{"id":31378,"date":"2019-05-31T18:36:26","date_gmt":"2019-05-31T09:36:26","guid":{"rendered":"https:\/\/jirak.net\/wp\/gpu-%eb%94%a5%eb%9f%ac%eb%8b%9d-%eb%aa%a8%eb%8d%b8-%ed%95%99%ec%8a%b5%ec%9d%84-%ec%9c%84%ed%95%9c-amazon-ec2-%ec%8a%a4%ed%8c%9f-%ec%9d%b8%ec%8a%a4%ed%84%b4%ec%8a%a4-%ed%99%9c%ec%9a%a9%eb%b2%95\/"},"modified":"2019-05-31T19:34:32","modified_gmt":"2019-05-31T10:34:32","slug":"gpu-%eb%94%a5%eb%9f%ac%eb%8b%9d-%eb%aa%a8%eb%8d%b8-%ed%95%99%ec%8a%b5%ec%9d%84-%ec%9c%84%ed%95%9c-amazon-ec2-%ec%8a%a4%ed%8c%9f-%ec%9d%b8%ec%8a%a4%ed%84%b4%ec%8a%a4-%ed%99%9c%ec%9a%a9%eb%b2%95","status":"publish","type":"post","link":"https:\/\/jirak.net\/wp\/gpu-%eb%94%a5%eb%9f%ac%eb%8b%9d-%eb%aa%a8%eb%8d%b8-%ed%95%99%ec%8a%b5%ec%9d%84-%ec%9c%84%ed%95%9c-amazon-ec2-%ec%8a%a4%ed%8c%9f-%ec%9d%b8%ec%8a%a4%ed%84%b4%ec%8a%a4-%ed%99%9c%ec%9a%a9%eb%b2%95\/","title":{"rendered":"GPU \ub525\ub7ec\ub2dd \ubaa8\ub378 \ud559\uc2b5\uc744 \uc704\ud55c Amazon EC2 \uc2a4\ud31f \uc778\uc2a4\ud134\uc2a4 \ud65c\uc6a9\ubc95"},"content":{"rendered":"<p>GPU \ub525\ub7ec\ub2dd \ubaa8\ub378 \ud559\uc2b5\uc744 \uc704\ud55c Amazon EC2 \uc2a4\ud31f \uc778\uc2a4\ud134\uc2a4 \ud65c\uc6a9\ubc95<br \/>\n<img loading=\"lazy\" decoding=\"async\" src=\"https:\/\/jirak.net\/wp\/wp-content\/uploads\/2019\/05\/ec2-spot-instances-1-1.gif\" width=\"800\" height=\"429\"><\/p>\n<p>\uc5ec\ub7ec\ubd84\uc774 \ub370\uc774\ud130 \uc138\ud2b8\ub97c \uc218\uc9d1\ud558\uace0, \uc2ec\uce35 \uc2e0\uacbd\ub9dd \uc544\ud0a4\ud14d\ucc98\ub97c \uc124\uacc4\ud558\uace0, \ud559\uc2b5 \ub8e8\ud2f4 \ucf54\ub4dc\ub97c \uc791\uc131\uc744 \uc644\ub8cc \ud558\uc168\ub2e4\uba74 \uc9c0\uae08\ubd80\ud130\ub294 \uac15\ub825\ud55c GPU \uc778\uc2a4\ud134\uc2a4\ub97c \uc0ac\uc6a9\ud574 \uc5ec\ub7ec epoch\uc5d0 \uac78\uccd0 \ub300\uaddc\ubaa8 \ub370\uc774\ud130 \uc138\ud2b8\ub97c \uc774\uc6a9\ud55c \ud559\uc2b5\uc744 \uc2e4\ud589\ud560 \ucc28\ub840\uc785\ub2c8\ub2e4. \uc5ec\ub7ec\ubd84\uc740 \uc774\ubbf8 NVIDIA Tesla V100 GPU\uc5d0 \uae30\ubc18\ud55c Amazon EC2 P3 \uc778\uc2a4\ud134\uc2a4\uac00 \ucef4\ud4e8\ud305 \uc9d1\uc57d\uc801\uc778 \ub525\ub7ec\ub2dd \ud559\uc2b5 \uc791\uc5c5\uc5d0 \uc801\ud569\ud558\ub2e4\ub294 \uc0ac\uc2e4\uc744 \uc54c\uace0 \uc788\uc9c0\ub9cc&nbsp;\uc608\uc0b0\uc774 \ube60\ub4ef\ud55c \uad00\uacc4\ub85c \uac00\ub2a5\ud55c\ud55c \ud559\uc2b5 \ube44\uc6a9\uc740 \ub0ae\ucd94\uace0 \uc2f6\uc2b5\ub2c8\ub2e4.<\/p>\n<p>\uba87 \uc2dc\uac04 \ub610\ub294 \uba70\uce60\uc5d0 \uac78\uccd0 \ud559\uc2b5 \uc791\uc5c5\uc744 \uc2e4\ud589\ud558\ub294 \ub525\ub7ec\ub2dd \uc5f0\uad6c\uc790 \ubc0f \uac1c\ubc1c\uc790\ub77c\uba74, \uc2a4\ud31f \uc778\uc2a4\ud134\uc2a4 \uc694\uae08\uc744 \ud1b5\ud574 \ud6e8\uc52c \uacbd\uc81c\uc801\uc778 \ube44\uc6a9\uc73c\ub85c \uace0\uc131\ub2a5 CPU\ub97c \uc774\uc6a9\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \uc2a4\ud31f \uc778\uc2a4\ud134\uc2a4\ub97c \uc0ac\uc6a9\ud558\uba74 \uc628\ub514\ub9e8\ub4dc \uc694\uae08\uc5d0 \ube44\ud574 \ub300\ud3ed \ud560\uc778\ub41c \uac00\uaca9\uc73c\ub85c \uc5ec\ubd84\uc758 Amazon EC2 \ucef4\ud4e8\ud305 \uc6a9\ub7c9\uc5d0 \uc561\uc138\uc2a4\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \uc778\uc2a4\ud134\uc2a4 \ubc0f \ub9ac\uc804\ubcc4 \ucd5c\uadfc \uc694\uae08 \ubaa9\ub85d\uc744 \ubcf4\ub824\uba74 <a href=\"https:\/\/aws.amazon.com\/ec2\/spot\/bid-advisor\/\" target=\"_blank\" rel=\"noopener noreferrer\">\uc2a4\ud31f \uc778\uc2a4\ud134\uc2a4 \uc5b4\ub4dc\ubc14\uc774\uc800<\/a>\ub97c \ubc29\ubb38\ud558\uc2ed\uc2dc\uc624. \uc2a4\ud31f \uc778\uc2a4\ud134\uc2a4\uc640 \uc628\ub514\ub9e8\ub4dc \uc778\uc2a4\ud134\uc2a4 \uc0ac\uc774\uc758 \uc8fc\uc694 \ucc28\uc774\uc810\uc744 \uc54c\uc544\ubcf4\ub824\uba74 <a href=\"https:\/\/docs.aws.amazon.com\/AWSEC2\/latest\/UserGuide\/using-spot-instances.html\" target=\"_blank\" rel=\"noopener noreferrer\">\uc774 Amazon EC2 \uc0ac\uc6a9 \uc124\uba85\uc11c<\/a>\ub97c \uac80\ud1a0\ud574 \ubcf4\uc2dc\uae30 \ubc14\ub78d\ub2c8\ub2e4.<\/p>\n<p>\uc2a4\ud31f \uc778\uc2a4\ud134\uc2a4\ub294 \ub525\ub7ec\ub2dd \uc6cc\ud06c\ud50c\ub85c\uc6b0\uc5d0 \uc801\ud569\ud558\uc9c0\ub9cc, \uc628\ub514\ub9e8\ub4dc \uc778\uc2a4\ud134\uc2a4\uc640 \ube44\uad50\ud588\uc744 \ub54c \uc2a4\ud31f \uc778\uc2a4\ud134\uc2a4\ub97c \uc0ac\uc6a9\ud560 \uacbd\uc6b0 \uba87 \uac00\uc9c0 \ubb38\uc81c\uac00 \uc788\uc2b5\ub2c8\ub2e4. \uccab \ubc88\uc9f8\ub294, \uc2a4\ud31f \uc778\uc2a4\ud134\uc2a4\ub294 \ubbf8\ub9ac \ud655\ubcf4\ub420 \uc218 \uc788\uc73c\uba70, \ubd88\uacfc2\ubd84 \uc804\uc5d0 \ud1b5\uc9c0\ud558\uc5ec \uc885\ub8cc\ub420 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \uc989, \uc778\uc2a4\ud134\uc2a4\uc5d0\uc11c \ud559\uc2b5 \uc791\uc5c5\uc744 \uc644\ub8cc \ub2e8\uacc4\uae4c\uc9c0 \uc2e4\ud589\ud560 \uc218 \uc788\uc744\uc9c0\uc5d0 \ub300\ud574 \uc2e0\ub8b0\ud560 \uc218 \uc5c6\uc2b5\ub2c8\ub2e4. \ub530\ub77c\uc11c \uc2dc\uac04\uc5d0 \ubbfc\uac10\ud55c \uc6cc\ud06c\ub85c\ub4dc\uc5d0 \ub300\ud574\uc11c\ub294 \uad8c\uc7a5\ub418\uc9c0 \uc54a\uc2b5\ub2c8\ub2e4. \ub450 \ubc88\uc9f8\ub294, \ud559\uc2b5 \uc9c4\ud589 \uc0c1\ud669\uc774 \uc801\uc808\ud788 \uc800\uc7a5\ub418\uc9c0 \uc54a\uc740 \uacbd\uc6b0 \uc778\uc2a4\ud134\uc2a4 \uc885\ub8cc\ub85c \uc778\ud574 \ub370\uc774\ud130\uac00 \uc190\uc2e4\ub420 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \uc138 \ubc88\uc9f8\ub294, \uc2a4\ud31f \uc778\uc2a4\ud134\uc2a4\ub97c \uc2dc\uc791\ud55c \ud6c4 \uc560\ud50c\ub9ac\ucf00\uc774\uc158\uc744 \uc911\ub2e8\ud558\uc9c0 \uc54a\ub3c4\ub85d \uacb0\uc815\ud55c \uacbd\uc6b0 \uc0ac\uc6a9\uc790\uc758 \uc720\uc77c\ud55c \uc635\uc158\uc740 \uc2a4\ud31f \uc778\uc2a4\ud134\uc2a4\ub97c \uc911\uc9c0\ud558\uace0 \uc628\ub514\ub9e8\ub4dc\ub098 \uc608\uc57d \uc778\uc2a4\ud134\uc2a4\ub85c \ub2e4\uc2dc \uc2dc\uc791\ud558\ub294 \uac83\ubfd0\uc785\ub2c8\ub2e4.<\/p>\n<p>\uc774\ub7ec\ud55c \ubb38\uc81c\ub97c \ud574\uacb0\ud558\uae30 \uc704\ud574 \uc5ec\uae30\uc5d0\uc11c\ub294 \uc2a4\ud31f \uc911\ub2e8\uc774 \ubc1c\uc0dd\ud55c \uacbd\uc6b0 \ud559\uc2b5 \uc9c4\ud589 \uc0c1\ud669 \uc190\uc2e4\uc744 \ucd5c\uc18c\ud654\ud558\uba74\uc11c \ub525\ub7ec\ub2dd \ud559\uc2b5 \uc6cc\ud06c\ud50c\ub85c\uc6b0\ub97c \uc704\ud574 \uc2a4\ud31f \uc778\uc2a4\ud134\uc2a4\ub97c \uc124\uc815\ud558\ub294 \ubc29\ubc95\uc5d0 \ub300\ud55c \ub2e8\uacc4\ubcc4 \uc790\uc2b5\uc11c\ub97c \uc81c\uacf5\ud569\ub2c8\ub2e4. \uc81c \ubaa9\ud45c\ub294 \ub2e4\uc74c \ud2b9\uc131\uc744 \ubc18\uc601\ud55c \uc124\uc815\uc744 \uad6c\ud604\ud558\ub294 \uac83\uc785\ub2c8\ub2e4.<\/p>\n<ul>\n<li><strong>\ucef4\ud4e8\ud305, \uc2a4\ud1a0\ub9ac\uc9c0, \ucf54\ub4dc \uc544\ud2f0\ud329\ud2b8\ub97c \ubd84\ub9ac\ud558\uace0, \ucef4\ud4e8\ud305 \uc778\uc2a4\ud134\uc2a4\uc5d0 \uc0c1\ud0dc\ub97c \uc800\uc7a5\ud558\uc9c0 \uc54a\uc2b5\ub2c8\ub2e4<\/strong>. \uadf8\ub7ec\uba74 \uc778\uc2a4\ud134\uc2a4\uac00 \uc885\ub8cc\ub418\uace0 \ub300\uccb4\ub41c \uacbd\uc6b0\uc5d0\ub3c4 \uac04\ub2e8\ud55c \ubcf5\uad6c\uc640 \ud559\uc2b5 \uc0c1\ud0dc \ubcf5\uc6d0\uc744 \ubcf4\uc7a5\ud569\ub2c8\ub2e4.<\/li>\n<li><strong>\ub370\uc774\ud130 \uc138\ud2b8, \ud559\uc2b5 \uc9c4\ud589 \uc0c1\ud669(\uccb4\ud06c\ud3ec\uc778\ud2b8) \ubc0f \ub85c\uadf8\uc5d0 \ub300\ud574 \uc804\uc6a9 \ubcfc\ub968\uc744 \uc0ac\uc6a9\ud569\ub2c8\ub2e4<\/strong>. \uc774 \ubcfc\ub968\uc740 \uc601\uad6c \ubcfc\ub968\uc774\uba70, \uc778\uc2a4\ud134\uc2a4 \uc885\ub8cc\ub85c \uc601\ud5a5\uc744 \ubc1b\uc9c0 \uc54a\uc544\uc57c \ud569\ub2c8\ub2e4.<\/li>\n<li><strong>\ud559\uc2b5 \ucf54\ub4dc\uc5d0 \ub300\ud574 \ubc84\uc804 \uc81c\uc5b4 \uc2dc\uc2a4\ud15c(\uc608: Git)\uc744 \uc0ac\uc6a9\ud569\ub2c8\ub2e4<\/strong>. \ud559\uc2b5\uc744 \uc2dc\uc791\/\uc7ac\uac1c\ud558\ub824\uba74 \uc774 \ub9ac\ud3ec\uc9c0\ud1a0\ub9ac\ub97c \ubcf5\uc81c\ud574\uc57c \ud569\ub2c8\ub2e4. \uadf8\ub7ec\uba74 \uc778\uc2a4\ud134\uc2a4\uac00 \uc885\ub8cc\ub418\uc5b4\ub3c4 \ucd94\uc801 \uae30\ub2a5\uc744 \ud1b5\ud574 \ucf54\ub4dc \ubcc0\uacbd \uc0ac\ud56d\uc774 \uc190\uc2e4\ub418\uc9c0 \uc54a\ub3c4\ub85d \ubc29\uc9c0\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4.<\/li>\n<li><strong>\ud559\uc2b5 \uc2a4\ud06c\ub9bd\ud2b8\uc5d0 \ub300\ud55c \ucf54\ub4dc \ubcc0\uacbd\uc744 \ucd5c\uc18c\ud654\ud569\ub2c8\ub2e4<\/strong>. \uc774\ub97c \ud1b5\ud574 \ud559\uc2b5 \uc2a4\ud06c\ub9bd\ud2b8\ub97c \ub3c5\ub9bd\uc801\uc73c\ub85c \uac1c\ubc1c\ud558\uace0, \ud559\uc2b5 \ucf54\ub4dc \uc678\ubd80\uc5d0\uc11c \ubc31\uc5c5 \ubc0f \uc2a4\ub0c5\uc0f7 \uc791\uc5c5\uc744 \uc218\ud589\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4.<\/li>\n<li><strong>\ubaa8\ub4e0 \uc791\uc5c5\uc744 \uc790\ub3d9\ud654\ud569\ub2c8\ub2e4. <\/strong>\uc885\ub8cc \ud6c4 \ub300\uccb4 \uc778\uc2a4\ud134\uc2a4 \uc0dd\uc131, \uc2dc\uc791 \uc2dc \ub370\uc774\ud130 \uc138\ud2b8 \ubc0f \uccb4\ud06c\ud3ec\uc778\ud2b8 EBS \ubcfc\ub968\uc758 \uc5f0\uacb0, \uac00\uc6a9 \uc601\uc5ed\uc5d0\uc11c \ubcfc\ub968 \uc774\ub3d9, \uc778\uc2a4\ud134\uc2a4 \uc0c1\ud0dc \ubcf5\uc6d0 \uc218\ud589, \ud559\uc2b5 \uc7ac\uac1c, \ud559\uc2b5 \uc885\ub8cc \uc2dc \uc778\uc2a4\ud134\uc2a4 \uc885\ub8cc \uc791\uc5c5\uc744 \uc790\ub3d9\ud654\ud569\ub2c8\ub2e4.<\/li>\n<\/ul>\n<h2>TensorFlow \ubc0f AWS Deep Learning AMI\ub97c \uc0ac\uc6a9\ud558\uc5ec \uc2a4\ud31f \uc778\uc2a4\ud134\uc2a4\uc5d0\uc11c \ub525\ub7ec\ub2dd \uc218\ud589<\/h2>\n<p>\uc774 \uc608\uc81c\uc5d0\uc11c\ub294 \uc2a4\ud31f \uc778\uc2a4\ud134\uc2a4\uc640 AWS Deep Learning AMI\ub97c \uc0ac\uc6a9\ud558\uc5ec <a href=\"https:\/\/www.cs.toronto.edu\/~kriz\/cifar.html\" target=\"_blank\" rel=\"noopener noreferrer\">CIFAR10 \ub370\uc774\ud130 \uc138\ud2b8<\/a>\uc5d0\uc11c <a href=\"https:\/\/arxiv.org\/abs\/1512.03385\" target=\"_blank\" rel=\"noopener noreferrer\">ResNet50 \ubaa8\ub378<\/a>\uc744 \ud559\uc2b5\ud569\ub2c8\ub2e4.