{"id":12182,"date":"2016-11-25T00:39:46","date_gmt":"2016-11-24T15:39:46","guid":{"rendered":"https:\/\/jirak.net\/wp\/aws%eb%a1%9c-%eb%94%a5-%eb%9f%ac%eb%8b%9d%ec%9d%84-%ec%9c%84%ed%95%9c-%ed%94%84%eb%a0%88%ec%9e%84%ec%9b%8c%ed%81%ac-mxnet-%ed%99%9c%ec%9a%a9%ed%95%98%ea%b8%b0\/"},"modified":"2016-11-25T11:35:09","modified_gmt":"2016-11-25T02:35:09","slug":"aws%eb%a1%9c-%eb%94%a5-%eb%9f%ac%eb%8b%9d%ec%9d%84-%ec%9c%84%ed%95%9c-%ed%94%84%eb%a0%88%ec%9e%84%ec%9b%8c%ed%81%ac-mxnet-%ed%99%9c%ec%9a%a9%ed%95%98%ea%b8%b0","status":"publish","type":"post","link":"https:\/\/jirak.net\/wp\/aws%eb%a1%9c-%eb%94%a5-%eb%9f%ac%eb%8b%9d%ec%9d%84-%ec%9c%84%ed%95%9c-%ed%94%84%eb%a0%88%ec%9e%84%ec%9b%8c%ed%81%ac-mxnet-%ed%99%9c%ec%9a%a9%ed%95%98%ea%b8%b0\/","title":{"rendered":"AWS\ub85c \ub525 \ub7ec\ub2dd\uc744 \uc704\ud55c \ud504\ub808\uc784\uc6cc\ud06c MxNet \ud65c\uc6a9\ud558\uae30"},"content":{"rendered":"<p>AWS\ub85c \ub525 \ub7ec\ub2dd\uc744 \uc704\ud55c \ud504\ub808\uc784\uc6cc\ud06c MxNet \ud65c\uc6a9\ud558\uae30<br \/>\n<img loading=\"lazy\" decoding=\"async\" src=\"https:\/\/jirak.net\/wp\/wp-content\/uploads\/2016\/11\/mxnetbanner.png\" width=\"650\" height=\"269\"><\/p>\n<p>\uba38\uc2e0 \ub7ec\ub2dd\uc5d0 \ub300\ud55c \uad00\uc2ec\uc774 \ud3ed\ubc1c\uc801\uc73c\ub85c \uc99d\uac00\ud588\uc2b5\ub2c8\ub2e4. \ud2b9\ud788, \uad6d\ub0b4\uc5d0\uc11c\ub294 \uc62c\ud574 \uc54c\ud30c\uace0\ub85c \uc778\ud574 \ub525\ub7ec\ub2dd(Deep Learning)\uc5d0 \ub300\ud55c \uad00\uc2ec\uc774 \ud06c\uac8c \uc99d\uac00\ud558\uc600\uc2b5\ub2c8\ub2e4. \uc778\uacf5 \uc2e0\uacbd\ub9dd\uc744 \uc774\uc6a9\ud55c \ub525 \ub7ec\ub2dd \uae30\ubc95\uc740 \ud558\ub4dc\uc6e8\uc5b4 \uc131\ub2a5\uc758 \ube44\uc57d\uc801\uc778 \uac1c\uc120\uacfc \uc2e0\uacbd\ub9dd \uc54c\uace0\ub9ac\uc998\uc758 \uac1c\uc120\uc73c\ub85c \uc778\ud574 \uc2e4\uc81c \ud65c\uc6a9 \uac00\ub2a5\ud55c \uc218\uc900\uc73c\ub85c \ube60\ub974\uac8c \ubcc0\ud654\ud558\uc600\uc2b5\ub2c8\ub2e4.<\/p>\n<p>\uc774\ub7ec\ud55c \uad00\uc2ec\uc73c\ub85c \uc778\ud574 \ubd84\uc0b0 \ub525 \ub7ec\ub2dd \ud504\ub808\uc784\uc6cc\ud06c(distributed deep-learning framework)\uac00 \ub9ce\uc774 \uac1c\ubc1c\ub418\uc5b4 \uc624\ud508\uc18c\uc2a4 \ud615\uc2dd\uc73c\ub85c \uacf5\uac1c\ub418\uace0 \uc788\ub294 \uc0c1\ud669\uc785\ub2c8\ub2e4. \uac8c\uc784 \uc11c\ube44\uc2a4 \uc601\uc5ed\uc5d0\uc11c\ub3c4 \ucc28\uce30 \ub525 \ub7ec\ub2dd\uc744 \ud65c\uc6a9\ud55c \uc11c\ube44\uc2a4\ub4e4\uc774 \uc9c0\uc18d\uc801\uc73c\ub85c \ub098\ud0c0\ub098\uace0 \uc788\uc2b5\ub2c8\ub2e4. \uc778\uacf5 \uc9c0\ub2a5 \uc5f0\uad6c \ubc0f \uc11c\ube44\uc2a4 \uac1c\ubc1c\uc790\uc758 \uc694\uad6c \uc0ac\ud56d, \uc9c0\uc6d0 \uc5b8\uc5b4 \ubc0f \ud558\ub4dc\uc6e8\uc5b4 \ub530\ub77c \uc5ec\ub7ec \uc885\ub958\uc758 \ub525 \ub7ec\ub2dd \ud504\ub808\uc784\uc6cc\ud06c\uac00 \uac1c\ubc1c\ub418\uc5b4 \uacf5\uac1c\ub418\uc5c8\uc2b5\ub2c8\ub2e4. \ub300\ud45c\uc801\uc778 \uac83\uc774 <a href=\"http:\/\/mxnet.io\/\">MXNet<\/a>, &nbsp; <a href=\"http:\/\/caffe.berkeleyvision.org\/\">Caffe<\/a>&nbsp;, <a href=\"http:\/\/deeplearning.net\/software\/theano\/\">Theano<\/a>&nbsp;, <a href=\"https:\/\/www.tensorflow.org\/\">TensorFlow<\/a>\u2122, <a href=\"http:\/\/torch.ch\/\">Torch<\/a>&nbsp; \ub4f1\uc774 