\u4e00\u5b9a\u7684\u6b21\u6570\u540e\u964d\u4f4e\u5b66\u4e60\u7387\uff0c\u8fd8\u6709\u53e6\u4e00\u79cd\u65b9\u5f0f\uff0c\u6211\u4eec\u53ef\u4ee5\u7531\u4e00\u4e9b\u56e0\u7d20\u51cf\u5c11\u5b66\u4e60\u7387\uff0c\u5982\u679c\u9a8c\u8bc1\u635floss\u5728\u67d0\u4e9bepoch(\u6bd4\u59825)\u505c\u6b62\u6539\u5584\uff0c\u51cf\u5c0f\u5b66\u4e60\u7387\u548c\u5982\u679c\u9a8c\u8bc1\u635f\u5931\u505c\u6b62\u6539\u5584\u5728\u67d0\u4e9bepoch(\u6bd4\u598210)\uff0c\u505c\u6b62\u8bad\u7ec3\u8fc7\u7a0b\u3002\u8fd9\u53ef\u4ee5\u901a\u8fc7\u5728Keras\u4e2d\u4f7f\u7528early stop\u7684ReduceLROnPlateau\u5f88\u5bb9\u6613\u505a\u5230\u3002<\/li><li>15\u3001\u5982\u679c\u60a8\u5728dense prediction\u9886\u57df\u5de5\u4f5c\uff0c\u5982\u524d\u666f\u5206\u5272\u6216\u8bed\u4e49\u5206\u5272\uff0c\u60a8\u5e94\u8be5\u4f7f\u7528\u8df3\u8fc7\u8fde\u63a5\uff0c\u56e0\u4e3a\u5bf9\u8c61\u8fb9\u754c\u6216\u6709\u7528\u7684\u4fe1\u606f\u4f1a\u7531\u4e8e\u6700\u5927\u6c60\u5316\u64cd\u4f5c\u6216strided convolutions\u800c\u4e22\u5931\u3002\u8fd9\u4e5f\u53ef\u4ee5\u5e2e\u52a9\u60a8\u7684\u7f51\u7edc\u8f7b\u677e\u5730\u5b66\u4e60\u7279\u5f81\u7a7a\u95f4\u5230\u56fe\u50cf\u7a7a\u95f4\u7684\u7279\u5f81\u6620\u5c04\uff0c\u6709\u52a9\u4e8e\u7f13\u89e3\u7f51\u7edc\u4e2d\u7684\u6d88\u5931\u68af\u5ea6\u95ee\u9898\u3002<\/li><li>16\u3001\u6570\u636e\u8d8a\u591a\u8d8a\u597d!\u603b\u662f\u4f7f\u7528\u6570\u636e\u589e\u5f3a\uff0c\u5982\u6c34\u5e73\u7ffb\u8f6c\uff0c\u65cb\u8f6c\uff0c\u7f29\u653e\u88c1\u526a\u7b49\u3002\u8fd9\u53ef\u4ee5\u5e2e\u52a9\u5927\u5e45\u5ea6\u63d0\u9ad8\u7cbe\u786e\u5ea6\u3002<\/li><li>17\u3001\u4f60\u5fc5\u987b\u8981\u6709\u4e00\u4e2a\u9ad8\u901f\u7684GPU\u6765\u8fdb\u884c\u8bad\u7ec3\uff0c\u4f46\u662f\u8fd9\u6709\u70b9\u6602\u8d35\u3002\u5982\u679c\u4f60\u60f3\u4f7f\u7528\u514d\u8d39\u7684\u4e91GPU\uff0c\u6211\u63a8\u8350\u4f7f\u7528\u8c37\u6b4cColab\u3002\u5982\u679c\u4f60\u4e0d\u77e5\u9053\u4ece\u54ea\u91cc\u5f00\u59cb\uff0c\u770b\u770b\u6211\u4e4b\u524d\u7684\u6587\u7ae0\u6216\u8005\u5c1d\u8bd5\u5404\u79cd\u4e91GPU\u5e73\u53f0\uff0c\u5982Floydhub\u6216Paperspace\u7b49\u3002<\/li><li>18\u3001<\/li><\/ul>\n\n\n\n<p>\u5728ReLU\u4e4b\u524d\u4f7f\u7528\u6700\u5927\u6c60\u5316\u6765\u8282\u7701\u4e00\u4e9b\u8ba1\u7b97\u3002\u7531\u4e8eReLU\u9608\u503c\u7684\u503c\u4e3a0\uff1af(x)=max(0,x)\u548c\u6700\u5927\u6c60\u5316\u53ea\u6709max\u6fc0\u6d3b\uff1af(x)=max(x1,x2\uff0c\u2026\uff0cxi)\uff0c\u4f7f\u7528Conv > MaxPool > ReLU \u800c\u4e0d\u662fConv > ReLU > MaxPool\u3002\u4f8b\u5982\uff0c\u5047\u8bbe\u6211\u4eec\u6709\u4e24\u4e2a\u4eceConv\u6765\u7684\u6fc0\u6d3b\u503c(\u53730.5\u548c-0.5):\u56e0\u6b64MaxPool > ReLU = max(0, max(0.5\uff0c-0.5)) = 0.5\u548cReLU > MaxPool = max(max(0,0.5), max(0,-0.5)) = 0.5\u770b\u5230\u4e86\u5417?\u8fd9\u4e24\u4e2a\u64cd\u4f5c\u7684\u8f93\u51fa\u4ecd\u7136\u662f0.5\u3002\u5728\u8fd9\u79cd\u60c5\u51b5\u4e0b\uff0c\u4f7f\u7528MaxPool > ReLU\u53ef\u4ee5\u8282\u7701\u4e00\u4e2amax \u64cd\u4f5c\u3002<\/p>\n\n\n\n<p>19\u3001 \u8003\u8651\u91c7\u7528<strong>\u6df1\u5ea6\u53ef\u5206\u79bb\u5377\u79ef\u8fd0\u7b97<\/strong>\uff0c\u4e0e\u5e38\u89c4\u7684\u5377\u79ef\u8fd0\u7b97\u76f8\u6bd4\uff0c\u8be5\u8fd0\u7b97\u901f\u5ea6\u5feb\uff0c\u4e14\u53c2\u6570\u6570\u91cf\u5927\u5927\u51cf\u5c11\u3002Depthwise Separable Convolution\uff1ahttps:\/\/arxiv.org\/abs\/1610.02357<\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u4ee5\u4e0b\u6587\u7ae0\u6765\u6e90\u4e8eAI\u7b97\u6cd5\u4e0e\u56fe\u50cf\u5904\u7406 \uff0c\u4f5c\u8005AI_study \u6700\u8fd1\u5728\u8dd1\u6a21\u578b\uff0c\u53d1\u73b0\u81ea\u5df1\u8dd1\u51fa\u6765\u7684\u6027\u80fd\u603b\u662f\u8ddf\u8bba\u6587\u91cc\u7684\u6709\u4e9b &hellip; <a href=\"http:\/\/139.9.1.231\/index.php\/2022\/05\/03\/deeplearning\/\" class=\"more-link\">\u7ee7\u7eed\u9605\u8bfb<span class=\"screen-reader-text\">\u6df1\u5ea6\u5b66\u4e60\u8c03\u53c2\u6280\u5de7<\/span><\/a><\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":[],"categories":[4,12],"tags":[],"_links":{"self":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/3969"}],"collection":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/comments?post=3969"}],"version-history":[{"count":12,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/3969\/revisions"}],"predecessor-version":[{"id":3982,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/3969\/revisions\/3982"}],"wp:attachment":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/media?parent=3969"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/categories?post=3969"},{"taxonomy":"post_tag","embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/tags?post=3969"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}