&nbsp;\uc800\ub294 AWS Deep Learning AMI \ubc84\uc804 21\uc5d0\uc11c \uc0ac\uc6a9 \uac00\ub2a5\ud55c CUDA 9\ub85c \uad6c\uc131\ub41c TensorFlow 1.12\ub97c \uc0ac\uc6a9\ud569\ub2c8\ub2e4. AWS Deep Learning AMI\ub294 \uc790\uc8fc \uc5c5\ub370\uc774\ud2b8\ub418\ubbc0\ub85c, \uba3c\uc800 <a href=\"http:\/\/aws.amazon.com\/partners\/aws-marketplace\" target=\"_blank\" rel=\"noopener noreferrer\">AWS Marketplace<\/a>\uc5d0\uc11c \ud559\uc2b5 \ucf54\ub4dc\uc640 \ud638\ud658 \uac00\ub2a5\ud55c \ucd5c\uc2e0 \ubc84\uc804\uc744 \uc0ac\uc6a9\ud558\ub294\uc9c0 \ud655\uc778\ud558\uc2ed\uc2dc\uc624. TensorFlow 1.13 \ubc0f CUDA 10\uc758 \uacbd\uc6b0 <a href=\"https:\/\/aws.amazon.com\/marketplace\/pp\/B077GCH38C\" target=\"_blank\" rel=\"noopener noreferrer\"> \uc774 AWS Deep Learning AMI<\/a>\ub97c \uc0ac\uc6a9\ud569\ub2c8\ub2e4.<\/p>\n<p>\uc800\ub294 \uc5ec\uae30\uc5d0\uc11c \ub525\ub7ec\ub2dd \ud559\uc2b5 \uc791\uc5c5\uc5d0 \ub300\ud55c \uc2a4\ud31f \ud50c\ub9bf \uc694\uccad\uc744 \uc124\uc815\ud558\ub294 \ubc29\ubc95\uc744 \ubcf4\uc5ec\ub4dc\ub9b4 \uac83\uc785\ub2c8\ub2e4. \uadf8\ub7ec\uba74 \uc0ac\uc6a9\uc790\uc758 \ud2b9\uc815 \ub370\uc774\ud130 \uc138\ud2b8 \ubc0f \ubaa8\ub378\uc5d0 \ub9de\ub294 \ubc29\ubc95\uc744 \uad6c\ud604\ud560 \ub54c \uc774 \ubc29\ubc95\uc744 \uc2dc\uc791\uc810\uc73c\ub85c \ud65c\uc6a9\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n<p>\uc774 \uacfc\uc815\uc5d0\uc11c\ub294 \uc0ac\uc6a9\uc790\uac00 \ub2e4\uc74c \uc804\uc81c \uc870\uac74\uc744 \ucda9\uc871\ud588\ub2e4\uace0 \uac00\uc815\ud569\ub2c8\ub2e4.<\/p>\n<ol>\n<li><a href=\"https:\/\/aws.amazon.com\/free\/\" target=\"_blank\" rel=\"noopener noreferrer\">AWS \uacc4\uc815<\/a>\uc774 \uc788\uace0 \ud638\uc2a4\ud2b8\uc5d0 <a href=\"https:\/\/aws.amazon.com\/cli\/\" target=\"_blank\" rel=\"noopener noreferrer\">AWS CLI \ub3c4\uad6c<\/a>\uac00 \uc124\uce58\ub418\uc5b4 \uc788\uc74c<\/li>\n<li>Python\uacfc \ud558\ub098 \uc774\uc0c1\uc758 \ub525\ub7ec\ub2dd \ud504\ub808\uc784\uc6cc\ud06c\uc5d0 \uce5c\uc219\ud568<\/li>\n<\/ol>\n<p>\uad6c\ud604 \uc138\ubd80 \uc815\ubcf4\ub97c \uac80\ud1a0\ud558\uba74 \ud544\uc694\ud55c \ubaa8\ub4e0 \uc815\ubcf4\ub97c \ubc30\uc6b8 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \ubaa8\ub4e0 \ucf54\ub4dc, \uad6c\uc131 \ud30c\uc77c \ubc0f AWS CLI \uba85\ub839\uc740 <a href=\"https:\/\/github.com\/awslabs\/ec2-spot-labs\/tree\/master\/ec2-spot-deep-learning-training\" target=\"_blank\" rel=\"noopener noreferrer\">GitHub\uc5d0\uc11c \uc0ac\uc6a9 \uac00\ub2a5<\/a>\ud569\ub2c8\ub2e4.<\/p>\n<p>\ub2e4\uc74c\uacfc \uac19\uc740 AWS \ubc0f \uc624\ud508 \uc18c\uc2a4 \uc11c\ube44\uc2a4\/\uac1c\ub150\uc744 \uc0ac\uc6a9\ud569\ub2c8\ub2e4. \uadf8\ub9bc 1\uc5d0\uc11c\ub294 \uc608\uc81c\uc5d0\uc11c \uc774 \ubaa8\ub4e0 \uc694\uc18c\ub97c \uc801\uc808\ud788 \uc870\ud569\ud574 \uc0ac\uc6a9\ud558\ub294 \ubc29\ubc95\uc744 \ubcf4\uc5ec\uc90d\ub2c8\ub2e4.<\/p>\n<ul>\n<li><a href=\"https:\/\/aws.amazon.com\/cli\/\" target=\"_blank\" rel=\"noopener noreferrer\">AWS CLI<\/a>: CLI\ub97c \uc0ac\uc6a9\ud558\uc5ec AWS \uc11c\ube44\uc2a4\uc640 \uc0c1\ud638 \uc791\uc6a9\ud569\ub2c8\ub2e4. CLI\uc5d0\uc11c \uc218\ud589\ud560 \uc218 \uc788\ub294 \ubaa8\ub4e0 \uc791\uc5c5\uc740 AWS \ucf58\uc194\uc5d0\uc11c\ub3c4 \uc218\ud589\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4. CLI\ub97c \uc0ac\uc6a9\ud558\uba74 \uc790\ub3d9\ud654\uac00 \uac00\ub2a5\ud558\uace0, \ubc14\ub85c \uc774 \uc810\uc774 \uc774 \uc608\uc81c\uc5d0\uc11c \uc81c\uac00 \ucd94\uad6c\ud558\ub294 \ubaa9\ud45c\uc774\uae30\ub3c4 \ud569\ub2c8\ub2e4.<\/li>\n<li><a href=\"https:\/\/docs.aws.amazon.com\/AWSEC2\/latest\/UserGuide\/spot-requests.html\" target=\"_blank\" rel=\"noopener noreferrer\">Amazon EC2 \uc2a4\ud31f \uc778\uc2a4\ud134\uc2a4 \ubc0f \uc2a4\ud31f \uc778\uc2a4\ud134\uc2a4 \uc694\uccad<\/a>: \uc2a4\ud31f \uc694\uccad\uc744 \ud1b5\ud574 \uc9c0\uc815\ub41c \uc2a4\ud31f \uc778\uc2a4\ud134\uc2a4\ub9cc \uc2e4\ud589\ub418\ub3c4\ub85d \ud569\ub2c8\ub2e4. \uc2a4\ud31f \ud50c\ub9bf\uc740 \ub300\uc0c1 \uc6a9\ub7c9\uc744 \ucda9\uc871\ud558\ub3c4\ub85d \uc2a4\ud31f \uc694\uccad\uc744 \uc81c\ucd9c\ud558\uace0 \uc911\ub2e8\ub41c \uc778\uc2a4\ud134\uc2a4\ub97c \uc790\ub3d9\uc73c\ub85c \ubcf4\ucda9\ud569\ub2c8\ub2e4.<\/li>\n<li><a href=\"https:\/\/docs.aws.amazon.com\/dlami\/latest\/devguide\/what-is-dlami.html\" target=\"_blank\" rel=\"noopener noreferrer\">AWS Deep Learning AMI<\/a>: \ub525\ub7ec\ub2dd \ud504\ub808\uc784\uc6cc\ud06c\uac00 \ubbf8\ub9ac \uc124\uce58\ub41c Amazon Machine Image\uc785\ub2c8\ub2e4. \uc774 \uc608\uc81c\uc5d0\uc11c\ub294 \ud559\uc2b5\uc744 \uc704\ud574 GPU \uac00\uc18d\uc744 \uc9c0\uc6d0\ud558\ub294 TensorFlow \ud504\ub808\uc784\uc6cc\ud06c\ub97c \uc0ac\uc6a9\ud569\ub2c8\ub2e4.<\/li>\n<li><a href=\"https:\/\/docs.aws.amazon.com\/AWSEC2\/latest\/UserGuide\/AmazonEBS.html\" target=\"_blank\" rel=\"noopener noreferrer\">Amazon Elastic Block Storage(EBS)<\/a>: \ub370\uc774\ud130 \uc138\ud2b8, \uccb4\ud06c\ud3ec\uc778\ud2b8 \ubc0f \ub85c\uadf8\ub97c \uc800\uc7a5\ud560 \uc601\uad6c \ubcfc\ub968\uc785\ub2c8\ub2e4. \uc774 \ubcfc\ub968\uc740 \ud604\uc7ac \uc2e4\ud589 \uc911\uc778 \uc778\uc2a4\ud134\uc2a4\uc5d0 \uc5f0\uacb0\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4.<\/li>\n<li><a href=\"https:\/\/docs.aws.amazon.com\/AWSEC2\/latest\/UserGuide\/EBSSnapshots.html\" target=\"_blank\" rel=\"noopener noreferrer\">Amazon EBS \uc2a4\ub0c5\uc0f7<\/a>: \uc2a4\ub0c5\uc0f7\uc744 \uc0ac\uc6a9\ud558\uba74 Amazon EBS \ubcfc\ub968\uc758 \ub370\uc774\ud130\ub97c Amazon S3\uc5d0 \ubc31\uc5c5\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \uc2a4\ub0c5\uc0f7\uc740 \uc0c8 EBS \ubcfc\ub968\uc73c\ub85c \ub370\uc774\ud130\ub97c \ubcf5\uc6d0\ud558\ub294 \ub370 \ud544\uc694\ud55c \ubaa8\ub4e0 \uc815\ubcf4\ub97c \ud3ec\ud568\ud558\uba70, \uc0c8 \uac00\uc6a9 \uc601\uc5ed\uc73c\ub85c \ubcfc\ub968\uc744 \ub9c8\uc774\uadf8\ub808\uc774\uc158\ud558\ub294 \ub370 \uc0ac\uc6a9\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4.<\/li>\n<li><a href=\"https:\/\/docs.aws.amazon.com\/AWSEC2\/latest\/UserGuide\/ec2-instance-metadata.html\" target=\"_blank\" rel=\"noopener noreferrer\">Amazon EC2 \uc0ac\uc6a9\uc790 \ub370\uc774\ud130 \ubc0f \uc778\uc2a4\ud134\uc2a4 \uba54\ud0c0\ub370\uc774\ud130<\/a>: \uc778\uc2a4\ud134\uc2a4\uac00 \uc2dc\uc791\ub418\uba74 \ubcfc\ub968 \uc5f0\uacb0, \ud559\uc2b5 \uc2dc\uc791 \ubc0f \uc815\ub9ac\uc640 \uac19\uc740 \uc791\uc5c5\uc744 \uc218\ud589\ud558\uae30 \uc704\ud574 \uc0ac\uc6a9\uc790 \ub370\uc774\ud130 \uc258 \uc2a4\ud06c\ub9bd\ud2b8\ub97c \uc2e4\ud589\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \uc778\uc2a4\ud134\uc2a4 \uba54\ud0c0\ub370\uc774\ud130\ub97c \ud1b5\ud574 \uc778\uc2a4\ud134\uc2a4\ub294 \uc0ac\uc6a9\uc790 \ub370\uc774\ud130 \uc258 \uc2a4\ud06c\ub9bd\ud2b8\uc640 \ud568\uaed8 \uc0ac\uc6a9\ud560 \uc778\uc2a4\ud134\uc2a4 ID\uc640 \uac19\uc774 \ud574\ub2f9 \uc778\uc2a4\ud134\uc2a4\uc5d0 \ub300\ud55c \uc790\uccb4 \uc815\ubcf4\ub97c \ucffc\ub9ac\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4.<\/li>\n<li><a href=\"https:\/\/docs.aws.amazon.com\/AWSEC2\/latest\/UserGuide\/iam-roles-for-amazon-ec2.html\" target=\"_blank\" rel=\"noopener noreferrer\">Amazon IAM \uc5ed\ud560 \ubc0f \uc815\ucc45<\/a>: \uc790\ub3d9\uc73c\ub85c AWS \uc11c\ube44\uc2a4\ub97c \uc0ac\uc6a9\ud558\ub3c4\ub85d EC2 \uc778\uc2a4\ud134\uc2a4 \uad8c\ud55c\uc744 \ubd80\uc5ec\ud569\ub2c8\ub2e4. \ubaa8\ub4e0 \uac83\uc744 \uc790\ub3d9\ud654\ud558\ub294 \uac83\uc774 \ud575\uc2ec\uc785\ub2c8\ub2e4.<\/li>\n<\/ul>\n<div id=\"attachment_8039\" class=\"wp-caption alignnone\" style=\"width: 810px\">\n<p><img loading=\"lazy\" decoding=\"async\" class=\"size-full wp-image-8039\" src=\"https:\/\/d2908q01vomqb2.cloudfront.net\/f1f836cb4ea6efb2a0b1b99f41ad8b103eff4b59\/2019\/03\/26\/ec2-spot-instances-1.gif\" alt=\"\" width=\"800\" height=\"429\" \/><\/p>\n<p class=\"wp-caption-text\">\uadf8\ub9bc 1: \ub525\ub7ec\ub2dd \uc6cc\ud06c\ud50c\ub85c\uc6b0\uc5d0\uc11c \uc2a4\ud31f \uc778\uc2a4\ud134\uc2a4 \uc0ac\uc6a9\uc5d0 \ub300\ud55c \ucc38\uc870 \uc544\ud0a4\ud14d\ucc98<\/p>\n<\/div>\n<h3>1\ub2e8\uacc4: \ubc94\uc6a9 \uc778\uc2a4\ud134\uc2a4\ub97c \uc0ac\uc6a9\ud558\uc5ec \ub370\uc774\ud130 \uc138\ud2b8 \ubc0f \uccb4\ud06c\ud3ec\uc778\ud2b8\uc5d0 \ub300\ud55c \uc804\uc6a9 EBS \ubcfc\ub968 \uc124\uc815<\/h3>\n<p>\uccab \ubc88\uc9f8 \ub2e8\uacc4\ub294 \ub85c\uadf8 \ubc0f \uae30\ud0c0 \uba54\ud0c0\ub370\uc774\ud130\uc640 \uac19\uc774 \uc601\uad6c\uc801\uc774\uc5b4\uc57c \ud558\ub294 \ub370\uc774\ud130 \uc138\ud2b8, \uccb4\ud06c\ud3ec\uc778\ud2b8 \ubc0f \uae30\ud0c0 \uc815\ubcf4\ub97c \uc800\uc7a5\ud560 \uc804\uc6a9 EBS \ubcfc\ub968\uc744 \uc124\uc815\ud558\ub294 \uac83\uc785\ub2c8\ub2e4. \uc774 \ub2e8\uacc4\ub294 \ud55c \ubc88\ub9cc \uc218\ud589\ub429\ub2c8\ub2e4. \uadf8\ub798\uc11c \ucc98\uc74c\uc5d0\ub294 \uc628\ub514\ub9e8\ub4dc <code class=\"lang-code\">m4.xlarge<\/code> \uc778\uc2a4\ud134\uc2a4\ub85c \uc2dc\uc791\ud569\ub2c8\ub2e4. \ub370\uc774\ud130 \uc138\ud2b8\uac00 \uc791\uace0 \uc900\ube44 \uc911\uc5d0 \uc0ac\uc804 \ucc98\ub9ac \ub2e8\uacc4\ub85c \uc218\ud589\ud558\uc9c0 \uc54a\uc73c\ub824\ub294 \uacbd\uc6b0 \uba54\ubaa8\ub9ac\uc640 \uc131\ub2a5\uc744 \ub354 \uc801\uac8c \uc18c\ube44\ud558\uba70(\ube44\uc6a9\uc774 \ub354 \uc904\uc5b4\ub4e4 \uc218 \uc788\uc74c) \uc778\uc2a4\ud134\uc2a4\ub97c \uc2dc\uc791\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \uc774\ubbf8\uc9c0\ub97c \ud2b8\ub79c\uc2a4\ucf54\ub529\ud558\uac70\ub098 \uae30\ud0c0 \uba40\ud2f0\uc2a4\ub808\ub4dc \uc0ac\uc804 \ucc98\ub9ac \ub8e8\ud2f4\uc744 \uc2e4\ud589\ud558\ub824\ub294 \uacbd\uc6b0 GPU \uc9c0\uc6d0 \ub610\ub294 \ucef4\ud4e8\ud305\uc5d0 \ucd5c\uc801\ud654\ub41c CPU \uc778\uc2a4\ud134\uc2a4\ub97c \uc120\ud0dd\ud569\ub2c8\ub2e4.