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n<p><strong>\ub525\ub7ec\ub2dd \ud504\ub808\uc784\uc6cc\ud06c \uc18c\uac1c<\/strong><br \/> DeepLearning 4J\uc5d0\uc11c \ube44\uad50\ud55c <a href=\"https:\/\/deeplearning4j.org\/compare-dl4j-torch7-pylearn\">\ub525\ub7ec\ub2dd \ud504\ub808\uc784\uc6cc\ud06c \ube44\uad50 \uae30\uc0ac<\/a>\ub97c \ubcf4\uba74 \uc544\ub798\uc640 \uac19\uc740 \uc7a5\ub2e8\uc810\uc744 \ub098\uc5f4\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4. <em>(\ubcf8 \uc790\ub8cc\ub294 AWS\uc758 \uacf5\uc2dd\uc801\uc778 \uc790\ub8cc\uac00 \uc544\ub2c8\uba70, DeepLearning 4J\uc758 \uc758\uacac\uc785\ub2c8\ub2e4.)<\/em><\/p>\n<table style=\"height: 882px\" border=\"1\" width=\"702\">\n<tbody>\n<tr>\n<td style=\"background-color: #ccc\">\ud504\ub808\uc784\uc6cc\ud06c<\/td>\n<td style=\"background-color: #ccc\">\uc7a5\uc810<\/td>\n<td style=\"background-color: #ccc\">\ub2e8\uc810<\/td>\n<\/tr>\n<tr>\n<td style=\"background-color: #eee\" width=\"82\"><strong>Theano<\/strong><\/td>\n<td width=\"162\">\n<ul>\n<li>Python \uc9c0\uc6d0<\/li>\n<li>Wrapper \ub97c \ud1b5\ud55c \ub192\uc740 \ucd94\uc0c1\ud654\ub85c \uc0ac\uc6a9\uc131 \ud3b8\ub9ac<\/li>\n<li>\uc5ec\ub7ec \uc5d0\ucf54\uc2dc\uc2a4\ud15c\uc774 \uc874\uc7ac<\/li>\n<li>&nbsp;\uc5f0\uad6c\uc6a9\uc73c\ub85c \ub9ce\uc774 \uc0ac\uc6a9\ub428<\/li>\n<\/ul>\n<\/td>\n<td width=\"182\">\n<ul>\n<li>Theano\uc790\uccb4\ub294 \ub85c\uc6b0\ub808\ubca8 \ub77c\uc774\ube0c\ub7ec\ub9ac<\/li>\n<li>\ud070 \uaddc\ubaa8 \ubaa8\ub378\uc5d0 \ub9ce\uc740 \ucef4\ud30c\uc77c \uc2dc\uac04<\/li>\n<li>torch\uc5d0 \ube44\ud574 \ub9e4\uc6b0 \ud070 \ub77c\uc774\ube0c\ub7ec\ub9ac<\/li>\n<li>\uc5d0\ub7ec\uba54\uc2dc\uc9c0\uac00 \ubd80\uc815\ud655<\/li>\n<\/ul>\n<\/td>\n<\/tr>\n<tr>\n<td style=\"background-color: #eee\" width=\"82\"><strong>Torch<\/strong><\/td>\n<td width=\"162\">\n<ul>\n<li>\ubaa8\ub4c8\ud654\ub41c \ub77c\uc774\ube0c\ub7ec\ub9ac\ub85c \uc0c1\ud638 \uc5f0\uacc4\uac00 \uc26c\uc6c0<\/li>\n<li>GPU\uc9c0\uc6d0, \ubcf8\uc778 \ub808\uc774\uc5b4 \ud0c0\uc785 \uc791\uc131\uc774 \ud3b8\ub9ac<\/li>\n<li>\uc120\ud6c8\ub828\ub41c \ubaa8\ub378\ub4e4\uc774 \ub9ce\uc74c<\/li>\n<\/ul>\n<\/td>\n<td width=\"182\">\n<ul>\n<li>Lua \uae30\ubc18<\/li>\n<li>\ud68c\uadc0 \ub274\ub7f4 \ub124\ud2b8\uc6cc\ud06c\uc5d0 \uc801\ud569\ud558\uc9c0 \uc54a\uc74c<\/li>\n<li>\ubb38\uc11c\ud654 \ubd80\uc2e4<\/li>\n<\/ul>\n<\/td>\n<\/tr>\n<tr>\n<td style=\"background-color: #eee\" width=\"82\"><strong>TensorFlow<\/strong><\/td>\n<td width=\"162\">\n<ul>\n<li>Python + Numpy<\/li>\n<li>\ucef4\ud4e8\ud305 \uadf8\ub798\ud504 \ucd94\uc0c1\ud654<\/li>\n<li>Theano\ubcf4\ub2e4 \ube60\ub978 \ucef4\ud30c\uc77c<\/li>\n<li>\uc2dc\uac01\ud654\ub97c \uc704\ud55c TensorBoard<\/li>\n<li>\ub370\uc774\ud130\uc640 \ubaa8\ub378\uc758 \ubcd1\ub82c\ud654<\/li>\n<\/ul>\n<\/td>\n<td width=\"182\">\n<ul>\n<li>\ub2e4\ub978 \ud504\ub808\uc784\uc6cc\ud06c\ubcf4\ub2e4 \ub290\ub9bc<\/li>\n<li>Torch\ubcf4\ub2e4 \ud6e8\uc52c \ud070 \ub77c\uc774\ube0c\ub7ec\ub9ac<\/li>\n<li>\uc120 \ud6c8\ub828\ub41c \ubaa8\ub378\uc774 \uc801\uc74c<\/li>\n<li>\uacc4\uc0b0 \uadf8\ub798\ud504\uac00 Python\uc73c\ub85c \ub418\uc5b4 \uc788\uc5b4\uc11c \ub290\ub9bc<\/li>\n<li>\ub3c4\uad6c\ub85c\uc11c\uc758 \uae30\ub2a5\uc774 \uc57d\ud568<\/li>\n<\/ul>\n<\/td>\n<\/tr>\n<tr>\n<td