<\/p>\n<p>AWS CLI\ub97c \uc0ac\uc6a9\ud558\uc5ec \ud130\ubbf8\ub110\uc5d0\uc11c \ub2e4\uc74c \uba85\ub839\uc744 \uc2e4\ud589\ud569\ub2c8\ub2e4. \uc544\ub798 \ub098\uc628 \ubaa8\ub4e0 \uba85\ub839\uc740 MacOS\uc5d0\uc11c \ud14c\uc2a4\ud2b8\ub418\uc5c8\uc2b5\ub2c8\ub2e4.<\/p>\n<div class=\"hide-language\">\n<pre><code class=\"lang-bash\">aws ec2 run-instances \n--image-id ami-0027dfad6168539c7 \n--security-group-ids &lt;SECURITY_GROUP_ID&gt; \n--count 1 \n--instance-type m4.xlarge \n--key-name &lt;KEYPAIR_NAME&gt; \n--subnet-id &lt;SUBNET_ID&gt; \n--query &quot;Instances[0].InstanceId&quot;\n<\/code><\/pre>\n<\/div>\n<p><code class=\"lang-code\">image-id<\/code>\ub294 Deep Learning AMI Ubuntu \uc778\uc2a4\ud134\uc2a4\ub97c \ub098\ud0c0\ub0c5\ub2c8\ub2e4. \uc778\uc2a4\ud134\uc2a4\uc5d0 \ub300\ud55c SSH \uc5f0\uacb0\uc744 \ud5c8\uc6a9\ud558\ub824\uba74 \ubcf4\uc548 \uadf8\ub8f9, \ud0a4 ID \ubc0f \uc11c\ube0c\ub137 ID\ub97c \uc5c5\ub370\uc774\ud2b8\ud574\uc57c \ud569\ub2c8\ub2e4. \uc790\uc138\ud55c \ub0b4\uc6a9\uc740 \uc774 <a href=\"https:\/\/docs.aws.amazon.com\/AWSEC2\/latest\/UserGuide\/AccessingInstancesLinux.html\" target=\"_blank\" rel=\"noopener noreferrer\">\ubb38\uc11c \ud398\uc774\uc9c0<\/a>\ub97c \ucc38\uc870\ud558\uc2ed\uc2dc\uc624.<\/p>\n<p><strong>\uc911\uc694<\/strong>: \ud2b9\uc815 \uac00\uc6a9 \uc601\uc5ed\uc5d0\uc11c \uc11c\ube0c\ub137\uc744 \uc0dd\uc131\ud558\uace0 \uc120\ud0dd\ud55c \ub0b4\uc6a9\uc744 \uae30\uc5b5\ud574 \ub450\uc2ed\uc2dc\uc624. EBS \ubcfc\ub968\uc740 \ub3d9\uc77c\ud55c \uc11c\ube0c\ub137\uc758 \uc778\uc2a4\ud134\uc2a4\uc5d0\ub9cc \uc5f0\uacb0\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \uad00\ub828 \uc124\uba85\uc740 \uadf8\ub9bc 1\uc744 \ucc38\uc870\ud558\uc2ed\uc2dc\uc624. \uc774 \uc608\uc81c\uc5d0\uc11c\ub294 \uc124\uc815\uc744 \uc704\ud574 \uac00\uc6a9 \uc601\uc5ed\uc73c\ub85c <code class=\"lang-code\">us-west-2b<\/code>\ub97c \uc0ac\uc6a9\ud569\ub2c8\ub2e4. 3\ub2e8\uacc4\uc5d0\uc11c\ub294 EBS \uc2a4\ub0c5\uc0f7\uc744 \uc0ac\uc6a9\ud558\uc5ec \uc5ec\ub7ec \uac00\uc6a9 \uc601\uc5ed \uac04\uc5d0 EBS \ubcfc\ub968 \ub9c8\uc774\uadf8\ub808\uc774\uc158\uc744 \uc790\ub3d9\ud654\ud558\ub294 \ubc29\ubc95\uc744 \ubcf4\uc5ec\uc90d\ub2c8\ub2e4.<\/p>\n<p>\uc774 \uc608\uc81c\uc5d0\uc11c, \uae30\uc6b8\uc784\uaf34\ub85c \ud45c\uc2dc\ub41c \ubaa8\ub4e0 \ud56d\ubaa9\uc740 \uc0ac\uc6a9\uc790 \uc124\uc815\uc5d0 \ud2b9\uc815\ud55c \uac12\uc73c\ub85c \ub300\uccb4\ud574\uc57c \ud558\uba70, \ub098\uba38\uc9c0\ub294 \uadf8\ub300\ub85c \ubcf5\uc0ac\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n<p>\ub2e4\uc74c\uc73c\ub85c, \ub370\uc774\ud130 \uc138\ud2b8 \ubc0f \uccb4\ud06c\ud3ec\uc778\ud2b8\uc5d0 \ub300\ud55c EBS \ubcfc\ub968\uc744 \uc0dd\uc131\ud569\ub2c8\ub2e4. \uc5ec\uae30\uc5d0\uc11c\ub294 100GiB\ub97c \uc694\uccad\ud569\ub2c8\ub2e4. \uc5ec\ub7ec\ubd84\uc740 \ub370\uc774\ud130 \uc138\ud2b8 \uc694\uad6c\uc5d0 \uc801\ud569\ud55c \uac12\uc744 \uc120\ud0dd\ud574\uc57c \ud569\ub2c8\ub2e4. EBS \ubcfc\ub968\uc740 \uc778\uc2a4\ud134\uc2a4\uc640 \ub3d9\uc77c\ud55c \uac00\uc6a9 \uc601\uc5ed\uc5d0 \uc788\uc5b4\uc57c \ud569\ub2c8\ub2e4.&nbsp;\ubcfc\ub968\uc744 \uc0dd\uc131\ud55c \ud6c4\uc5d0\ub294 \uc778\uc2a4\ud134\uc2a4\uc5d0 \uc5f0\uacb0\ud569\ub2c8\ub2e4. <code class=\"lang-code\">run-instances<\/code> \ubc0f <code class=\"lang-code\">create-volume<\/code> \uba85\ub839 \ucd9c\ub825\uc5d0\uc11c ID \uc138\ubd80 \uc815\ubcf4\ub97c \uc9c0\uc815\ud569\ub2c8\ub2e4.<\/p>\n<div class=\"hide-language\">\n<pre><code class=\"lang-bash\">aws ec2 create-volume \n--size 100 \n--region &lt;AWS_REGION&gt; \n--availability-zone &lt;INSTANCE_AZ&gt; \n--volume-type gp2 \n--tag-specifications 'ResourceType=volume,Tags=[{Key=Name,Value=DL-datasets-checkpoints}]'\n\naws ec2 attach-volume \n--volume-id vol-&lt;your_volume_id&gt; \n--instance-id i-&lt;your_instance_id&gt; \n--device \/dev\/sdf\n<\/code><\/pre>\n<\/div>\n<p>\ubb38\uc11c\uc758 \ub2e8\uacc4\uc5d0 \ub530\ub77c <a href=\"https:\/\/docs.aws.amazon.com\/AWSEC2\/latest\/UserGuide\/AccessingInstancesLinux.html\" target=\"_blank\" rel=\"noopener noreferrer\">SSH\ub97c \uc0ac\uc6a9\ud558\uc5ec \uc778\uc2a4\ud134\uc2a4<\/a>\uc5d0 \uc5f0\uacb0\ud558\uace0 <a href=\"https:\/\/docs.aws.amazon.com\/AWSEC2\/latest\/UserGuide\/ebs-using-volumes.html\" target=\"_blank\" rel=\"noopener noreferrer\">\uc5f0\uacb0\ub41c \ubcfc\ub968\uc744 \ud3ec\ub9f7\ud558\uace0 \ub9c8\uc6b4\ud2b8<\/a>\ud569\ub2c8\ub2e4. \uc774 \uc608\uc81c\uc5d0\uc11c\ub294 \uc774\ub984\uc774 <code class=\"lang-code\">\/dltraining<\/code>\uc778 \ub8e8\ud2b8\uc5d0\uc11c \ub9c8\uc6b4\ud2b8 \uc9c0\uc810 \ub514\ub809\ud130\ub9ac\ub97c \uc0ac\uc6a9\ud569\ub2c8\ub2e4.<\/p>\n<p>\uc774 \ub2e8\uacc4\ub294 \ud55c \ubc88\ub9cc \uc218\ud589\ud569\ub2c8\ub2e4. 3\ub2e8\uacc4 \ud6c4\ubc18\uc73c\ub85c \uac00\uba74, \ud559\uc2b5\uc744 \uc704\ud574 \ub370\uc774\ud130 \uc138\ud2b8 \ubc0f \uccb4\ud06c\ud3ec\uc778\ud2b8\ub97c \uc0ac\uc6a9\ud560 \uc218 \uc788\ub3c4\ub85d, \uc0c8\ub85c\uc6b4 \uac01 \uc2a4\ud31f \uc778\uc2a4\ud134\uc2a4\uac00 \uc2dc\uc791 \uc2dc \ubcfc\ub968\uc744 \uc790\ub3d9\uc73c\ub85c \uc790\uccb4 \ub9c8\uc6b4\ud2b8\ud558\ub294 \ubc29\ubc95\uc744 \ud655\uc778\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n<p>\uc774 \uc608\uc81c\uc5d0\uc11c\ub294 \ub2e4\uc74c \uacbd\ub85c\ub97c \uc0ac\uc6a9\ud569\ub2c8\ub2e4.<\/p>\n<ul>\n<li>\ub370\uc774\ud130 \uc138\ud2b8: <code class=\"lang-code\">\/dltraining\/datasets<\/code><\/li>\n<li>\ud559\uc2b5 \uc9c4\ud589 \uc0c1\ud669 \uccb4\ud06c\ud3ec\uc778\ud2b8: <code class=\"lang-code\">\/dltraining\/checkpoints<\/code><\/li>\n<\/ul>\n<div class=\"hide-language\">\n<pre><code class=\"lang-bash\">sudo mkdir \/dltraining\nsudo mkfs -t xfs \/dev\/xvdf\nsudo mount \/dev\/xvdf \/dltraining\nsudo chown -R ubuntu: \/dltraining\/\ncd \/dltraining\nmkdir datasets\nmkdir checkpoints\n#\n# Optional: Run commands to move your custom datasets into the Datasets directory.\n#\n<\/code><\/pre>\n<\/div>\n<p>\uc774 \uc608\uc81c\ub97c \ub530\ub77c \uc9c4\ud589\ud558\uba74 \uc774 \ub514\ub809\ud130\ub9ac\ub97c \uc0dd\uc131\ud558\uace0 \ube48 \uc0c1\ud0dc\ub85c \ub458 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \ud559\uc2b5 \uc2a4\ud06c\ub9bd\ud2b8 <code class=\"lang-code\">ec2_spot_keras_training.py<\/code>\ub294 Keras\ub97c \uc0ac\uc6a9\ud558\uc5ec CIFAR10 \ub370\uc774\ud130 \uc138\ud2b8\ub97c \ub2e4\uc6b4\ub85c\ub4dc\ud558\uba70, \ucd5c\ucd08 \ud559\uc2b5\uc774 \uc2dc\uc791\ub429\ub2c8\ub2e4.<\/p>\n<p>\ub2e4\uc74c \uba85\ub839\uc744 \uc0ac\uc6a9\ud558\uc5ec \uc774 \uc778\uc2a4\ud134\uc2a4\ub97c \uc885\ub8cc\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \uc774\uc81c \ubcfc\ub968 \uc124\uc815\uc774 \uc644\ub8cc\ub418\uc5c8\uc73c\uba70, \uc774 \ubcfc\ub968\uc774 \uc0dd\uc131\ub41c \uac00\uc6a9 \uc601\uc5ed\uc5d0\uc11c \uc601\uad6c \ubcfc\ub968\uc774 \ub429\ub2c8\ub2e4.<\/p>\n<div class=\"hide-language\">\n<pre><code class=\"lang-bash\">aws ec2 terminate-instances \n--instance-ids i-&lt;your_instance_id&gt; \n--output text<\/code><\/pre>\n<\/div>\n<h3>2\ub2e8\uacc4: \uc778\uc2a4\ud134\uc2a4 \uad8c\ud55c\uc744 \ubd80\uc5ec\ud560 IAM \uc5ed\ud560 \ubc0f \uc815\ucc45 \uc0dd\uc131<\/h3>\n<p>\ud074\ub77c\uc6b0\ub4dc\ub97c \ucc98\uc74c \uc811\ud558\ub294 \uacbd\uc6b0 <a href=\"http:\/\/aws.amazon.com\/iam\" target=\"_blank\" rel=\"noopener noreferrer\">AWS Identity and Access Management<\/a>(IAM) \uac1c\ub150\ub3c4 \uc0c8\ub85c\uc6b8 \uc218 \uc788\uc2b5\ub2c8\ub2e4. <a href=\"https:\/\/docs.aws.amazon.com\/AWSEC2\/latest\/UserGuide\/iam-roles-for-amazon-ec2.html\" target=\"_blank\" rel=\"noopener noreferrer\">IAM \uc5ed\ud560 \ubc0f \uc815\ucc45<\/a>\uc740 \uc790\ub3d9\uc73c\ub85c \ub2e4\ub978 AWS \uc11c\ube44\uc2a4\uc5d0 \uc561\uc138\uc2a4\ud558\ub3c4\ub85d \ud5c8\uc6a9\ud558\ub294 \ud2b9\uc815 \uad8c\ud55c\uc744 \uc778\uc2a4\ud134\uc2a4\uc5d0 \ubd80\uc5ec\ud558\ub294 \ub370 \uc0ac\uc6a9\ub429\ub2c8\ub2e4.<\/p>\n<p>\uc800\ub294 \uc2a4\ud31f \uc778\uc2a4\ud134\uc2a4\ub97c \uc0ac\uc6a9\ud558\uc5ec \ud559\uc2b5 \uc911\uc5d0 1\ub2e8\uacc4\uc5d0\uc11c \uc0dd\uc131\ud55c EBS \ubcfc\ub968\uc758 \ub370\uc774\ud130 \uc138\ud2b8\uc640 \uccb4\ud06c\ud3ec\uc778\ud2b8\uc5d0 \uc561\uc138\uc2a4\ud558\ub824\uace0 \ud569\ub2c8\ub2e4. \ud558\uc9c0\ub9cc \uc778\uc2a4\ud134\uc2a4\uc640 \ub3d9\uc77c\ud55c \uac00\uc6a9 \uc601\uc5ed\uc758 \ubcfc\ub968\ub9cc \uc778\uc2a4\ud134\uc2a4\uc5d0 \uc5f0\uacb0\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \ubcfc\ub968\uacfc \uc778\uc2a4\ud134\uc2a4\uac00 \uc11c\ub85c \ub2e4\ub978 \uac00\uc6a9 \uc601\uc5ed\uc5d0 \uc788\uc73c\uba74 Amazon S3\uc5d0 \uc800\uc7a5\ub41c \ubcfc\ub968\uc758 \uc2a4\ub0c5\uc0f7\uc744 \uc0ac\uc6a9\ud558\uc5ec \uc0c8 \ubcfc\ub968\uc744 \uc0dd\uc131\ud574\uc57c \ud569\ub2c8\ub2e4.<\/p>\n<p>\uc774 \ubaa8\ub4e0 \ub2e8\uacc4\ub294 AWS CLI \ubc0f <a href=\"https:\/\/docs.aws.amazon.com\/AWSEC2\/latest\/UserGuide\/ec2-instance-metadata.html\" target=\"_blank\" rel=\"noopener noreferrer\">\uc0ac\uc6a9\uc790 \ub370\uc774\ud130 bash \uc2a4\ud06c\ub9bd\ud2b8<\/a>\ub97c \uc0ac\uc6a9\ud558\uc5ec \uc778\uc2a4\ud134\uc2a4\ub97c \uc2dc\uc791\ud560 \ub54c \uc218\ud589\ud560 \uc218 \uc788\uc73c\uba70, 3\ub2e8\uacc4\uc5d0\uc11c \uadf8 \ubc29\ubc95\uc744 \ud655\uc778\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \ub2e4\uc74c\uc740 \uc778\uc2a4\ud134\uc2a4\ub97c \uc2dc\uc791\ud560 \ub54c \uc2e4\ud589\ud574\uc57c \ud558\ub294 \ubaa8\ub4e0 AWS CLI \uba85\ub839\uc785\ub2c8\ub2e4.