style=\"background-color: #eee\" width=\"82\"><strong>Caffe<\/strong><\/td>\n<td width=\"162\">\n<ul>\n<li>\uc774\ubbf8\uc9c0 \ud504\ub85c\uc138\uc2f1\uc5d0 \uc801\ud569<\/li>\n<li>\uc798 \ud29c\ub2dd\ub41c \ub124\ud2b8\uc6cc\ud06c<\/li>\n<li>\ucf54\ub4dc \uc791\uc131\uc5c6\uc774 \ubaa8\ub378 \ud2b8\ub808\uc774\ub2dd \uac00\ub2a5<\/li>\n<li>Python\uc778\ud130\ud398\uc774\uc2a4\uac00 \uc720\uc6a9<\/li>\n<\/ul>\n<\/td>\n<td width=\"182\">\n<ul>\n<li>GPU\ub97c \uc704\ud574\uc11c\ub294 C++\/CUDA\uc791\uc131 \ud544\uc694<\/li>\n<li>\ud68c\uadc0 \ub124\ud2b8\uc6cc\ud06c\uc5d0\ub294 \ubd80\uc801\ud569<\/li>\n<li>\ud070 \ub124\ud2b8\uc6cc\ud06c\uc5d0\ub294 \ubd80\uc801\uc808<\/li>\n<li>\ud655\uc7a5\uc131\uc774 \ub5a8\uc5b4\uc9d0<\/li>\n<\/ul>\n<\/td>\n<\/tr>\n<tr>\n<td style=\"background-color: #eee\" width=\"82\"><strong>MxNet<\/strong><\/td>\n<td width=\"162\">\n<ul>\n<li>\ud63c\ud569 \ud328\ub7ec\ub2e4\uc784 \uc9c0\uc6d0(symbolic\/imperative)<\/li>\n<li>\uc790\ub3d9 \ubbf8\ubd84\ud654<\/li>\n<li>GPU, mobile\uc5d0\uc11c\ub3c4 \ub3d9\uc791<\/li>\n<li>\uc5ec\ub7ec \uc5b8\uc5b4 \uc9c0\uc6d0(C++, Python, R, Scala, Julia, Matlab and Javascript)<\/li>\n<li>\ucd5c\uc801\ud654\ub41c C++ \uc5d4\uc9c4\uc73c\ub85c \uc88b\uc740 \uc131\ub2a5<\/li>\n<\/ul>\n<\/td>\n<td width=\"182\">\n<ul>\n<li>\ub85c\uc6b0 \ub808\ubca8 \ud150\uc11c \uc5f0\uc0b0\uc790\uac00 \uc801\uc74c<\/li>\n<li>\ud750\ub984 \uc81c\uc5b4 \uc5f0\uc0b0\uc790 \uc9c0\uc6d0\ud558\uc9c0 \uc54a\uc74c<\/li>\n<li>\ucef4\ud30c\uc77c \uc138\ud305\uc5d0 \ub530\ub77c \uacb0\uacfc\uac00 \ub2ec\ub77c\uc9d0.<\/li>\n<li>\uc790\uc2e0\uc758 \ucee4\uc2a4\ucef4 \ub808\uc774\uc5b4 \uc0dd\uc131\uc744 \uc704\ud574\uc11c\ub294 \uc5b4\ub290\uc815\ub3c4 \ubc31\uc5d4\ub4dc \ud150\uc11c \ub77c\uc774\ube0c\ub7ec\ub9ac \uc774\ud574\uac00 \ud544\uc694<\/li>\n<\/ul>\n<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p>\uc544\ub9c8\uc874\uc758 CTO\uc778 Werner Vogels \ubc15\uc0ac\uaed8\uc11c\ub294 \ucd5c\uadfc <a href=\"http:\/\/www.allthingsdistributed.com\/2016\/11\/mxnet-default-framework-deep-learning-aws.html\">MXNet \u2013 Deep Learning Framework of Choice at AWS<\/a>\ub77c\ub294 \uae00\uc5d0\uc11c \ud655\uc7a5 \ub2a5\ub825, \uac1c\ubc1c \uc18d\ub3c4, \uc774\ub3d9\uc131 \ub4f1\uc758 \ub2e4\uc591\ud55c \uc694\uc778\uc744 \ube44\ucd94\uc5b4 \ubcfc \ub54c, MXNet\uc774 \uac00\uc7a5 \uc88b\uc740 \ub525\ub7ec\ub2dd \ud504\ub808\uc784\uc6cd\uc774\ub77c\uace0 \ud310\ub2e8\ud558\uace0, \uc774\ub97c \uae30\ubc18\ud55c \ub525\ub7ec\ub2dd \uc11c\ube44\uc2a4 \uac1c\ubc1c \uc9c0\uc6d0 \ubc0f \uc624\ud508 \uc18c\uc2a4 \uc9c0\uc6d0\uc5d0 \ub300\ud55c \uc758\uc9c0\ub97c \ud53c\ub825\ud55c \ubc14 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n<p><img decoding=\"async\" class=\"aligncenter\" src=\"http:\/\/www.allthingsdistributed.com\/images\/mxnetbanner.png\" alt=\"\" width=\"650\" \/><\/p>\n<p>MxNet\uc740 \uc624\ud508\uc18c\uc2a4\ub85c \uc5ec\ub7ec \uc5b8\uc5b4\ub97c \uc9c0\uc6d0\ud558\uace0 \ubaa8\ubc14\uc77c\ubd80\ud130 \uc11c\ubc84\uae4c\uc9c0 \ub2e4\uc591\ud55c \ub514\ubc14\uc774\uc2a4\ub97c \uc9c0\uc6d0\ud558\ub294 \ub525 \ub7ec\ub2dd \ud504\ub808\uc784\uc6cc\ud06c \uc785\ub2c8\ub2e4. CPU\uc640 GPU \uc5f0\uc0b0\uc744 \uc9c0\uc6d0\ud558\uace0, \uc2ec\ubcfc\ub9ad\uacfc \uba85\ub839\uc801(imperative) \ud504\ub85c\uadf8\ub798\ubc0d\uc758 \ud63c\ud569 \ubc29\uc2dd \uae4c\uc9c0 \uc9c0\uc6d0\ud558\uba70 \ucd5c\uc801\ud654\ub41c \uc5d4\uc9c4\uc744 \uc0ac\uc6a9\ud574\uc11c \uc131\ub2a5\uc774 \ub6f0\uc5b4\ub0a9\ub2c8\ub2e4.