<\/p>\n<ul>\n<li>\uc774\ub984 \ud0dc\uadf8\uac00 <em>DL-datasets-checkpoints<\/em>\uc778 \ubcfc\ub968 \ucffc\ub9ac(\ud558\ub098\ub9cc \uc874\uc7ac\ud574\uc57c \ud568)<\/li>\n<li>\ud0dc\uadf8\uac00 <em>DL-datasets-checkpoints-snapshot<\/em>\uc778 \uc774 \ubcfc\ub968\uc758 \uc2a4\ub0c5\uc0f7 \uc0dd\uc131<\/li>\n<li>\uc778\uc2a4\ud134\uc2a4\uc640 \ubcfc\ub968\uc774 \ub3d9\uc77c\ud55c \uac00\uc6a9 \uc601\uc5ed\uc5d0 \uc788\uc73c\uba74 \uc778\uc2a4\ud134\uc2a4\uc5d0 \ubcfc\ub968 \uc5f0\uacb0<\/li>\n<li>\uc778\uc2a4\ud134\uc2a4\uc640 \ubcfc\ub968\uc774 \uc11c\ub85c \ub2e4\ub978 \uac00\uc6a9 \uc601\uc5ed\uc5d0 \uc788\uc73c\uba74, \uc774\ub984\uc774 <em>DL-datasets-checkpoints<\/em>\uc778 \uc778\uc2a4\ud134\uc2a4\uc758 \uac00\uc6a9 \uc601\uc5ed\uc5d0\uc11c \uc2a4\ub0c5\uc0f7\uc744 \ud1b5\ud574 \uc0c8 \ubcfc\ub968\uc744 \uc0dd\uc131\ud558\uace0 \uc778\uc2a4\ud134\uc2a4\uc5d0 \uc5f0\uacb0\ud569\ub2c8\ub2e4. \ud558\ub098\uc758 \uc0ac\ubcf8\ub9cc \uc874\uc7ac\ud558\ub3c4\ub85d \ub2e4\ub978 \uac00\uc6a9 \uc601\uc5ed\uc758 \ubcfc\ub968\uc740 \uc0ad\uc81c\ud569\ub2c8\ub2e4.<\/li>\n<li>\ud559\uc2b5\uc774 \uc644\ub8cc\ub418\uba74 \uc2a4\ud31f \ud50c\ub9bf \uc694\uccad\uc744 \ucde8\uc18c\ud558\uace0 \ubaa8\ub4e0 \ud559\uc2b5 \uc778\uc2a4\ud134\uc2a4\ub97c \uc885\ub8cc\ud569\ub2c8\ub2e4.<\/li>\n<\/ul>\n<p>\uc778\uc2a4\ud134\uc2a4\uc5d0\uc11c \uc774\ub7ec\ud55c \uc791\uc5c5\uc744 \uc218\ud589\ud560 \uc218 \uc788\ub3c4\ub85d \uc790\ub3d9\uc73c\ub85c \uc774\ub97c \uc218\ud589\ud560 \uad8c\ud55c\uc744 \uc778\uc2a4\ud134\uc2a4\uc5d0 \ubd80\uc5ec\ud574\uc57c \ud569\ub2c8\ub2e4.&nbsp;\uc774 \ubc29\ubc95\uc5d0\uc11c\ub294 \uc7a0\uc7ac\uc801\uc73c\ub85c \ub0a8\uc6a9\ub420 \uc18c\uc9c0\uac00 \uc788\uc73c\ubbc0\ub85c, \uc0ac\uc6a9\uc790\uc778 \uc81c\uac00 \uac16\uace0 \uc788\ub294 \ub3d9\uc77c\ud55c \ubaa8\ub4e0 \uad8c\ud55c\uc744 \uc778\uc2a4\ud134\uc2a4\uc5d0 \ubd80\uc5ec\ud558\uc9c0 \uc54a\uc2b5\ub2c8\ub2e4.<\/p>\n<p>\uba3c\uc800, IAM \uc5ed\ud560\uc774\ub77c\uace0 \ud558\ub294, Amazon EC2 \uc778\uc2a4\ud134\uc2a4\uc5d0 \ub300\ud55c \uc5ed\ud560\ubd80\ud130 \uc0dd\uc131\ud569\ub2c8\ub2e4. \uc774\ud6c4\uc5d0\ub294 \uc815\ucc45\uc774\ub77c\uace0 \ud558\ub294 \ud56d\ubaa9\uc744 \uc0dd\uc131\ud558\uc5ec \uc774 \uc5ed\ud560\uc5d0 \ud2b9\uc815 \uad8c\ud55c\uc744 \ubd80\uc5ec\ud569\ub2c8\ub2e4.&nbsp; \ub2e4\uc74c \uba85\ub839\uc744 \uc2e4\ud589\ud558\uc5ec \uc0c8 IAM \uc5ed\ud560\uc744 \uc0dd\uc131\ud569\ub2c8\ub2e4. \uc5ec\uae30\uc11c \uc5ed\ud560 \uc774\ub984\uc740 <code class=\"lang-code\">DL-Training<\/code>\uc774\ub77c\uace0 \uc9c0\uc815\ud588\uc9c0\ub9cc, \ub2e4\ub978 \uc774\ub984\uc744 \uc120\ud0dd\ud574\ub3c4 \uc88b\uc2b5\ub2c8\ub2e4.<\/p>\n<div class=\"hide-language\">\n<pre><code class=\"lang-bash\">aws iam create-role \n--role-name DL-Training \n--assume-role-policy-document '{&quot;Version&quot;:&quot;2012-10-17&quot;,&quot;Statement&quot;:[{&quot;Sid&quot;:&quot;&quot;,&quot;Effect&quot;:&quot;Allow&quot;,&quot;Principal&quot;:{&quot;Service&quot;:&quot;ec2.amazonaws.com&quot;},&quot;Action&quot;:&quot;sts:AssumeRole&quot;}]}'<\/code><\/pre>\n<\/div>\n<p>\ub2e4\uc74c\uc73c\ub85c, \uc778\uc2a4\ud134\uc2a4\uc5d0 \ub2e4\uc74c \uad8c\ud55c\uc744 \ubd80\uc5ec\ud558\ub294 \uc815\ucc45\uc744 \uc0dd\uc131\ud574 \uc5f0\uacb0\ud569\ub2c8\ub2e4.<\/p>\n<ol>\n<li>\ubcfc\ub968 \uc124\uba85, \uc0dd\uc131, \uc5f0\uacb0 \ubc0f \uc0ad\uc81c<\/li>\n<li>\ubcfc\ub968\uc5d0 \ub300\ud55c \uc2a4\ub0c5\uc0f7 \uc0dd\uc131<\/li>\n<li>\uc2a4\ud31f \uc778\uc2a4\ud134\uc2a4 \uc124\uba85<\/li>\n<li>\uc2a4\ud31f \ud50c\ub9bf \uc694\uccad \ucde8\uc18c \ubc0f \uc778\uc2a4\ud134\uc2a4 \uc885\ub8cc<\/li>\n<\/ol>\n<p>\uc0ac\uc6a9\uc790 \uc560\ud50c\ub9ac\ucf00\uc774\uc158\uc5d0\uc11c \uc0ac\uc6a9\ud558\ub824\ub294 \uacbd\uc6b0 \ub2e4\ub978 AWS \uc11c\ube44\uc2a4\uc5d0 \uc561\uc138\uc2a4\ud560 \uad8c\ud55c\uc744 \ubd80\uc5ec\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \uc77c\ubc18\uc801\uc73c\ub85c \uc791\uc5c5\uc774 \uad6c\uccb4\uc801\uc77c\uc218\ub85d \uc778\uc2a4\ud134\uc2a4 \uc791\uc5c5 \ud6a8\uc728\uc131\ub3c4 \ub192\uc544\uc9d1\ub2c8\ub2e4. \uad8c\ud55c\uc740 \uc608\uc81c GitHub \ub9ac\ud3ec\uc9c0\ud1a0\ub9ac\uc758 <a href=\"https:\/\/github.com\/awslabs\/ec2-spot-labs\/tree\/master\/ec2-spot-deep-learning-training\" target=\"_blank\" rel=\"noopener noreferrer\">ec2-permissions-dl-training.json<\/a> \ud30c\uc77c\uc5d0 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n<div class=\"hide-language\">\n<pre><code class=\"lang-json\">{\n&quot;Version&quot;: &quot;2012-10-17&quot;,\n&quot;Statement&quot;: [\n{\n&quot;Effect&quot;: &quot;Allow&quot;,\n&quot;Action&quot;: [\n&quot;ec2:AttachVolume&quot;,\n&quot;ec2:DeleteVolume&quot;,\n&quot;ec2:DescribeVolumeStatus&quot;,\n&quot;ec2:CancelSpotFleetRequests&quot;,\n&quot;ec2:CreateTags&quot;,\n&quot;ec2:DescribeVolumes&quot;,\n&quot;ec2:CreateSnapshot&quot;,\n&quot;ec2:DescribeSpotInstanceRequests&quot;,\n&quot;ec2:DescribeSnapshots&quot;,\n&quot;ec2:CreateVolume&quot;\n],\n&quot;Resource&quot;: &quot;*&quot;\n}\n]\n}\n<\/code><\/pre>\n<\/div>\n<p>\uadf8\ub9ac\uace0 \ub2e4\uc74c\uc744 \uc2e4\ud589\ud558\uc5ec \uc815\ucc45\uc744 \uc0dd\uc131\ud558\uace0 \uc774\ub97c IAM \uc5ed\ud560\uc5d0 \uc5f0\uacb0\ud569\ub2c8\ub2e4.<\/p>\n<div class=\"hide-language\">\n<pre><code class=\"lang-bash\">aws iam create-policy \n--policy-name ec2-permissions-dl-training \n--policy-document file:\/\/ec2-permissions-dl-training.json\n&nbsp;\naws iam attach-role-policy \n--policy-arn arn:aws:iam::&lt;account_id&gt;:policy\/ec2-permissions-dl-training \n--role-name DL-Training<\/code><\/pre>\n<\/div>\n<p>attach-role-policy \uba85\ub839\uc5d0\uc11c <em>&lt;account_id&gt;<\/em>\ub97c \uc0ac\uc6a9\uc790\uc758 AWS \uacc4\uc815 ID\ub85c \ub300\uccb4\ud574\uc57c \ud569\ub2c8\ub2e4.<\/p>\n<h3>3\ub2e8\uacc4: EC2 \uc0ac\uc6a9\uc790 \ub370\uc774\ud130 bash \uc2a4\ud06c\ub9bd\ud2b8 \uc0dd\uc131<\/h3>\n<p>\ub2e4\uc74c\uc73c\ub85c, \ud559\uc2b5\uc744 \uc2e4\ud589\ud558\ub824\ub294 \uc778\uc2a4\ud134\uc2a4\uc5d0 \ub300\ud55c \uc138\ubd80 \uc815\ubcf4\ub97c \uc0ac\uc6a9\ud558\uc5ec \uc2dc\uc791 \uc0ac\uc591 \ud30c\uc77c\uc744 \uc0dd\uc131\ud569\ub2c8\ub2e4. \uc774 \uc608\uc81c\uc5d0\uc11c\ub294 <code class=\"lang-code\">p3.2xlarge<\/code>\ub97c \uc0ac\uc6a9\ud558\ub824\uace0 \ud569\ub2c8\ub2e4. \ub2e4\uc911 GPU \uae30\ubc18 \ud559\uc2b5 \uc791\uc5c5\uc744 \uc2e4\ud589 \uc911\uc774\ub77c\uba74 GPU\uac00 \ub354 \ub9ce\uc740 \uc778\uc2a4\ud134\uc2a4\ub97c \uc694\uccad\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \uc5ec\uae30\uc11c, \ub2e4\uc911 GPU \uae30\ubc18 \uc791\uc5c5\uc740 \ub3d9\uc77c\ud55c \uc778\uc2a4\ud134\uc2a4\uc5d0 \uc788\ub294 \ub2e4\uc911 GPU\ub97c \ub9d0\ud569\ub2c8\ub2e4. \ud604\uc7ac \ub2e8\uc77c \uc778\uc2a4\ud134\uc2a4\uc5d0\uc11c \uac00\uc838\uc62c \uc218 \uc788\ub294 \ucd5c\ub300 GPU \uc218\ub294 <code class=\"lang-code\">p3.16xlarge<\/code> \ub610\ub294 <code class=\"lang-code\">p3dn.24xlarge<\/code>\uc5d0\uc11c 8\uac1c GPU\uc785\ub2c8\ub2e4. \ubd84\uc0b0\/\ub2e4\uc911 \ub178\ub4dc \ud559\uc2b5 \uc0ac\uc6a9 \uc0ac\ub840\ub294 \ud5a5\ud6c4 \ube14\ub85c\uadf8 \uac8c\uc2dc\ubb3c\uc5d0\uc11c \ub2e4\ub8f0 \uc608\uc815\uc785\ub2c8\ub2e4.<\/p>\n<p>2\ub2e8\uacc4\uc5d0\uc11c \uc124\uba85\ud55c \ub300\ub85c, Amazon EC2\uc5d0\uc11c\ub294 \uc2dc\uc791 \uc2dc \uc2e4\ud589\ud560 \uc778\uc2a4\ud134\uc2a4\ub85c \uc0ac\uc6a9\uc790 \ub370\uc774\ud130 \uc258 \uc2a4\ud06c\ub9bd\ud2b8\ub97c \uc804\ub2ec\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \uc774\uc81c \uc0ac\uc6a9\uc790 \ub370\uc774\ud130 \uc258 \uc2a4\ud06c\ub9bd\ud2b8\ub97c \uc0b4\ud3b4\ubcf4\uaca0\uc2b5\ub2c8\ub2e4. \uc804\uccb4 \uc2a4\ud06c\ub9bd\ud2b8(<a href=\"https:\/\/github.com\/awslabs\/ec2-spot-labs\/tree\/master\/ec2-spot-deep-learning-training\" target=\"_blank\" rel=\"noopener noreferrer\">user_data_script.sh<\/a>)\ub294 GitHub\uc5d0\uc11c \uc0ac\uc6a9\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n<p>\ud30c\uc77c\uc5d0\ub294 4\uac1c\uc758 \uc8fc\uc694 \uc139\uc158\uc774 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n<p><strong>&nbsp;<\/strong><strong>\uc778\uc2a4\ud134\uc2a4 ID \ubc0f \ucffc\ub9ac \ubcfc\ub968 \uac00\uc838\uc624\uae30<\/strong><\/p>\n<p>\uc774 \uc139\uc158\uc5d0\uc11c, \uc2a4\ud06c\ub9bd\ud2b8\ub294 \uc774 \uc2a4\ud06c\ub9bd\ud2b8\uac00 \uc2e4\ud589 \uc911\uc778 ID \uc778\uc2a4\ud134\uc2a4\uc5d0 \uc561\uc138\uc2a4\ud558\uae30 \uc704\ud574 \uc778\uc2a4\ud134\uc2a4 \uba54\ud0c0\ub370\uc774\ud130 API\ub97c \ucffc\ub9ac\ud569\ub2c8\ub2e4. \uadf8\ub9ac\uace0 \uc774 \uc815\ubcf4\ub97c \uc0ac\uc6a9\ud558\uc5ec \ud0dc\uadf8\uac00 <code class=\"lang-code\">DL-datasets-checkpoints<\/code>\uc778 \ub370\uc774\ud130 \uc138\ud2b8 \ubc0f \uccb4\ud06c\ud3ec\uc778\ud2b8 \ubcfc\ub968\uc744 \uac80\uc0c9\ud569\ub2c8\ub2e4.<\/p>\n<div class=\"hide-language\">\n<pre><code class=\"lang-bash\">#!\/bin\/bash\n\n# Get instance ID \nINSTANCE_ID=$(curl -s http:\/\/169.254.169.254\/latest\/meta-data\/instance-id)\nINSTANCE_AZ=$(curl -s http:\/\/169.254.169.254\/latest\/meta-data\/placement\/availability-zone)\nAWS_REGION=us-west-2\n\n# Get Volume Id and availability zone\nVOLUME_ID=$(aws ec2 describe-volumes --region $AWS_REGION --filter &quot;Name=tag:Name,Values=DL-datasets-checkpoints&quot; --query &quot;Volumes[].VolumeId&quot; --output text)\nVOLUME_AZ=$(aws ec2 describe-volumes --region $AWS_REGION --filter &quot;Name=tag:Name,Values=DL-datasets-checkpoints&quot; --query &quot;Volumes[].AvailabilityZone&quot; --output text)\n<\/code><\/pre>\n<\/div>\n<p><strong>\ubcfc\ub968 \ubc0f \uc778\uc2a4\ud134\uc2a4\uac00 \ub3d9\uc77c\ud55c \uac00\uc6a9 \uc601\uc5ed\uc5d0 \uc788\ub294\uc9c0 \ud655\uc778<\/strong><\/p>\n<p>\uc774 \uc139\uc158\uc5d0\uc11c\ub294, \uc2a4\ud06c\ub9bd\ud2b8\uac00 \ubcfc\ub968 \ubc0f \uc778\uc2a4\ud134\uc2a4\uac00 \ub3d9\uc77c\ud55c \uac00\uc6a9 \uc601\uc5ed\uc5d0 \uc788\ub294\uc9c0 \ud655\uc778\ud569\ub2c8\ub2e4. \ub2e4\ub978 \uac00\uc6a9 \uc601\uc5ed\uc5d0 \uc788\uc73c\uba74 \uba3c\uc800 Amazon S3\uc5d0\uc11c \ubcfc\ub968\uc758 \ud2b9\uc815 \uc2dc\uc810 \uc2a4\ub0c5\uc0f7\uc744 \uc0dd\uc131\ud569\ub2c8\ub2e4. \uc2a4\ub0c5\uc0f7\uc774 \uc0dd\uc131\ub418\uba74 \ubcfc\ub968\uc744 \uc0ad\uc81c\ud558\uace0 \uc778\uc2a4\ud134\uc2a4\uc758 \uac00\uc6a9 \uc601\uc5ed\uc5d0\uc11c \uc2a4\ub0c5\uc0f7\uc744 \ud1b5\ud574 \uc0c8 \ubcfc\ub968\uc744 \uc0dd\uc131\ud569\ub2c8\ub2e4. \uadf8\ub9bc 2\uc5d0\uc11c\ub294 \ub450 \uac1c\uc758 \ud328\ud134\uc744 \ubcf4\uc5ec\uc90d\ub2c8\ub2e4.<\/p>\n<p><code class=\"lang-code\">aws ec2 wait<\/code> \uba85\ub839\uc744 \uc0ac\uc6a9\ud558\uba74 \ub2e4\uc74c \uba85\ub839\uc73c\ub85c \uc9c4\ud589\ud558\uae30 \uc804\uc5d0 \uc2a4\ub0c5\uc0f7 \ubc0f \ubcfc\ub968 \uc0dd\uc131 \uc644\ub8cc\ub97c \ubcf4\uc7a5\ud569\ub2c8\ub2e4.<\/p>\n<div id=\"attachment_8040\" class=\"wp-caption alignnone\" style=\"width: 810px\">\n<p><img loading=\"lazy\" decoding=\"async\" class=\"size-full wp-image-8040\" src=\"https:\/\/d2908q01vomqb2.cloudfront.net\/f1f836cb4ea6efb2a0b1b99f41ad8b103eff4b59\/2019\/03\/26\/ec2-spot-instances-2.gif\" alt=\"\" width=\"800\" height=\"411\" \/><\/p>\n<p class=\"wp-caption-text\">\uadf8\ub9bc 2: \uc2a4\ud31f \uc778\uc2a4\ud134\uc2a4 \uc885\ub8cc \uc2dc \ub2e4\ub978 \uac00\uc6a9 \uc601\uc5ed\uc5d0\uc11c \uc0c8 \uc2a4\ud31f \uc778\uc2a4\ud134\uc2a4\uac00 \uc2dc\uc791\ub418\uba74 (a) EBS \ubcfc\ub968 \uc2a4\ub0c5\uc0f7\uc774 S3\uc5d0 \uc800\uc7a5\ub418\uace0 \uc778\uc2a4\ud134\uc2a4\uc758 \uac00\uc6a9 \uc601\uc5ed\uc5d0\uc11c \uc2a4\ub0c5\uc0f7\uc744 \ud1b5\ud574 \uc0c8 \ubcfc\ub968\uc774 \uc0dd\uc131\ub429\ub2c8\ub2e4. \ubcfc\ub968\uacfc \ub3d9\uc77c\ud55c \uac00\uc6a9 \uc601\uc5ed\uc5d0\uc11c \uc0c8 \uc2a4\ud31f \uc778\uc2a4\ud134\uc2a4\uac00 \uc2dc\uc791\ub418\uba74 (b) \ub3d9\uc77c\ud55c EBS \ubcfc\ub968\uc774 \uc0c8 \uc778\uc2a4\ud134\uc2a4\uc5d0 \uc5f0\uacb0\ub429\ub2c8\ub2e4.<\/p>\n<\/div>\n<div class=\"hide-language\">\n<pre><code class=\"lang-bash\">if [ $VOLUME_AZ != $INSTANCE_AZ ]; then\n&nbsp;&nbsp;&nbsp;&nbsp;SNAPSHOT_ID=$(aws ec2 create-snapshot \n&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;--region $AWS_REGION \n&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;--volume-id $VOLUME_ID \n&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;--description &quot;`date +&quot;%D %T&quot;`&quot; \n&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;--tag-specifications 'ResourceType=snapshot,Tags=[{Key=Name,Value=DL-datasets-checkpoints-snapshot}]' \n&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;--query SnapshotId --output text)\n&nbsp;&nbsp;&nbsp;&nbsp;aws ec2 wait --region $AWS_REGION snapshot-completed --snapshot-ids $SNAPSHOT_ID\n&nbsp;&nbsp;&nbsp;&nbsp;aws ec2 --region $AWS_REGION delete-volume --volume-id $VOLUME_ID\n&nbsp;&nbsp;&nbsp;&nbsp;VOLUME_ID=$(aws ec2 create-volume \n&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;--region $AWS_REGION \n&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;--availability-zone $INSTANCE_AZ \n&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;--snapshot-id $SNAPSHOT_ID \n&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;--volume-type gp2 \n&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;--tag-specifications 'ResourceType=volume,Tags=[{Key=Name,Value=DL-datasets-checkpoints}]' \n&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;--query VolumeId --output text)\n&nbsp;&nbsp;&nbsp;&nbsp;aws ec2 wait volume-available --region $AWS_REGION --volume-id $VOLUME_ID\nfi\n<\/code><\/pre>\n<\/div>\n<p><strong>\ubcfc\ub968 \uc5f0\uacb0 \ubc0f \ub9c8\uc6b4\ud2b8:<\/strong> \uc774 \uc139\uc158\uc5d0\uc11c\ub294, \uc2a4\ud06c\ub9bd\ud2b8\uac00 \uba3c\uc800 \uc778\uc2a4\ud134\uc2a4\uc640 \ub3d9\uc77c\ud55c \uac00\uc6a9 \uc601\uc5ed\uc5d0 \uc788\ub294 \ubcfc\ub968\uc744 \uc5f0\uacb0\ud569\ub2c8\ub2e4. \uadf8\ub7ec\uba74 \/dltraining\uc758 \ub9c8\uc6b4\ud2b8 \uc9c0\uc810 \ub514\ub809\ud130\ub9ac\uc5d0 \uc5f0\uacb0\ub41c \ubcfc\ub968\uc744 \ub9c8\uc6b4\ud2b8\ud569\ub2c8\ub2e4. \uadf8\ub9ac\uace0 \uc0ac\uc6a9\uc790 \ub370\uc774\ud130 \uc2a4\ud06c\ub9bd\ud2b8\uac00 \ub8e8\ud2b8\ub85c \uc2e4\ud589\ub418\ubbc0\ub85c Ubuntu \uc0ac\uc6a9\uc790\uc5d0 \ub300\ud55c \uc18c\uc720\uad8c\uc744 \uc5c5\ub370\uc774\ud2b8\ud569\ub2c8\ub2e4.<\/p>\n<div class=\"hide-language\">\n<pre><code class=\"lang-bash\">aws ec2 attach-volume \n--region $AWS_REGION --volume-id $VOLUME_ID \n--instance-id $INSTANCE_ID --device \/dev\/sdf\nsleep 10\n\n# Mount volume and change ownership, since this script is run as root\nmkdir \/dltraining\nmount \/dev\/xvdf \/dltraining\nchown -R ubuntu: \/dltraining\/\ncd \/home\/ubuntu\/\n<\/code><\/pre>\n<\/div>\n<p><strong>\ud559\uc2b5 \uc2a4\ud06c\ub9bd\ud2b8 \uac00\uc838\uc624\uae30:<\/strong>&nbsp;\uc774 \uc139\uc158\uc5d0\uc11c\ub294, \uc2a4\ud06c\ub9bd\ud2b8\uac00 \ud559\uc2b5 \ucf54\ub4dc git \ub9ac\ud3ec\uc9c0\ud1a0\ub9ac\ub97c \ubcf5\uc81c\ud569\ub2c8\ub2e4.<\/p>\n<div class=\"hide-language\">\n<pre><code class=\"lang-bash\"># Get training code\ngit clone https:\/\/github.com\/awslabs\/ec2-spot-labs.git\nchown -R ubuntu: ec2-spot-labs\ncd ec2-spot-labs\/ec2-spot-deep-learning-training\/\n<\/code><\/pre>\n<\/div>\n<p><strong>\ud559\uc2b5 \uc2dc\uc791\/\uc7ac\uac1c<\/strong>: \uc2a4\ud06c\ub9bd\ud2b8\uac00 <code class=\"lang-code\">tensorflow_p36<\/code> Conda \ud658\uacbd\uc744 \ud65c\uc131\ud654\ud558\uace0 Ubuntu \uc0ac\uc6a9\uc790\ub85c \ud559\uc2b5 \uc2a4\ud06c\ub9bd\ud2b8\ub97c \uc2e4\ud589\ud569\ub2c8\ub2e4. \ud559\uc2b5 \uc2a4\ud06c\ub9bd\ud2b8\ub294 Amazon EBS \ubcfc\ub968\uc758 \ub370\uc774\ud130 \uc138\ud2b8 \ub85c\ub4dc \uc791\uc5c5\uacfc \uccb4\ud06c\ud3ec\uc778\ud2b8\uc5d0\uc11c \ud559\uc2b5 \uc7ac\uac1c \uc791\uc5c5\uc744 \uad00\ub9ac\ud569\ub2c8\ub2e4. 4\ub2e8\uacc4\uc5d0\uc11c\ub294 \ud559\uc2b5 \uc2a4\ud06c\ub9bd\ud2b8\uc5d0 \ud544\uc694\ud55c \uc218\uc815\uc744 \uc2dc\uc791\ud569\ub2c8\ub2e4.<\/p>\n<div class=\"hide-language\">\n<pre><code class=\"lang-bash\"># Initiate training using the tensorflow_36 conda environment\nsudo -H -u ubuntu bash -c &quot;source \/home\/ubuntu\/anaconda3\/bin\/activate tensorflow_p36; python ec2_spot_keras_training.py &quot;\n<\/code><\/pre>\n<\/div>\n<p><strong>\uc815\ub9ac:<\/strong>&nbsp;\ud559\uc2b5\uc774 \uc644\ub8cc\ub418\uba74 \uc2a4\ud06c\ub9bd\ud2b8\ub294 \ud604\uc7ac \uc778\uc2a4\ud134\uc2a4\uc5d0 \uc5f0\uacb0\ub41c \uc2a4\ud31f \ud50c\ub9bf \uc694\uccad\uc744 \ucde8\uc18c\ud558\uc5ec \uc815\ub9ac\ud569\ub2c8\ub2e4. \ub610\ud55c <code class=\"lang-code\">cancel-spot-fleet-requests<\/code>\ub97c \ud1b5\ud574 \ud50c\ub9bf\uc774 \uad00\ub9ac\ud558\ub294 \uc778\uc2a4\ud134\uc2a4\ub97c \uc885\ub8cc\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n<div class=\"hide-language\">\n<pre><code class=\"lang-bash\"># After training, clean up by cancelling spot fleet requests\nSPOT_FLEET_REQUEST_ID=$(aws ec2 describe-spot-instance-requests --region $AWS_REGION --filter &quot;Name=instance-id,Values='$INSTANCE_ID'&quot; --query &quot;SpotInstanceRequests[].Tags[?Key=='aws:ec2spot:fleet-request-id'].Value[]&quot; --output text)\n\naws ec2 cancel-spot-fleet-requests --region $AWS_REGION --spot-fleet-request-ids $SPOT_FLEET_REQUEST_ID --terminate-instances\n<\/code><\/pre>\n<\/div>\n<h3>4\ub2e8\uacc4: \uc2a4\ud31f \ud50c\ub9bf \uc694\uccad \uad6c\uc131 \ud30c\uc77c \uc0dd\uc131<\/h3>\n<p>\ub2e4\uc74c\uc73c\ub85c, \ub300\uc0c1 \uc6a9\ub7c9(\uc608\uc81c\uc5d0\uc11c\ub294 \uc778\uc2a4\ud134\uc2a4 1\uac1c), \uc778\uc2a4\ud134\uc2a4\uc5d0 \ub300\ud55c \uc2dc\uc791 \uc0ac\uc591 \ubc0f \uc9c0\ubd88\ud560 \uc6a9\uc758\uac00 \uc788\ub294 \ucd5c\ub300 \uac00\uaca9\uc744 \ud3ec\ud568\ud558\ub294 \uc2a4\ud31f \ud50c\ub9bf \uad6c\uc131 \ud30c\uc77c\uc744 \uc0dd\uc131\ud569\ub2c8\ub2e4.&nbsp; \uc2a4\ud31f \ud50c\ub9bf\uc740 \ub300\uc0c1 \uc6a9\ub7c9\uc744 \ucda9\uc871\ud558\ub3c4\ub85d \uc694\uccad\uc744 \uc81c\ucd9c\ud558\uace0 \uc911\ub2e8\ub41c \uc778\uc2a4\ud134\uc2a4\ub97c \uc790\ub3d9\uc73c\ub85c \ubcf4\ucda9\ud569\ub2c8\ub2e4.<\/p>\n<p><code class=\"lang-code\">LaunchSpecifications<\/code> \uc139\uc158 \uc544\ub798\uc5d0 \ub450 \uac1c\uc758 \uc11c\ub85c \ub2e4\ub978 \uc0ac\uc591\uc744 \ud3ec\ud568\ud558\uc600\uc2b5\ub2c8\ub2e4.<\/p>\n<ol>\n<li>us-west-2 \ub9ac\uc804 \ub0b4 \ubaa8\ub4e0 \uac00\uc6a9 \uc601\uc5ed\uc5d0 \ubc30\uce58\ud560 \uc218 \uc788\ub294 p3.2xlarge \uc778\uc2a4\ud134\uc2a4 \uc720\ud615<\/li>\n<li>us-west-2 \ub9ac\uc804 \ub0b4 \ubaa8\ub4e0 \uac00\uc6a9 \uc601\uc5ed\uc5d0 \ubc30\uce58\ud560 \uc218 \uc788\ub294 p2.xlarge \uc778\uc2a4\ud134\uc2a4 \uc720\ud615<\/li>\n<\/ol>\n<p>\uc2a4\ud31f \ud50c\ub9bf \uad6c\uc131\uc740 \uc608\uc81c GitHub \ub9ac\ud3ec\uc9c0\ud1a0\ub9ac\uc758 <a href=\"https:\/\/github.com\/awslabs\/ec2-spot-labs\/tree\/master\/ec2-spot-deep-learning-training\" target=\"_blank\" rel=\"noopener noreferrer\">spot_fleet_config.json<\/a> \ud30c\uc77c\uc5d0 \uc788\uc2b5\ub2c8\ub2e4. \uc2a4\ud31f \ud50c\ub9bf \uad6c\uc131 \ud30c\uc77c\uc744 \uc0ac\uc6a9\ud558\uba74 \uc778\uc2a4\ud134\uc2a4 \uc720\ud615\uacfc \uac00\uc6a9 \uc601\uc5ed\uc744 \ud0c4\ub825\uc801\uc73c\ub85c \uc870\ud569\ud558\uc5ec \uc0ac\uc6a9\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \uc0ac\uc6a9\uc790\uc758 \ud559\uc2b5 \uc2a4\ud06c\ub9bd\ud2b8\uc5d0\uc11c NVIDIA Tesla V100\uc758 \ud63c\ud569 \uc815\ubc00\ub3c4 Tensor Core\ub97c \uc0ac\uc6a9\ud55c\ub2e4\uba74 \uc778\uc2a4\ud134\uc2a4 \uc720\ud615\uc744 <code class=\"lang-code\">p3.2xlarge<\/code>\ub85c\ub9cc \uc81c\ud55c\ud560 \uc218\ub3c4 \uc788\uc2b5\ub2c8\ub2e4. NVIDIA Tesla K80 \uae30\ubc18 <code class=\"lang-code\">p2.xlarge<\/code>\ub294 \ub2e8\uc77c \uc815\ubc00\ub3c4(FP32) \ubc0f \ubc30\uc815\ubc00\ub3c4(FP64)\ub9cc \uc9c0\uc6d0\ud558\uba70, \ub525\ub7ec\ub2dd \ud559\uc2b5\uc5d0 \ub300\ud574 V100\ubcf4\ub2e4 \ub290\ub9ac\uc9c0\ub9cc \ub354 \uc800\ub834\ud569\ub2c8\ub2e4. \uc694\uad6c \uc0ac\ud56d\uc5d0 \uc801\ud569\ud55c \uc870\ud569\uc744 \uc120\ud0dd\ud569\ub2c8\ub2e4.