<\/p>\n<p>\ub610\ud55c \uc2e4\ubb34\uc801\uc73c\ub85c \ub9ce\uc774 \uc0ac\uc6a9\ud558\ub294 Python, C++, R, Scala, Julia, Matlab, and JavaScript\uc744 \uc9c0\uc6d0\ud558\ub294 \ub4f1 \uc0b0\uc5c5\uacc4\uc5d0\uc11c \uc751\uc6a9\ud558\uae30\uc5d0 \ub9e4\uc6b0 \uc801\ud569\ud55c \ub525 \ub7ec\ub2dd \ud504\ub808\uc784\uc6cc\ud06c\uc785\ub2c8\ub2e4.<\/p>\n<p>\uc544\ub798 \uadf8\ub9bc\uc5d0\uc11c \ubcf4\uc2dc\ub2e4\uc2dc\ud53c, Inception v3 (MXNet \ubc0f P2 \uc778\uc2a4\ud134\uc2a4 \uc0ac\uc6a9)\ub97c \ud1b5\ud574 GPU \uc22b\uc790\ub97c \uc99d\uac00\uc2dc\ucf30\uc744 \ub54c, \ub2e4\ub978 \ub77c\uc774\ube0c\ub7ec\ub9ac \ubcf4\ub2e4 \ube60\ub978 \ucc98\ub9ac\ub7c9\uc744 \uac00\uc9d0\uacfc \ub3d9\uc2dc\uc5d0 GPU \uc22b\uc790\uac00 \uc99d\uac00\ud558\ub294 \ud655\uc7a5 \uc0c1\ud669\uc5d0\uc11c\ub3c4 \ucc98\ub9ac\ub7c9\uc758 \ud6a8\uc728\uc774 85%\uc5d0 \ub2ec\ud560 \uc815\ub3c4\ub85c \ub6f0\uc5b4\ub09c \uc131\ub2a5\uc744 \ubcf4\uc5ec \uc8fc\uace0 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n<p><img decoding=\"async\" class=\"aligncenter\" src=\"http:\/\/www.allthingsdistributed.com\/images\/mxnetgraph.png\" alt=\"\" width=\"650\" \/><\/p>\n<p>\ucef4\ud4e8\ud305 \ucc98\ub9ac\ub7c9 \ubc0f \ud6a8\uc728 \ubfd0\ub9cc \uc544\ub2c8\ub77c \uba54\ubaa8\ub9ac \uc0ac\uc6a9\ub7c9\ub3c4 \uc911\uc694\ud569\ub2c8\ub2e4. MXNet\uc740 1,000\uac1c\uc758 \uc2e0\uacbd\ub9dd \ub808\uc774\uc5b4\ub97c \uc0ac\uc6a9\ud560 \ub54c 4GB \uc774\ud558\uc758 \uba54\ubaa8\ub9ac\ub97c \uc0ac\uc6a9\ud558\uace0, \uc774\uc2dd\uc131 \uba74\uc5d0\uc11c\ub3c4 \ub2e4\uc591\ud55c \ud50c\ub7ab\ud3fc\uc744 \uc9c0\uc6d0\ud569\ub2c8\ub2e4. \uc548\ub4dc\ub85c\uc774\ub4dc\ub098 iOS\uc5d0\uc11c\ub3c4 \ud65c\uc6a9 \uac00\ub2a5\ud558\uace0, \uc2ec\uc9c0\uc5b4 \uc790\ubc14\uc2a4\ud06c\ub9bd\ud2b8 \ud655\uc7a5 \uae30\ub2a5\uc73c\ub85c \uc6f9 \ube0c\ub77c\uc6b0\uc800\uc5d0\uc11c\ub3c4 \uc2e4\ud589 \ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n<p><strong>DeepLearning AMI\ub97c \ud1b5\ud574 MXNet \uc2e4\ud589\ud558\uae30<\/strong><br \/> \uc774 \uae00\uc5d0\uc11c\ub294 <a href=\"https:\/\/aws.amazon.com\/ko\/blogs\/korea\/new-p2-instance-type-for-amazon-ec2-up-to-16-gpus\/\">Amazon EC2\uc758 \uc2e0\uaddc GPU \uae30\ubc18 P2 \uc778\uc2a4\ud134\uc2a4<\/a> \ubc0f G2 \uc778\uc2a4\ud134\uc2a4\ub97c \ud1b5\ud574 \ub525\ub7ec\ub2dd API\ub97c \uae30\ubc18\uc73c\ub85c MXNet\uc744 \uac04\ub2e8\ud788 \uc2e4\ud589\ud574 \ubcf4\uaca0\uc2b5\ub2c8\ub2e4. (P2 \ubc0f G2 \uc778\uc2a4\ud134\uc2a4\ub294 \uc2dc\uac04\ub2f9 \uac00\uaca9\uc774 \ub2e4\ub978 \uc778\uc2a4\ud134\uc2a4\uc5d0 \ube44\ud574 \ub192\uc73c\ubbc0\ub85c, \uc9e7\uc740 \uc2dc\uac04 \ud14c\uc2a4\ud2b8\ub97c \uc704\ud574\uc11c\ub294 \uc2a4\ud31f \uc778\uc2a4\ud134\uc2a4\ub97c \uc774\uc6a9\ud574 \ubcf4\ub294 \uac83\ub3c4 \uad8c\uc7a5\ud569\ub2c8\ub2e4.)