<\/p>\n<div class=\"hide-language\">\n<pre><code class=\"lang-bash\">{\n&quot;TargetCapacity&quot;: 1,\n&quot;AllocationStrategy&quot;: &quot;lowestPrice&quot;,\n&quot;IamFleetRole&quot;: &quot;arn:aws:iam::&lt;ACCOUNT_NUMBER&gt;:role\/DL-Training-Spot-Fleet-Role&quot;,\n&quot;LaunchSpecifications&quot;: [\n{\n&quot;ImageId&quot;: &quot;ami-0027dfad6168539c7&quot;,\n&quot;KeyName&quot;: &quot;&lt;KEYPAIR_NAME&gt;&quot;,\n&quot;SecurityGroups&quot;: [\n{\n&quot;GroupId&quot;: &lt;SECURITY_GROUP_ID&gt;\n}\n],\n&quot;InstanceType&quot;: &quot;p3.2xlarge&quot;,\n&quot;Placement&quot;: {\n&quot;AvailabilityZone&quot;: &quot;us-west-2a, us-west-2b, us-west-2c, us-west-2d&quot;\n},\n&quot;UserData&quot;: &quot;<strong>base64_encoded_bash_script<\/strong>&quot;,\n&quot;IamInstanceProfile&quot;: {\n&quot;Arn&quot;: &quot;arn:aws:iam::&lt;ACCOUNT_NUMBER&gt;:instance-profile\/DL-Training&quot;\n}\n},\n{\n&quot;ImageId&quot;: &quot;ami-0027dfad6168539c7&quot;,\n&quot;KeyName&quot;: &quot;&lt;KEYPAIR_NAME&gt;&quot;,\n&quot;SecurityGroups&quot;: [\n{\n&quot;GroupId&quot;: &lt;SECURITY_GROUP_ID&gt;\n}\n],\n&quot;InstanceType&quot;: &quot;p2.xlarge&quot;,\n&quot;Placement&quot;: {\n&quot;AvailabilityZone&quot;: &quot;us-west-2a, us-west-2b, us-west-2c, us-west-2d&quot;\n},\n&quot;UserData&quot;: &quot;<strong>base64_encoded_bash_script<\/strong>&quot;,\n&quot;IamInstanceProfile&quot;: {\n&quot;Arn&quot;: &quot;arn:aws:iam::&lt;ACCOUNT_NUMBER&gt;:instance-profile\/DL-Training&quot;\n}\n}\n<\/code><\/pre>\n<\/div>\n<p>\uc218\ub3d9\uc73c\ub85c \uc9c4\ud589 \uc0c1\ud669\uc744 \ud655\uc778\ud558\uace0 \ub514\ubc84\uae45\ud558\ub294 \uacbd\uc6b0\uc5d0 SSH\ub97c \ud1b5\ud574 \uc778\uc2a4\ud134\uc2a4\uc5d0 \uc5f0\uacb0\ud560 \uc218 \uc788\ub294 \ubcf4\uc548 \uadf8\ub8f9\uc744 \uc0ac\uc6a9\ud558\uace0, \uc778\uc99d\uc744 \uc704\ud574 \ud0a4 \ud398\uc5b4 \uc774\ub984\uc744 \uc0ac\uc6a9\ud574\uc57c \ud569\ub2c8\ub2e4. IAM \uc778\uc2a4\ud134\uc2a4 \ud504\ub85c\ud30c\uc77c \uc544\ub798\uc5d0\uc11c, 2\ub2e8\uacc4\uc5d0\uc11c \uc0dd\uc131\ud55c IAM \uc5ed\ud560\uc744 \uc5c5\ub370\uc774\ud2b8\ud558\uc5ec \ud544\uc694\ud55c \uad8c\ud55c\uc744 \uc778\uc2a4\ud134\uc2a4\uc5d0 \ubd80\uc5ec\ud569\ub2c8\ub2e4.<\/p>\n<p>\uc2a4\ud31f \ud50c\ub9bf \uc694\uccad\uc744 \uc0ac\uc6a9\ud558\ub824\uba74 \ub2e4\uc74c \uba85\ub839\uc744 \uc2e4\ud589\ud558\uc5ec IAM \ud50c\ub9bf \uc5ed\ud560\uc744 \uc0dd\uc131\ud569\ub2c8\ub2e4.<\/p>\n<div class=\"hide-language\">\n<pre><code class=\"lang-bash\">aws iam create-role \n--role-name DL-Training-Spot-Fleet-Role \n--assume-role-policy-document '{&quot;Version&quot;:&quot;2012-10-17&quot;,&quot;Statement&quot;:[{&quot;Sid&quot;:&quot;&quot;,&quot;Effect&quot;:&quot;Allow&quot;,&quot;Principal&quot;:{&quot;Service&quot;:&quot;spotfleet.amazonaws.com&quot;},&quot;Action&quot;:&quot;sts:AssumeRole&quot;}]}'\n\naws iam attach-role-policy \n--policy-arn arn:aws:iam::aws:policy\/service-role\/AmazonEC2SpotFleetTaggingRole --role-name DL-Training-Spot-Fleet-Role\n<\/code><\/pre>\n<\/div>\n<p>\uc704 \uad6c\uc131\uc758 \uc9e7\uc740 \ucf54\ub4dc \uc870\uac01\uc5d0\uc11c, \uc0ac\uc6a9\uc790 \ub370\uc774\ud130 \uc544\ub798\uc5d0 \uc788\ub294 <strong><em>base64_encoded_bash_script<\/em><\/strong> \ud14d\uc2a4\ud2b8\ub294 <a href=\"https:\/\/docs.aws.amazon.com\/AWSEC2\/latest\/UserGuide\/ec2-instance-metadata.html?shortFooter=true#instancedata-add-user-data\" target=\"_blank\" rel=\"noopener noreferrer\">base64\ub85c \uc778\ucf54\ub529\ub41c \uc0ac\uc6a9\uc790 \ub370\uc774\ud130<\/a>&nbsp;\uc258 \uc2a4\ud06c\ub9bd\ud2b8\ub85c \ub300\uccb4\ud574\uc57c \ud569\ub2c8\ub2e4. \uc774\ub97c \uc218\ud589\ud558\uae30 \uc704\ud574 Mac \ubc0f Linux \uae30\ubc18 OS\uc5d0\uc11c \uc0ac\uc6a9 \uac00\ub2a5\ud55c <code class=\"lang-code\">base64<\/code> \uc720\ud2f8\ub9ac\ud2f0\ub97c \uc0ac\uc6a9\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \ub2e4\uc74c\uc740 Mac\uc5d0\uc11c \uc791\ub3d9\ud569\ub2c8\ub2e4. Linux\uc5d0\uc11c \uc791\ub3d9\ud558\ub824\uba74 <code class=\"lang-code\">-b<\/code>\ub97c <code class=\"lang-code\">-w<\/code>\ub85c \ub300\uccb4\ud558\uc5ec \uc904 \ubc14\uafc8\uc744 \uc81c\uac70\ud569\ub2c8\ub2e4. <code class=\"lang-code\">sed<\/code> \uba85\ub839\uc740 <code class=\"lang-code\">base64_encoded_bash_script<\/code> \ubb38\uc790\uc5f4\uc758 \ubaa8\ub4e0 \ud56d\ubaa9\uc744 base64\ub85c \uc778\ucf54\ub529\ub41c bash \uc2a4\ud06c\ub9bd\ud2b8\ub85c \ub300\uccb4\ud569\ub2c8\ub2e4.<\/p>\n<div class=\"hide-language\">\n<pre><code class=\"lang-bash\">USER_DATA=`base64 user_data_script.sh -b0`\nsed -i '' &quot;s|base64_encoded_bash_script|$USER_DATA|g&quot; spot_fleet_config.json \n<\/code><\/pre>\n<\/div>\n<h3>5\ub2e8\uacc4: \ub525\ub7ec\ub2dd \ud559\uc2b5 \uc2a4\ud06c\ub9bd\ud2b8 \uc5c5\ub370\uc774\ud2b8<\/h3>\n<p>\ub9c8\uc9c0\ub9c9 \ub2e8\uacc4\uc5d0\uc11c\ub294, \ub370\uc774\ud130 \uc138\ud2b8\ub97c \uc5f0\uacb0\ub41c Amazon EBS \ubcfc\ub968\uc5d0\uc11c \ub85c\ub4dc\ud558\uace0 \uc774 \ubcfc\ub968\uc5d0 \uccb4\ud06c\ud3ec\uc778\ud2b8\ub97c \uc800\uc7a5\ud558\ub3c4\ub85d \ub525\ub7ec\ub2dd \ud559\uc2b5 \uc2a4\ud06c\ub9bd\ud2b8\ub97c \uc5c5\ub370\uc774\ud2b8\ud569\ub2c8\ub2e4. \uc774 \uc608\uc81c\uc5d0\uc11c\ub294 <a href=\"https:\/\/www.cs.toronto.edu\/~kriz\/cifar.html\" target=\"_blank\" rel=\"noopener noreferrer\">CIFAR10 \ub370\uc774\ud130 \uc138\ud2b8<\/a>\uc5d0\uc11c <a href=\"https:\/\/arxiv.org\/abs\/1512.03385\" target=\"_blank\" rel=\"noopener noreferrer\">ResNet50 \ubaa8\ub378<\/a>\uc744 \ud559\uc2b5\ud569\ub2c8\ub2e4.&nbsp;\uc77c\ubc18\uc801\uc778 \ub525\ub7ec\ub2dd \ud559\uc2b5 \uc2a4\ud06c\ub9bd\ud2b8\ub294 \ub2e4\uc74c \ub2e8\uacc4\ub97c \uc218\ud589\ud560 \uc218\ub3c4 \uc788\uc2b5\ub2c8\ub2e4. \uc544\ub798\uc758 \uc720\uc0ac \ucf54\ub4dc\uc5d0\ub294, \uc6b0\ub9ac\uac00 \ub2e4\ub8ec \uc124\uc815\uacfc \ud568\uaed8 \uc0ac\uc6a9\ud558\ub3c4\ub85d \ud559\uc2b5 \uc2a4\ud06c\ub9bd\ud2b8\uc5d0\uc11c \uc218\uc815\ud574\uc57c \ud558\ub294 \ubcc0\uacbd \uc0ac\ud56d\uc774 \ub098\uc640 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n<div class=\"hide-language\">\n<pre><code class=\"lang-bash\"># Prepare datasets \/ setup dataset loaders\ndataset = load_data(ebs_mount_point_dataset)\n\n# Define model\nif exists(ebs_mount_point_checkpoints)\ncheckpoint, checkpoint_epoch = get_latest_checkpoint(ebs_mount_point_checkpoints)\nmodel = load_model(checkpoint)\nelse\nmodel = define_model()\ncheckpoint_epoch = 0\n&nbsp;\n# Define training parameters\n\n# Execute training loop\nfor i = checkpoint_epoch to max_epoch\n...\n...\n...\n# Avoid corrupted checkpoints due to termination\nstatus = get_spot_termination_status()\nif status == \u201cTerminating\u201d\npause_training()\n# Save checkpoints and progress\nsave_model_checkpoint(model, ebs_mount_point_checkpoints)\nsave_progress_logs(ebs_mount_point)\nend\n<\/code><\/pre>\n<\/div>\n<p>\uc694\uc57d\ud558\uba74,<\/p>\n<ul>\n<li>\ub9c8\uc6b4\ud2b8\ub41c Amazon EBS \ubcfc\ub968(\uc608\uc81c\uc5d0\uc11c\ub294, <code class=\"lang-code\">\/dltraining<\/code>)\uc5d0\uc11c \ub370\uc774\ud130\ub97c \ub85c\ub4dc\ud569\ub2c8\ub2e4.<\/li>\n<li>\uccb4\ud06c\ud3ec\uc778\ud2b8\uac00 \uc788\ub294\uc9c0 \ud655\uc778\ud558\uace0, \uccb4\ud06c\ud3ec\uc778\ud2b8\ub97c \ub85c\ub4dc\ud558\uace0 epoch \ubc88\ud638\ub97c \uc5c5\ub370\uc774\ud2b8\ud558\uc5ec \ud559\uc2b5\uc744 \uc7ac\uac1c\ud569\ub2c8\ub2e4. \uadf8\ub807\uc9c0 \uc54a\uc73c\uba74, \ubaa8\ub378 \uc544\ud0a4\ud14d\ucc98\ub97c \uc815\uc758\ud558\uace0 \ucc98\uc74c\ubd80\ud130 \ud559\uc2b5\uc744 \uc0c8\ub85c \uc2dc\uc791\ud569\ub2c8\ub2e4.<\/li>\n<li>\ud559\uc2b5 \ub8e8\ud504\uc5d0\uc11c \uc885\ub8cc \uacf5\uc9c0\uac00 \ubc1c\ud589\ub418\uc5c8\ub294\uc9c0 \ud655\uc778\ud569\ub2c8\ub2e4. \uadf8\ub807\ub2e4\uba74, \uc190\uc0c1\ub418\uac70\ub098 \ubd88\uc644\uc804\ud55c \uccb4\ud06c\ud3ec\uc778\ud2b8\ub97c \ubc29\uc9c0\ud558\uae30 \uc704\ud574 \uccb4\ud06c\ud3ec\uc778\ud2b8 \uc124\uc815 \uc911\uc5d0 \uc885\ub8cc\ub418\uc9c0 \uc54a\ub3c4\ub85d \ud559\uc2b5\uc744 \uc77c\uc2dc \uc911\uc9c0\ud569\ub2c8\ub2e4.<\/li>\n<li>\uc885\ub8cc \uacf5\uc9c0\uac00 \ubc1c\ud589\ub418\uc9c0 \uc54a\uc740 \uacbd\uc6b0 <code class=\"lang-code\">\/dltraining\/checkpoints\/<\/code>\uc5d0 \ubaa8\ub378 \uccb4\ud06c\ud3ec\uc778\ud2b8\ub97c \uc800\uc7a5\ud569\ub2c8\ub2e4.<\/li>\n<\/ul>\n<p>\uc774 \uc608\uc81c\uc758 \ud559\uc2b5 \uc2a4\ud06c\ub9bd\ud2b8\ub294 <code class=\"lang-code\">ec2_spot_keras_training.py<\/code>\uc774\uba70, \uc608\uc81c \ub9ac\ud3ec\uc9c0\ud1a0\ub9ac\uc5d0\uc11c \uc0ac\uc6a9 \uac00\ub2a5\ud569\ub2c8\ub2e4. \ub2e4\uc74c\uc740 \uc6b0\ub9ac\uac00 \ub2e4\ub8ec \ud559\uc2b5 \uc2a4\ud06c\ub9bd\ud2b8\uc758 \uc9e7\uc740 \ucf54\ub4dc \uc870\uac01\uc785\ub2c8\ub2e4. <code class=\"lang-code\">load_checkpoint_model()<\/code> \ud568\uc218\ub294 \ud559\uc2b5\uc744 \uc7ac\uac1c\ud558\ub3c4\ub85d \ucd5c\uc2e0 \uccb4\ud06c\ud3ec\uc778\ud2b8\ub97c \ub85c\ub4dc\ud569\ub2c8\ub2e4.<\/p>\n<div class=\"hide-language\">\n<pre><code class=\"lang-bash\">def load_checkpoint_model(checkpoint_path, checkpoint_names):\nlist_of_checkpoint_files = glob.glob(os.path.join(checkpoint_path, '*'))\ncheckpoint_epoch_number = max([int(file.split(&quot;.&quot;)[1]) for file in list_of_checkpoint_files])\ncheckpoint_epoch_path = os.path.join(checkpoint_path,\ncheckpoint_names.format(epoch=checkpoint_epoch_number))\nresume_model = load_model(checkpoint_epoch_path)\nreturn resume_model, checkpoint_epoch_number\n<\/code><\/pre>\n<\/div>\n<p>\uc800\ub294 TensorFlow \ubc31\uc5d4\ub4dc\uc5d0\uc11c Keras\ub97c \uc0ac\uc6a9\ud588\uae30 \ub54c\ubb38\uc5d0 \ud559\uc2b5 \ub8e8\ud504\ub97c \uba85\uc2dc\uc801\uc73c\ub85c \uc791\uc131\ud558\uc9c0 \uc54a\uc544\ub3c4 \ub418\uc5c8\uc2b5\ub2c8\ub2e4. Keras\uc5d0\uc11c\ub294 \uac01 epoch \uc774\ud6c4\uc5d0 \uccb4\ud06c\ud3ec\uc778\ud2b8\ub97c \uc800\uc7a5\ud558\uace0 \uc9c4\ud589 \uc0c1\ud669\uc744 \uae30\ub85d\ud558\ub294 \uac04\ud3b8\ud55c \ucf5c\ubc31 \ud568\uc218\ub97c \uc81c\uacf5\ud569\ub2c8\ub2e4.<\/p>\n<p>\ucc38\uace0: TensorFlow\uc758 \ub85c\uc6b0 \ub808\ubca8 API, PyTorch \ub610\ub294 \uae30\ud0c0 \ud504\ub808\uc784\uc6cc\ud06c\ub97c \uc0ac\uc6a9\ud558\uc5ec \uc0ac\uc6a9\uc790\ub9cc\uc758 \ud559\uc2b5 \ub8e8\ud504\ub97c \uad6c\ud604\ud558\ub294 \uacbd\uc6b0 \uc0ac\uc6a9\uc790\uac00 \uc9c4\ud589 \uc0c1\ud669 \uccb4\ud06c\ud3ec\uc778\ud2b8\ub97c \uad00\ub9ac\ud574\uc57c \ud569\ub2c8\ub2e4. \ubb34\uc5c7\uc744 \ud560\uc9c0 \uc798 \ubaa8\ub974\ub294 \uacbd\uc6b0 \ub9e4\uc6b0 \uae4c\ub2e4\ub85c\uc6b8 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \ud559\uc2b5\uc744 \uc801\uc808\ud788 \uc7ac\uac1c\ud558\ub824\uba74 (1)&nbsp;\ubaa8\ub378\uc744 \uc7ac\uc815\uc758\ud558\uae30 \uc704\ud574 \ubaa8\ub378 \uc544\ud0a4\ud14d\ucc98, (2) \ud604\uc7ac epoch \uc885\ub8cc \uc2dc \ubaa8\ub378\uc758 \uc644\ub8cc\ub41c epoch \ubc88\ud638\uc640 \uac00\uc911\uce58, (3) \uc190\uc2e4 \ud568\uc218, \ucd5c\uc801\ud654 \ub3c4\uad6c, \ud559\uc2b5 \ube44\uc728 \uc77c\uc815 \ub4f1\uc758 \ud559\uc2b5 \ud558\uc774\ud37c \ud30c\ub77c\ubbf8\ud130 (4) epoch \uc885\ub8cc \uc2dc \ucd5c\uc801\ud654 \ub3c4\uad6c \uc0c1\ud0dc\ub97c \uc800\uc7a5\ud574\uc57c \ud569\ub2c8\ub2e4.<\/p>\n<p>\uc9c4\ud589 \uc0c1\ud669 \uccb4\ud06c\ud3ec\uc778\ud2b8\ub97c \uc124\uc815\ud558\uace0 \uc885\ub8cc \uc0c1\ud0dc\ub97c \ud655\uc778\ud558\uae30 \uc704\ud574 \uc0ac\uc6a9\ud558\ub294 Keras \ucf5c\ubc31\uc740 \ub2e4\uc74c\uacfc \uac19\uc2b5\ub2c8\ub2e4.<\/p>\n<div class=\"hide-language\">\n<pre><code class=\"lang-bash\">def define_callbacks(volume_mount_dir, checkpoint_path, checkpoint_names, today_date):\n\n# Model checkpoint callback\nif not os.path.isdir(checkpoint_path):\nos.makedirs(checkpoint_path)\nfilepath = os.path.join(checkpoint_path, checkpoint_names)\ncheckpoint_callback = ModelCheckpoint(filepath=filepath,\nsave_weights_only=False,\nmonitor='val_loss')\n\n# Loss history callback\nepoch_results_callback = CSVLogger(os.path.join(volume_mount_dir, \n'training_log_{}.csv'.format(today_date)),\nappend=True)\n\nclass SpotTermination(keras.callbacks.Callback):\ndef on_batch_begin(self, batch, logs={}):\nstatus_code = requests.get(&quot;http:\/\/169.254.169.254\/latest\/meta-data\/spot\/instance-action&quot;).status_code\nif status_code != 404:\ntime.sleep(150)\nspot_termination_callback = SpotTermination()\ncallbacks = [checkpoint_callback, epoch_results_callback]\nreturn callbacks\n<\/code><\/pre>\n<\/div>\n<h3>6\ub2e8\uacc4: \ud559\uc2b5\uc744 \uc2dc\uc791\ud558\ub3c4\ub85d \uc2a4\ud31f \uc694\uccad \uc2dc\uc791<\/h3>\n<p>\uc774\uc81c 4\ub2e8\uacc4\uc5d0\uc11c \uc0dd\uc131\ud55c <code class=\"lang-code\">spot_fleet_config.json<\/code> \uad6c\uc131 \ud30c\uc77c\uc744 \uc0ac\uc6a9\ud558\uc5ec \uc2a4\ud31f \ud50c\ub9bf \uc694\uccad\uc744 \uc81c\ucd9c\ud560 \uc900\ube44\uac00 \ub418\uc5c8\uc2b5\ub2c8\ub2e4.<\/p>\n<div class=\"hide-language\">\n<pre><code class=\"lang-bash\">aws ec2 request-spot-fleet --spot-fleet-request-config file:\/\/spot_fleet_config.json<\/code><\/pre>\n<\/div>\n<h3>\uc885\ud569<\/h3>\n<p>\uc9c0\uae08\uae4c\uc9c0 \ub9ce\uc740 \ucf54\ub4dc, \uad6c\uc131 \ud30c\uc77c, AWS CLI \uba85\ub839\uc744 \uc18c\uac1c\ud588\uc2b5\ub2c8\ub2e4. \uadf8\ub9bc 3\uc5d0\uc11c\ub294 \uc774\ub7ec\ud55c \ucf54\ub4dc\uc640 \uad6c\uc131 \uc544\ud2f0\ud329\ud2b8\ub97c \uc870\ud569\ud558\ub294 \ubc29\ubc95\uc744 \ubcf4\uc5ec\uc90d\ub2c8\ub2e4. \uc774\ub7ec\ud55c \uc694\uc18c\uac00 \ubaa8\ub450 \uc5b4\ub5bb\uac8c \uc5f0\uacb0\ub418\uc5c8\ub294\uc9c0 \ub354 \uc798 \ud30c\uc545\ud560 \uc218 \uc788\ub3c4\ub85d \uadf8 \uacfc\uc815\uc744 \uc0b4\ud3b4\ubcf4\uaca0\uc2b5\ub2c8\ub2e4.<\/p>\n<div id=\"attachment_8041\" class=\"wp-caption alignnone\" style=\"width: 810px\">\n<p><img loading=\"lazy\" decoding=\"async\" class=\"size-full wp-image-8041\" src=\"https:\/\/d2908q01vomqb2.cloudfront.net\/f1f836cb4ea6efb2a0b1b99f41ad8b103eff4b59\/2019\/03\/26\/ec2-spot-instances-3.gif\" alt=\"\" width=\"800\" height=\"471\" \/><\/p>\n<p class=\"wp-caption-text\">\uadf8\ub9bc 3: \ub370\uc774\ud130, \ucf54\ub4dc, \uad6c\uc131 \uc544\ud2f0\ud329\ud2b8 \uc885\uc18d\uc131 \ucc28\ud2b8<\/p>\n<\/div>\n<p>\uc5ec\ub7ec\ubd84\ub3c4 \uc2dc\uc791\ud574 \ubcf4\uc138\uc694!<\/p>\n<p>\ub525\ub7ec\ub2dd \uc5f0\uad6c\uc790\ub098 \uac1c\ubc1c\uc790\ub294 \uba3c\uc800 \ub85c\uceec\ub85c, \uc544\ub2c8\uba74 \uc800\ub834\ud55c CPU \uc804\uc6a9 Amazon EC2 \uc628\ub514\ub9e8\ub4dc \uc778\uc2a4\ud134\uc2a4\uc5d0\uc11c AWS Deep Learning AMI\ub97c \uc0ac\uc6a9\ud558\uc5ec \ubaa8\ub378 \ud504\ub85c\ud1a0\ud0c0\uc785\uc744 \ub9cc\ub4e4\uace0 \uac1c\ubc1c\ud569\ub2c8\ub2e4. GPU\uc5d0\uc11c \ud559\uc2b5\uc744 \uc2e4\ud589\ud560 \uc900\ube44\uac00 \ub418\uc5c8\ub2e4\uba74 \ud559\uc2b5 \uc2a4\ud06c\ub9bd\ud2b8\ub97c Git \ub9ac\ud3ec\uc9c0\ud1a0\ub9ac\ub85c \uc62e\uaca8\ubcf4\uc2ed\uc2dc\uc624.<\/p>\n<p>\uadf8\ub9ac\uace0 6\ub2e8\uacc4\uc5d0\uc11c \ubcf4\uc5ec\uc900 <code class=\"lang-code\">aws ec2 request-spot-fleet<\/code> \uba85\ub839\uc744 \uc0ac\uc6a9\ud558\uc5ec \uc2a4\ud31f \uc694\uccad\uc744 \uc81c\ucd9c\ud569\ub2c8\ub2e4. \uadf8\ub7ec\uba74 \uc774\uc81c \ubaa8\ub450 \uc791\ub3d9\ub429\ub2c8\ub2e4.<\/p>\n<p>\uc2a4\ud31f \uc694\uccad\uc740 \uc2a4\ud31f \ud50c\ub9bf \uad6c\uc131 \ud30c\uc77c <code class=\"lang-code\">spot_fleet_config.json<\/code>\uc744 \uc0ac\uc6a9\ud558\uc5ec \uc6d0\ud558\ub294 \uc2a4\ud31f \uc778\uc2a4\ud134\uc2a4 \uc720\ud615\uc744 \uc2dc\uc791\ud569\ub2c8\ub2e4. \uc774 \uc608\uc81c\uc5d0\uc11c\ub294 us-west-2 \ub9ac\uc804\uc758 \uac00\uc6a9 \uc601\uc5ed\uc5d0 \uc788\ub294 <code class=\"lang-code\">p3.2xlarge<\/code> \uc778\uc2a4\ud134\uc2a4\uc5d0\uc11c \ud559\uc2b5 \uc791\uc5c5\uc744 \uc2e4\ud589\ud569\ub2c8\ub2e4. \ud559\uc2b5 \uc2a4\ud06c\ub9bd\ud2b8\ub294 GPU\uc5d0 \ucd5c\uc801\ud654\ub41c TensorFlow \ud504\ub808\uc784\uc6cc\ud06c\ub97c \ud3ec\ud568\ud558\ub294, AWS Deep Learning AMI\ub97c \uc0ac\uc6a9\ud558\uc5ec \uc774\ubbf8\uc9c0\ub85c \uc0dd\uc131\ub41c \uc778\uc2a4\ud134\uc2a4\uc5d0\uc11c \uc2e4\ud589\ub429\ub2c8\ub2e4.<\/p>\n<p>\uc2a4\ud31f \ud50c\ub9ad \uad6c\uc131 \ud30c\uc77c\uc5d0\ub294 <code class=\"lang-code\">user_data_script.sh<\/code> bash \uc2a4\ud06c\ub9bd\ud2b8 \ud30c\uc77c\ub3c4 \uc788\uc2b5\ub2c8\ub2e4. \uc0ac\uc6a9\uc790 \ub370\uc774\ud130 bash \uc2a4\ud06c\ub9bd\ud2b8\ub294 \uc2dc\uc791 \uc2dc \uc2a4\ud31f \uc778\uc2a4\ud134\uc2a4\uc5d0\uc11c \uc2e4\ud589\ub429\ub2c8\ub2e4. \uc774 \uc2a4\ud06c\ub9bd\ud2b8\ub294 \ub370\uc774\ud130 \uc138\ud2b8 \ubc0f \uccb4\ud06c\ud3ec\uc778\ud2b8 \ubcfc\ub968 \ub9c8\uc6b4\ud2b8, \ud559\uc2b5 \uc2a4\ud06c\ub9bd\ud2b8 \ubcf5\uc81c \ubc0f 3\ub2e8\uacc4\uc5d0\uc11c \ubcf4\uc5ec\uc900 \ud559\uc2b5 \uc2dc\uc791 \uc791\uc5c5\uc744 \uad00\ub9ac\ud569\ub2c8\ub2e4.<\/p>\n<p>\uc6a9\ub7c9\uc774 \ubd80\uc871\ud558\uac70\ub098 \ube44\uc2fc \uc2a4\ud31f \uc778\uc2a4\ud134\uc2a4 \uc694\uae08 \ub54c\ubb38\uc5d0 <a href=\"https:\/\/docs.aws.amazon.com\/AWSEC2\/latest\/UserGuide\/spot-interruptions.html\" target=\"_blank\" rel=\"noopener noreferrer\">\uc2a4\ud31f \uc911\ub2e8<\/a>\uc774 \ubc1c\uc0dd\ud558\uba74, \uc778\uc2a4\ud134\uc2a4\uac00 \uc885\ub8cc\ub418\uace0 \ub370\uc774\ud130 \uc138\ud2b8 \ubc0f \uccb4\ud06c\ud3ec\uc778\ud2b8 Amazon EBS \ubcfc\ub968\uc774 \ubd84\ub9ac\ub429\ub2c8\ub2e4. \uadf8\ub9ac\uace0 \uc2a4\ud31f \ud50c\ub9bf\uc740 \uc911\ub2e8\ub41c \uc778\uc2a4\ud134\uc2a4\ub97c \uc790\ub3d9\uc73c\ub85c \ubcf4\ucda9\ud558\ub294 \ub610 \ub2e4\ub978 \uc694\uccad\uc744 \uc81c\ucd9c\ud569\ub2c8\ub2e4.<\/p>\n<p>\uc694\uccad\uc774 \ub2e4\uc2dc \uc774\ud589\ub418\uba74 \uc0c8 \uc2a4\ud31f \uc778\uc2a4\ud134\uc2a4\uac00 \uc2dc\uc791\ub418\uace0, \uc2dc\uc791 \uc2dc <code class=\"lang-code\">user_data_script.sh<\/code>\ub97c \uc2e4\ud589\ud569\ub2c8\ub2e4. \uc774 \uc2a4\ud06c\ub9bd\ud2b8\ub294 \ub370\uc774\ud130 \uc138\ud2b8 \ubc0f \uccb4\ud06c\ud3ec\uc778\ud2b8 \ubcfc\ub968\uc744 \ucffc\ub9ac\ud569\ub2c8\ub2e4. \ubcfc\ub968 \ubc0f \uc778\uc2a4\ud134\uc2a4\uac00 \uc11c\ub85c \ub2e4\ub978 \uac00\uc6a9 \uc601\uc5ed\uc5d0 \uc788\uc73c\uba74, \uba3c\uc800 \ubcfc\ub968\uc758 \uc2a4\ub0c5\uc0f7\uc744 \uc0dd\uc131\ud558\uace0 \ud604\uc7ac \uc778\uc2a4\ud134\uc2a4\uc758 \uac00\uc6a9 \uc601\uc5ed\uc5d0 \uc2a4\ub0c5\uc0f7\uc5d0 \uae30\ubc18\ud55c \uc0c8 \ubcfc\ub968\uc744 \uc0dd\uc131\ud569\ub2c8\ub2e4. \ub2e8\uc77c \ucd9c\ucc98\ub97c \ubcf4\uc7a5\ud558\uae30 \uc704\ud574 \uc774\uc804 \uac00\uc6a9 \uc601\uc5ed\uc758 \ubcfc\ub968\uc740 \uc0ad\uc81c\ub429\ub2c8\ub2e4.