<\/p>\n<p><img decoding=\"async\" class=\"aligncenter size-full wp-image-1921\" style=\"width: 80%\" src=\"https:\/\/d2908q01vomqb2.cloudfront.net\/7b52009b64fd0a2a49e6d8a939753077792b0554\/2016\/11\/23\/2016-11-mxnet-1.png\" alt=\"2016-11-mxnet-1\" \/><\/p>\n<p>\uba3c\uc800 AWS \ub9c8\ucf13 \ud50c\ub808\uc774\uc2a4\uc5d0\uc11c \uc81c\uacf5\ud558\ub294 <a href=\"https:\/\/aws.amazon.com\/marketplace\/pp\/B01M0AXXQB\">Amazon Deep Learning AMI<\/a>\uc744 \uae30\ubc18\uc73c\ub85c \uc778\uc2a4\ud134\uc2a4\ub97c \uc2e4\ud589\ud569\ub2c8\ub2e4. \ubcf8 AMI\uc5d0 \uc124\uce58\ub41c \ub525 \ub7ec\ub2dd \ud504\ub808\uc784\uc6cc\ud06c\ub294 Caffe, MxNet, TensorFlow, Theano, Torch \uc785\ub2c8\ub2e4. \uc6b0\ub9ac\ub294 \uc5ec\uae30\uc11c MxNet\uc744 \uc0ac\uc6a9\ud574 \ubcf4\uaca0\uc2b5\ub2c8\ub2e4. SSH\ub85c \uc811\uc18d\ud574\uc11c \uc778\uc2a4\ud134\uc2a4\uc758 src \ub514\ub809\ud1a0\ub9ac\uc5d0 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n<p><img decoding=\"async\" class=\"aligncenter size-full wp-image-1922\" style=\"width: 80%\" src=\"https:\/\/d2908q01vomqb2.cloudfront.net\/7b52009b64fd0a2a49e6d8a939753077792b0554\/2016\/11\/23\/2016-11-mxnet-2.png\" alt=\"2016-11-mxnet-2\" \/><\/p>\n<p>\uc778\uc2a4\ud134\uc2a4\uc758 src\/mxnet\/example \ub514\ub809\ud1a0\ub9ac\ub97c \ubcf4\uba74 \ub9ce\uc740 \uc608\uc81c\ub4e4\uc774 \uc874\uc7ac\ud569\ub2c8\ub2e4. \uc774\uc640 \uad00\ub828\ub41c \ud29c\ud1a0\ub9ac\uc5bc\uc740 <a href=\"http:\/\/mxnet.io\/tutorials\/index.html\">http:\/\/mxnet.io\/tutorials\/index.html<\/a> \ub9c1\ud06c\ub97c \ucc38\uc870\ud558\uba74 \ub429\ub2c8\ub2e4.<\/p>\n<p>\uc6b0\ub9ac\uac00 \uc2e4\ud589\uc744 \ud574\ubcfc \uac83\uc740 \uc22b\uc790 \uc774\ubbf8\uc9c0\ub97c \ud2b8\ub808\uc774\ub2dd \ud574\uc11c \ud2b9\uc815 \uc774\ubbf8\uc9c0 \ub0b4 \uc22b\uc790 \ub370\uc774\ud130\uac00 \uc5b4\ub5a4 \uac83\uc778\uc9c0 \ucc3e\uc544\ub0b4\ub294 \uc608\uc81c\uc785\ub2c8\ub2e4. \uac04\ub2e8\ud55c Python \ucf54\ub4dc\ub97c \ud1b5\ud574 \ud574 \ubcfc \uc218 \uc788\uc2b5\ub2c8\ub2e4. (<a href=\"http:\/\/mxnet.io\/tutorials\/python\/mnist.html\">http:\/\/mxnet.io\/tutorials\/python\/mnist.html<\/a>.)<\/p>\n<p>\ub2e4\ub9cc \uc5ec\uae30\uc11c \ud50c\ub86f\ud305\uc758 \uacbd\uc6b0 \ud604\uc7ac \uc2e4\ud589 \uc911\uc778 \uc778\uc2a4\ud134\uc2a4\uc5d0\uc11c \uad6c\ub3d9\uc774 \ub418\uc9c0 \uc54a\uae30 \ub54c\ubb38\uc5d0 \ud50c\ub86f(plot)\uc744 \uc774\ubbf8\uc9c0\ub85c \ucd9c\ub825\ud558\ub294 \ubd80\ubd84\uc740 \uc81c\uc678\ud558\uace0 \uc2e4\ud589\uc744 \ud574\uc57c \ud569\ub2c8\ub2e4. \ubb38\uc11c\uc5d0\ub294 \ub2e8\uacc4\ubcc4\ub85c \uc218\ud589\uc744 \ud558\uac8c \ub418\uc5b4 \uc788\uc2b5\ub2c8\ub2e4\ub9cc \uc18c\uc2a4 \ubd80\ubd84\uc744 \uace0\uccd0\uc11c \ud55c\ubc88\uc5d0 \uc2e4\ud589\ud558\uac8c \uc218\uc815\uc744 \ud55c \uac83\uc774 \ub2e4\uc74c Python \uc18c\uc2a4 \uc785\ub2c8\ub2e4.<\/p>\n<pre class=\" language-python\"><code class=\" languange-python\">import mxnet as mx\n\ndef to4d(img):\n&nbsp; &nbsp; return img.reshape(img.shape[0], 1, 28, 28).astype(np.float32)\/255\n\nbatch_size = 100\ntrain_iter = mx.io.NDArrayIter(to4d(train_img), train_lbl, batch_size, shuffle=True)\nval_iter = mx.io.NDArrayIter(to4d(val_img), val_lbl, batch_size)\n\n\n# Create a place holder variable for the input data\ndata = mx.sym.Variable('data')\n# Flatten the data from 4-D shape (batch_size, num_channel, width, height)&nbsp;\n# into 2-D (batch_size, num_channel*width*height)\ndata = mx.sym.Flatten(data=data)\n\n# The first fully-connected layer\nfc1&nbsp; = mx.sym.FullyConnected(data=data, name='fc1', num_hidden=128)\n# Apply relu to the output of the first fully-connnected layer\nact1 = mx.sym.Activation(data=fc1, name='relu1', act_type=&quot;relu&quot;)\n\n# The second fully-connected layer and the according activation function\nfc2&nbsp; = mx.sym.FullyConnected(data=act1, name='fc2', num_hidden = 64)\nact2 = mx.sym.Activation(data=fc2, name='relu2', act_type=&quot;relu&quot;)\n\n# The thrid fully-connected layer, note that the hidden size should be 10, which is the number of unique digits\nfc3&nbsp; = mx.sym.FullyConnected(data=act2, name='fc3', num_hidden=10)\n# The softmax and loss layer\nmlp&nbsp; = mx.sym.SoftmaxOutput(data=fc3, name='softmax')\n\n# We visualize the network structure with output size (the batch_size is ignored.)\nshape = {&quot;data&quot; : (batch_size, 1, 28, 28)}\nmx.viz.plot_network(symbol=mlp, shape=shape)\n\n\nimport logging\nlogging.getLogger().setLevel(logging.DEBUG)\n\nmodel = mx.model.FeedForward(\n&nbsp; &nbsp; symbol = mlp, &nbsp; &nbsp; &nbsp; # network structure\n&nbsp; &nbsp; num_epoch = 10, &nbsp; &nbsp; # number of data passes for training&nbsp;\n&nbsp; &nbsp; learning_rate = 0.1 # learning rate of SGD&nbsp;\n)\nmodel.fit(\n&nbsp; &nbsp; X=train_iter, &nbsp; &nbsp; &nbsp; # training data\n&nbsp; &nbsp; eval_data=val_iter, # validation data\n&nbsp; &nbsp; batch_end_callback = mx.callback.Speedometer(batch_size, 200) # output progress for each 200 data batches\n)\n\nprob = model.predict(val_img[0:1].astype(np.float32)\/255)[0]\nprint 'Classified as %d with probability %f' % (prob.argmax(), max(prob))<\/code><\/pre>\n<p>\uc704\uc758 \ucf54\ub4dc\ub97c \uc2e4\ud589\ud558\uba74 \ub2e4\uc74c\uacfc \uac19\uc774 \ubaa8\ub378\uc744 \ud6c8\ub828\ud558\uac8c \ub418\uace0, \uadf8 \ub2e4\uc74c\uc5d0 \uc22b\uc790 7 \uc774\ubbf8\uc9c0\ub97c \ud310\ub2e8\ud574\uc11c \uacb0\uacfc\ub97c \ucd9c\ub825\ud569\ub2c8\ub2e4.<\/p>\n<p><img decoding=\"async\" class=\"aligncenter size-full wp-image-1923\" style=\"width: 80%\" src=\"https:\/\/d2908q01vomqb2.cloudfront.net\/7b52009b64fd0a2a49e6d8a939753077792b0554\/2016\/11\/23\/2016-11-mxnet-3.png\" alt=\"2016-11-mxnet-3\" \/><\/p>\n<p>MxNet \uc758 <a href=\"https:\/\/github.com\/dmlc\/mxnet\/tree\/master\/example\">Github \ub808\ud3ec\uc9c0\ud130\ub9ac\uc758 example<\/a>\uc744 \ubcf4\uba74 \uc774 \uc678\uc5d0\ub3c4 \uc544\uc8fc \ub2e4\uc591\ud55c \uc0d8\ud50c \ucf54\ub4dc\uac00 \uc874\uc7ac\ud569\ub2c8\ub2e4.<\/p>\n<p>\ub525 \ub7ec\ub2dd\uc740 \uc774\uc81c \ub9ce\uc740 \ubd84\uc57c\uc5d0\uc11c \ube60\ub974\uac8c \uc751\uc6a9\ub418\uace0 \uc788\uc2b5\ub2c8\ub2e4. \uac8c\uc784 \uac1c\ubc1c\uc790 \ubd84\ub4e4\ub3c4 \uc774\ub7ec\ud55c \ub525 \ub7ec\ub2dd \uae30\ubc95\uc744 \ud65c\uc6a9\ud55c \ub2e4\uc591\ud55c \uac8c\uc784 AI\ub97c \ub9cc\ub4dc\uc154\uc11c \ud65c\uc6a9\ud558\uc2dc\uae30 \ubc14\ub78d\ub2c8\ub2e4. \ub9cc\uc57d \uc2e4\uc81c \uc11c\ube44\uc2a4\ub97c \uc704\ud574 \ud074\ub7ec\uc2a4\ud130\ub97c \uad6c\uc131\ud574\uc57c \ud55c\ub2e4\uba74, <a href=\"https:\/\/aws.amazon.com\/cloudformation\">AWS CloudFormation<\/a>\ub97c \ud1b5\ud574 AWS \uc790\uc6d0\uc744 \uc190\uc27d\uac8c \ub9cc\ub4e4\uace0 \uc6b4\uc601\ud560 \uc218 \uc788\ub294 \ubc29\uc2dd\uc744 \ud1b5\ud574 \ub525\ub7ec\ub2dd \ud074\ub7ec\uc2a4\ud130\ub97c \ub9cc\ub4e4 \uc218 \uc788\ub294 <a href=\"https:\/\/github.com\/dmlc\/mxnet\/tree\/master\/tools\/cfn\">MXNet\uc758 CF \ud15c\ud50c\ub9bf<\/a>\uc744 \uc18c\uac1c\ud569\ub2c8\ub2e4.