<\/p>\n<p>\uadf8\ub9ac\uace0 \uc2a4\ud06c\ub9bd\ud2b8\ub294 \uc778\uc2a4\ud134\uc2a4\uc5d0 \ubcfc\ub968\uc744 \uc5f0\uacb0\ud558\uace0 \ucd5c\uc2e0 \uccb4\ud06c\ud3ec\uc778\ud2b8\ubd80\ud130 \ud559\uc2b5\uc744 \uc7ac\uac1c\ud569\ub2c8\ub2e4. \ud559\uc2b5\uc774 \uc644\ub8cc\ub418\uba74 \uc2a4\ud31f \ud50c\ub9bf \uc694\uccad\uc774 \ucde8\uc18c\ub418\uace0 \ud604\uc7ac \uc2e4\ud589 \uc911\uc778 \uc778\uc2a4\ud134\uc2a4\ub294 \uc885\ub8cc\ub429\ub2c8\ub2e4.<\/p>\n<p>\ub354 \ube44\uc2fc \ucd5c\ub300 \uc2a4\ud31f \uc778\uc2a4\ud134\uc2a4 \uc694\uae08\uc744 \uc9c0\uc815\ud558\uac70\ub098 \uc778\uc2a4\ud134\uc2a4 \uc720\ud615 \ub610\ub294 \uac00\uc6a9 \uc601\uc5ed\uc744 \ubcc0\uacbd\ud558\ub824\ub294 \uacbd\uc6b0 <code class=\"lang-code\">aws ec2 cancel-spot-fleet-requests<\/code>\ub97c \uc2e4\ud589\ud558\uace0 \uc5c5\ub370\uc774\ud2b8\ub41c \uc2a4\ud31f \ud50c\ub9bf \uad6c\uc131 \ud30c\uc77c <code class=\"lang-code\">spot_fleet_config.json<\/code>\uc5d0\uc11c \uc0c8 \uc694\uccad\uc744 \uc2dc\uc791\ud558\uc5ec \uc2e4\ud589 \uc911\uc778 \uc2a4\ud31f \ud50c\ub9bf \uc694\uccad\uc744 \ucde8\uc18c\ud558\uae30\ub9cc \ud558\uba74 \ub429\ub2c8\ub2e4.<\/p>\n<h2>\uc694\uc57d<\/h2>\n<p>\ubcf8 \uac8c\uc2dc\ubb3c\uc740 \uc2a4\ud31f \uc778\uc2a4\ud134\uc2a4\ub97c \uc0ac\uc6a9\ud558\uc5ec \uc628\ub514\ub9e8\ub4dc \uc778\uc2a4\ud134\uc2a4\ubcf4\ub2e4 \ud6e8\uc52c \ub354 \uc800\ub834\ud55c \ube44\uc6a9\uc73c\ub85c GPU \uc778\uc2a4\ud134\uc2a4\uc5d0\uc11c \ub525\ub7ec\ub2dd \ud559\uc2b5 \uc2e4\ud5d8\uc744 \uc2e4\ud589\ud558\ub294 \ubc29\ubc95\uc5d0 \ub300\ud55c \uac1c\uc694\uc744 \uc124\uba85\ud569\ub2c8\ub2e4.<\/p>\n<p>\uc774 \ube14\ub85c\uadf8 \uac8c\uc2dc\ubb3c\uc5d0 \ub098\uc628 \uc124\uc815\uc740 \ub2e4\uc74c\uacfc \uac19\uc774 \uace0\uae09 \ub525\ub7ec\ub2dd \uc6cc\ud06c\ud50c\ub85c\uc6b0\ub97c \ub2e4\ub8e8\ub3c4\ub85d \ud655\uc7a5\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n<ul>\n<li>\ub2e4\uc911 GPU \ud559\uc2b5. \ub2e4\uc911 GPU \ud559\uc2b5\uc744 \uc9c0\uc6d0\ud558\ub3c4\ub85d \ud559\uc2b5 \uc2a4\ud06c\ub9bd\ud2b8 \uc5c5\ub370\uc774\ud2b8.<\/li>\n<li>\ud558\uc704 epoch \uc138\ubd84\ud654 \uccb4\ud06c\ud3ec\uc778\ud2b8 \uc9c0\uc815 \ubc0f \uc7ac\uac1c. \uc774 \uc608\uc81c\uc5d0\uc11c, \uccb4\ud06c\ud3ec\uc778\ud2b8\ub294 \uac01 epoch\uac00 \ub05d\ub098\ub294 \uc2dc\uc810\uc5d0\ub9cc \uc800\uc7a5\ub429\ub2c8\ub2e4. epoch\ub97c \uc644\ub8cc\ud558\ub294 \ub370 \ub354 \uc624\ub798 \uac78\ub9ac\ub294 \ud070 \ub370\uc774\ud130 \uc138\ud2b8\uc640 \ubcf5\uc7a1\ud55c \ubaa8\ub378\uc758 \uacbd\uc6b0 \uccb4\ud06c\ud3ec\uc778\ud2b8\ub97c \uc790\uc8fc \uc9c0\uc815\ud558\uba74 \uc911\ub2e8 \uc911\uc5d0 \uc9c4\ud589 \uc0c1\ud669 \uc190\uc2e4\uc744 \ucd5c\uc18c\ud654\ud569\ub2c8\ub2e4.<\/li>\n<li>\ub2e4\uc911 \ubcd1\ub82c \uc2e4\ud5d8. \uc11c\ub85c \ub2e4\ub978 \ud558\uc774\ud37c \ud30c\ub77c\ubbf8\ud130\ub97c \ud1b5\ud574 \uc5ec\ub7ec \uac1c\uc758 \ub3c5\ub9bd\ub41c \ud559\uc2b5 \uc791\uc5c5\uc744 \uc2e4\ud589\ud558\ub3c4\ub85d \uc2a4\ud31f \ud50c\ub9bf \ub300\uc0c1 \uc6a9\ub7c9\uc744 \ub298\ub9bd\ub2c8\ub2e4.<\/li>\n<\/ul>\n<p>\ubcf8 \uac8c\uc2dc\ubb3c\uc774 \uc5ec\ub7ec\ubd84\uc5d0\uac8c \ub3c4\uc6c0\uc774 \ub418\uc5c8\uae30\ub97c \ubc14\ub78d\ub2c8\ub2e4. \uad81\uae08\ud55c \uc810\uc774\ub098 \uc758\uacac\uc774 \uc788\uc73c\uc2dc\uba74 \uc544\ub798 \ub313\uae00 \uc139\uc158\uc744 \uc774\uc6a9\ud574 \uc8fc\uc2ed\uc2dc\uc624. \uc2a4\ud31f \ud559\uc2b5\uc744 \uc9c1\uc811 \uad6c\ud604\ud574 \ubcf4\uc2ed\uc2dc\uc624!<\/p>\n<hr \/>\n<h3>\uc791\uc131\uc790 \uc18c\uac1c<\/h3>\n<p><img loading=\"lazy\" decoding=\"async\" class=\"size-full wp-image-8054 alignleft\" src=\"https:\/\/d2908q01vomqb2.cloudfront.net\/f1f836cb4ea6efb2a0b1b99f41ad8b103eff4b59\/2019\/03\/26\/shashank-prasanna-100.jpg\" alt=\"\" width=\"100\" height=\"138\" \/>Shashank Prasanna \ub294 Amazon Web Services(AWS)\uc5d0 \uadfc\ubb34\ud558\ub294 AI \ubc0f Machine Learning \uae30\uc220 \uc5d0\ubc18\uc824\ub9ac\uc2a4\ud2b8\uc774\uba70, \uc5d4\uc9c0\ub2c8\uc5b4, \uac1c\ubc1c\uc790, \ub370\uc774\ud130 \uacfc\ud559\uc790\uac00 \uae30\uacc4 \ud559\uc2b5\uc5d0 \uad00\ud55c \ubb38\uc81c\ub97c \ud574\uacb0\ud560 \uc218 \uc788\ub3c4\ub85d \ub3c4\uc6c0\uc744 \uc8fc\ub294 \ub370 \uc8fc\ub825\ud558\uace0 \uc788\uc2b5\ub2c8\ub2e4. AWS\uc5d0 \uc785\uc0ac\ud558\uae30 \uc804\uc5d0\ub294, NVIDIA, MathWorks(MATLAB &amp; Simulink\uc758 \uc81c\uc791\uc0ac), \uadf8\ub9ac\uace0 Oracle\uc5d0\uc11c \uadfc\ubb34\ud558\uba70, \uc81c\ud488 \ub9c8\ucf00\ud305, \uc81c\ud488 \uad00\ub9ac, \uc18c\ud504\ud2b8\uc6e8\uc5b4 \uac1c\ubc1c \uc5ed\ud560\uc744 \ub9e1\uc558\uc2b5\ub2c8\ub2e4.<\/p>\n<p><em>\uc774 \uae00\uc740 <a href=\"https:\/\/aws.amazon.com\/blogs\/machine-learning\/\">AWS Machine Learning Blog<\/a>\uc758 <a href=\"https:\/\/aws.amazon.com\/blogs\/machine-learning\/train-deep-learning-models-on-gpus-using-amazon-ec2-spot-instances\/\">Train Deep Learning Models on GPUs using Amazon EC2 Spot Instances<\/a>\uc758 \ud55c\uad6d\uc5b4 \ubc88\uc5ed\uc73c\ub85c \uc815\ub3c4\ud604 AWS \ud14c\ud06c\ub2c8\uceec \ud2b8\ub808\uc774\ub108\uac00 \uac10\uc218\ud558\uc600\uc2b5\ub2c8\ub2e4.<\/em><\/p>\n<p>&nbsp;<\/p>\n<p>&nbsp;<\/p>\n<p>Source: <a href=\"https:\/\/aws.amazon.com\/ko\/blogs\/korea\/train-deep-learning-models-on-gpus-using-amazon-ec2-spot-instances\/\" target=\"_blank\" rel=\"noopener noreferrer\">GPU \ub525\ub7ec\ub2dd \ubaa8\ub378 \ud559\uc2b5\uc744 \uc704\ud55c Amazon EC2 \uc2a4\ud31f \uc778\uc2a4\ud134\uc2a4 \ud65c\uc6a9\ubc95<\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<div class=\"mh-excerpt\"><p>GPU \ub525\ub7ec\ub2dd \ubaa8\ub378 \ud559\uc2b5\uc744 \uc704\ud55c Amazon EC2 \uc2a4\ud31f \uc778\uc2a4\ud134\uc2a4 \ud65c\uc6a9\ubc95 \uc5ec\ub7ec\ubd84\uc774 \ub370\uc774\ud130 \uc138\ud2b8\ub97c \uc218\uc9d1\ud558\uace0, \uc2ec\uce35 \uc2e0\uacbd\ub9dd \uc544\ud0a4\ud14d\ucc98\ub97c \uc124\uacc4\ud558\uace0, \ud559\uc2b5 \ub8e8\ud2f4 \ucf54\ub4dc\ub97c \uc791\uc131\uc744 \uc644\ub8cc \ud558\uc168\ub2e4\uba74 \uc9c0\uae08\ubd80\ud130\ub294 \uac15\ub825\ud55c GPU \uc778\uc2a4\ud134\uc2a4\ub97c \uc0ac\uc6a9\ud574 \uc5ec\ub7ec epoch\uc5d0 \uac78\uccd0 \ub300\uaddc\ubaa8 \ub370\uc774\ud130 \uc138\ud2b8\ub97c \uc774\uc6a9\ud55c \ud559\uc2b5\uc744 \uc2e4\ud589\ud560 \ucc28\ub840\uc785\ub2c8\ub2e4. \uc5ec\ub7ec\ubd84\uc740 \uc774\ubbf8 NVIDIA Tesla V100 GPU\uc5d0 \uae30\ubc18\ud55c Amazon EC2 P3 \uc778\uc2a4\ud134\uc2a4\uac00 \ucef4\ud4e8\ud305 \uc9d1\uc57d\uc801\uc778 \ub525\ub7ec\ub2dd \ud559\uc2b5 \uc791\uc5c5\uc5d0 \uc801\ud569\ud558\ub2e4\ub294 \uc0ac\uc2e4\uc744 \uc54c\uace0 \uc788\uc9c0\ub9cc&nbsp;\uc608\uc0b0\uc774 \ube60\ub4ef\ud55c \uad00\uacc4\ub85c \uac00\ub2a5\ud55c\ud55c \ud559\uc2b5 \ube44\uc6a9\uc740 \ub0ae\ucd94\uace0 \uc2f6\uc2b5\ub2c8\ub2e4. \uba87 \uc2dc\uac04 \ub610\ub294 \uba70\uce60\uc5d0 \uac78\uccd0 \ud559\uc2b5 \uc791\uc5c5\uc744 \uc2e4\ud589\ud558\ub294 \ub525\ub7ec\ub2dd \uc5f0\uad6c\uc790 \ubc0f \uac1c\ubc1c\uc790\ub77c\uba74, \uc2a4\ud31f \uc778\uc2a4\ud134\uc2a4 \uc694\uae08\uc744 \ud1b5\ud574 \ud6e8\uc52c \uacbd\uc81c\uc801\uc778 \ube44\uc6a9\uc73c\ub85c \uace0\uc131\ub2a5 CPU\ub97c \uc774\uc6a9\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \uc2a4\ud31f \uc778\uc2a4\ud134\uc2a4\ub97c \uc0ac\uc6a9\ud558\uba74 \uc628\ub514\ub9e8\ub4dc \uc694\uae08\uc5d0 \ube44\ud574 \ub300\ud3ed \ud560\uc778\ub41c \uac00\uaca9\uc73c\ub85c \uc5ec\ubd84\uc758 Amazon EC2 \ucef4\ud4e8\ud305 \uc6a9\ub7c9\uc5d0 \uc561\uc138\uc2a4\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \uc778\uc2a4\ud134\uc2a4 \ubc0f \ub9ac\uc804\ubcc4 \ucd5c\uadfc \uc694\uae08 \ubaa9\ub85d\uc744 \ubcf4\ub824\uba74 \uc2a4\ud31f \uc778\uc2a4\ud134\uc2a4 \uc5b4\ub4dc\ubc14\uc774\uc800\ub97c \ubc29\ubb38\ud558\uc2ed\uc2dc\uc624. \uc2a4\ud31f \uc778\uc2a4\ud134\uc2a4\uc640 \uc628\ub514\ub9e8\ub4dc \uc778\uc2a4\ud134\uc2a4 \uc0ac\uc774\uc758 \uc8fc\uc694 \ucc28\uc774\uc810\uc744 \uc54c\uc544\ubcf4\ub824\uba74 \uc774 Amazon EC2 \uc0ac\uc6a9 \uc124\uba85\uc11c\ub97c \uac80\ud1a0\ud574 \ubcf4\uc2dc\uae30 \ubc14\ub78d\ub2c8\ub2e4. \uc2a4\ud31f \uc778\uc2a4\ud134\uc2a4\ub294 \ub525\ub7ec\ub2dd \uc6cc\ud06c\ud50c\ub85c\uc6b0\uc5d0 \uc801\ud569\ud558\uc9c0\ub9cc, \uc628\ub514\ub9e8\ub4dc <a class=\"mh-excerpt-more\" href=\"https:\/\/jirak.net\/wp\/gpu-%eb%94%a5%eb%9f%ac%eb%8b%9d-%eb%aa%a8%eb%8d%b8-%ed%95%99%ec%8a%b5%ec%9d%84-%ec%9c%84%ed%95%9c-amazon-ec2-%ec%8a%a4%ed%8c%9f-%ec%9d%b8%ec%8a%a4%ed%84%b4%ec%8a%a4-%ed%99%9c%ec%9a%a9%eb%b2%95\/\" title=\"GPU \ub525\ub7ec\ub2dd \ubaa8\ub378 \ud559\uc2b5\uc744 \uc704\ud55c Amazon EC2 \uc2a4\ud31f \uc778\uc2a4\ud134\uc2a4 \ud65c\uc6a9\ubc95\">[ more&#8230; ]<\/a><\/p>\n<\/div>","protected":false},"author":1,"featured_media":31379,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[169],"tags":[656],"class_list":["post-31378","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\/31378","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=31378"}],"version-history":[{"count":1,"href":"https:\/\/jirak.net\/wp\/wp-json\/wp\/v2\/posts\/31378\/revisions"}],"predecessor-version":[{"id":31380,"href":"https:\/\/jirak.net\/wp\/wp-json\/wp\/v2\/posts\/31378\/revisions\/31380"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/jirak.net\/wp\/wp-json\/wp\/v2\/media\/31379"}],"wp:attachment":[{"href":"https:\/\/jirak.net\/wp\/wp-json\/wp\/v2\/media?parent=31378"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/jirak.net\/wp\/wp-json\/wp\/v2\/categories?post=31378"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/jirak.net\/wp\/wp-json\/wp\/v2\/tags?post=31378"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}