<\/p>\n<p>\uc774 \ud15c\ud50c\ub9bf\uc744 \uc774\uc6a9\ud558\uc5ec <a href=\"https:\/\/aws.amazon.com\/marketplace\/pp\/B01M0AXXQB\">Amazon Deep Learning AMI<\/a>\uc744 <a href=\"https:\/\/aws.amazon.com\/ko\/blogs\/korea\/new-p2-instance-type-for-amazon-ec2-up-to-16-gpus\/\">Amazon EC2\uc758 \uc2e0\uaddc GPU P2 \uc778\uc2a4\ud134\uc2a4<\/a>\uc5d0 \uad6c\uc131\ud558\uc5ec, \uc790\ub3d9 \uc2a4\ucf00\uc77c\ub9c1\uc744 \uc9c0\uc6d0\ud558\ub294 \ubd84\uc0b0 \ub525\ub7ec\ub2dd \ud074\ub7ec\uc2a4\ud130\ub97c \ub9cc\ub4e4\uc5b4 \uc6b4\uc601\ud560 \uc218 \uc788\uac8c \ub429\ub2c8\ub2e4. (\uc790\uc138\ud55c \uac83\uc740 <a href=\"https:\/\/aws.amazon.com\/ko\/blogs\/korea\/distributed-deep-learning-made-easy\/\">AWS\ub97c \ud1b5\ud55c \ubd84\uc0b0 \ub525\ub7ec\ub2dd(Deep Learning) \uad6c\uc131\ud558\uae30<\/a>\ub97c \ucc38\uace0\ud558\uc138\uc694.)<\/p>\n<p><em>\ubcf8 \uae00\uc740 \uc544\ub9c8\uc874\uc6f9\uc11c\ube44\uc2a4 \ucf54\ub9ac\uc544\uc758 \uc194\ub8e8\uc158\uc988 \uc544\ud0a4\ud14d\ud2b8\uac00 \uad6d\ub0b4 \uace0\uac1d\uc744 \uc704\ud574 \uc804\ud574 \ub4dc\ub9ac\ub294 <a href=\"https:\/\/aws.amazon.com\/ko\/blogs\/korea\/category\/korea-techtips\/\">AWS \ud65c\uc6a9 \uae30\uc220 \ud301<\/a>\uc744 \ubcf4\ub0b4\ub4dc\ub9ac\ub294 \ucf54\ub108\ub85c\uc11c, \uc774\ubc88 \uae00\uc740 \ubc15\uc120\uc6a9 \uc194\ub8e8\uc158\uc988 \uc544\ud0a4\ud14d\ud2b8\uaed8\uc11c \uc791\uc131\ud574\uc8fc\uc168\uc2b5\ub2c8\ub2e4.<\/em><\/p>\n<p><img decoding=\"async\" src=\"http:\/\/d0.awsstatic.com\/events\/aws-hosted-events\/2016\/AWS%20Cloud%2006\/Speaker_1.jpg\" alt=\"\" \/><\/p>\n<p>Source: <a href=\"https:\/\/aws.amazon.com\/ko\/blogs\/korea\/aws-deep-learning-framework-mxnet\/\" target=\"_blank\">AWS\ub85c \ub525 \ub7ec\ub2dd\uc744 \uc704\ud55c \ud504\ub808\uc784\uc6cc\ud06c MxNet \ud65c\uc6a9\ud558\uae30<\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<div class=\"mh-excerpt\"><p>AWS\ub85c \ub525 \ub7ec\ub2dd\uc744 \uc704\ud55c \ud504\ub808\uc784\uc6cc\ud06c MxNet \ud65c\uc6a9\ud558\uae30 \uba38\uc2e0 \ub7ec\ub2dd\uc5d0 \ub300\ud55c \uad00\uc2ec\uc774 \ud3ed\ubc1c\uc801\uc73c\ub85c \uc99d\uac00\ud588\uc2b5\ub2c8\ub2e4. \ud2b9\ud788, \uad6d\ub0b4\uc5d0\uc11c\ub294 \uc62c\ud574 \uc54c\ud30c\uace0\ub85c \uc778\ud574 \ub525\ub7ec\ub2dd(Deep Learning)\uc5d0 \ub300\ud55c \uad00\uc2ec\uc774 \ud06c\uac8c \uc99d\uac00\ud558\uc600\uc2b5\ub2c8\ub2e4. \uc778\uacf5 \uc2e0\uacbd\ub9dd\uc744 \uc774\uc6a9\ud55c \ub525 \ub7ec\ub2dd \uae30\ubc95\uc740 \ud558\ub4dc\uc6e8\uc5b4 \uc131\ub2a5\uc758 \ube44\uc57d\uc801\uc778 \uac1c\uc120\uacfc \uc2e0\uacbd\ub9dd \uc54c\uace0\ub9ac\uc998\uc758 \uac1c\uc120\uc73c\ub85c \uc778\ud574 \uc2e4\uc81c \ud65c\uc6a9 \uac00\ub2a5\ud55c \uc218\uc900\uc73c\ub85c \ube60\ub974\uac8c \ubcc0\ud654\ud558\uc600\uc2b5\ub2c8\ub2e4. \uc774\ub7ec\ud55c \uad00\uc2ec\uc73c\ub85c \uc778\ud574 \ubd84\uc0b0 \ub525 \ub7ec\ub2dd \ud504\ub808\uc784\uc6cc\ud06c(distributed deep-learning framework)\uac00 \ub9ce\uc774 \uac1c\ubc1c\ub418\uc5b4 \uc624\ud508\uc18c\uc2a4 \ud615\uc2dd\uc73c\ub85c \uacf5\uac1c\ub418\uace0 \uc788\ub294 \uc0c1\ud669\uc785\ub2c8\ub2e4. \uac8c\uc784 \uc11c\ube44\uc2a4 \uc601\uc5ed\uc5d0\uc11c\ub3c4 \ucc28\uce30 \ub525 \ub7ec\ub2dd\uc744 \ud65c\uc6a9\ud55c \uc11c\ube44\uc2a4\ub4e4\uc774 \uc9c0\uc18d\uc801\uc73c\ub85c \ub098\ud0c0\ub098\uace0 \uc788\uc2b5\ub2c8\ub2e4. \uc778\uacf5 \uc9c0\ub2a5 \uc5f0\uad6c \ubc0f \uc11c\ube44\uc2a4 \uac1c\ubc1c\uc790\uc758 \uc694\uad6c \uc0ac\ud56d, \uc9c0\uc6d0 \uc5b8\uc5b4 \ubc0f \ud558\ub4dc\uc6e8\uc5b4 \ub530\ub77c \uc5ec\ub7ec \uc885\ub958\uc758 \ub525 \ub7ec\ub2dd \ud504\ub808\uc784\uc6cc\ud06c\uac00 \uac1c\ubc1c\ub418\uc5b4 \uacf5\uac1c\ub418\uc5c8\uc2b5\ub2c8\ub2e4. \ub300\ud45c\uc801\uc778 \uac83\uc774 MXNet, &nbsp; Caffe&nbsp;, Theano&nbsp;, TensorFlow\u2122, Torch&nbsp; \ub4f1\uc774 \uc788\uc2b5\ub2c8\ub2e4. \ub525\ub7ec\ub2dd \ud504\ub808\uc784\uc6cc\ud06c \uc18c\uac1c DeepLearning 4J\uc5d0\uc11c \ube44\uad50\ud55c \ub525\ub7ec\ub2dd \ud504\ub808\uc784\uc6cc\ud06c \ube44\uad50 \uae30\uc0ac\ub97c \ubcf4\uba74 \uc544\ub798\uc640 \uac19\uc740 \uc7a5\ub2e8\uc810\uc744 \ub098\uc5f4\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4. (\ubcf8 \uc790\ub8cc\ub294 AWS\uc758 \uacf5\uc2dd\uc801\uc778 \uc790\ub8cc\uac00 \uc544\ub2c8\uba70, DeepLearning 4J\uc758 \uc758\uacac\uc785\ub2c8\ub2e4.) \ud504\ub808\uc784\uc6cc\ud06c \uc7a5\uc810 \ub2e8\uc810 Theano Python \uc9c0\uc6d0 Wrapper \ub97c \ud1b5\ud55c \ub192\uc740 \ucd94\uc0c1\ud654\ub85c \uc0ac\uc6a9\uc131 \ud3b8\ub9ac <a class=\"mh-excerpt-more\" href=\"https:\/\/jirak.net\/wp\/aws%eb%a1%9c-%eb%94%a5-%eb%9f%ac%eb%8b%9d%ec%9d%84-%ec%9c%84%ed%95%9c-%ed%94%84%eb%a0%88%ec%9e%84%ec%9b%8c%ed%81%ac-mxnet-%ed%99%9c%ec%9a%a9%ed%95%98%ea%b8%b0\/\" title=\"AWS\ub85c \ub525 \ub7ec\ub2dd\uc744 \uc704\ud55c \ud504\ub808\uc784\uc6cc\ud06c MxNet \ud65c\uc6a9\ud558\uae30\">[ more&#8230; ]<\/a><\/p>\n<\/div>","protected":false},"author":1,"featured_media":12183,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[169],"tags":[656],"class_list":["post-12182","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-news","tag-aws"],"amp_enabled":true,"_links":{"self":[{"href":"https:\/\/jirak.net\/wp\/wp-json\/wp\/v2\/posts\/12182","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/jirak.net\/wp\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/jirak.net\/wp\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/jirak.net\/wp\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/jirak.net\/wp\/wp-json\/wp\/v2\/comments?post=12182"}],"version-history":[{"count":1,"href":"https:\/\/jirak.net\/wp\/wp-json\/wp\/v2\/posts\/12182\/revisions"}],"predecessor-version":[{"id":12184,"href":"https:\/\/jirak.net\/wp\/wp-json\/wp\/v2\/posts\/12182\/revisions\/12184"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/jirak.net\/wp\/wp-json\/wp\/v2\/media\/12183"}],"wp:attachment":[{"href":"https:\/\/jirak.net\/wp\/wp-json\/wp\/v2\/media?parent=12182"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/jirak.net\/wp\/wp-json\/wp\/v2\/categories?post=12182"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/jirak.net\/wp\/wp-json\/wp\/v2\/tags?post=12182"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}