\u30cb\u30e5\u30fc\u30e9\u30eb\u30cd\u30c3\u30c8\u30ef\u30fc\u30af\u306e\u5165\u9580\u66f8\u306b\u8f09\u3063\u3066\u3044\u308b\u666e\u901a\u306e\u30b7\u30f3\u30d7\u30eb\u306a\u30cd\u30c3\u30c8\u30ef\u30fc\u30af\u69cb\u6210\u3067\u5b9f\u88c5\u3057\u3066\u307f\u305f\u3002<\/p>\n
\u30d7\u30ed\u30b0\u30e9\u30e0\u306e\u4ed5\u69d8\u306f\u4ee5\u4e0b\u306e\u901a\u308a\u3002
\n(1) MNIST\u753b\u50cf\u30c7\u30fc\u30bf60,000\u679a\u3092\u5b66\u7fd2
\n(2) MNIST\u30c6\u30b9\u30c8\u30c7\u30fc\u30bf 10,000\u679a\u3092\u30c6\u30b9\u30c8
\n(3) \u5165\u529b\u5c64\u306e\u30e6\u30cb\u30c3\u30c8\u6570\u306f28×28\u306e784\u500b
\n(4) \u51fa\u529b\u5c64\u306e\u30e6\u30cb\u30c3\u30c8\u6570\u306f\u6570\u5b570\uff5e9\u306b\u5bfe\u5fdc\u3059\u308b10\u500b
\n(5) \u96a0\u308c\u5c64\u306e\u5c64\u6570\u3068\u5404\u5c64\u306e\u30e6\u30cb\u30c3\u30c8\u6570\u306f\u5f15\u6570\u3067\u4efb\u610f\u306b\u6307\u5b9a\u53ef\u80fd
\n(6) \u6570\u5b57\u5225\u306b\u6b63\u89e3\u7387\u3092\u8868\u793a
\n(7) \u5404\u5c64\u306e\u6d3b\u6027\u5316\u95a2\u6570\u306fSigmoid
\n(8) \u8aa4\u5dee\u95a2\u6570\u306f\u4e8c\u4e57\u548c\u8aa4\u5dee
\n(9) Epoch\u6570\u306f1
\n(10) 1\u5b66\u7fd2\u30c7\u30fc\u30bf\u306e\u5b66\u7fd2\u3054\u3068\u306b\u30d1\u30e9\u30e1\u30fc\u30bf\u30fc\u3092\u66f4\u65b0<\/p>\n
\u9069\u5f53\u306b\u9078\u3093\u3060\u51689\u30d1\u30bf\u30fc\u30f3\u306e\u5c64\u69cb\u6210\u3092\u8a66\u3057\u305f\u7d50\u679c\u306f\u4ee5\u4e0b\u306e\u3088\u3046\u306b\u306a\u3063\u305f\u3002 (1) 3\u5c64 [784]-[16]-[10] 89.0%<\/span><\/p>\n (2) 3\u5c64 [784]-[24]-[10] 89.9%<\/span><\/p>\n (3) 3\u5c64 [784]-[32]-[10] 90.0%<\/span><\/p>\n (4) 3\u5c64 [784]-[48]-[10] 90.6%<\/span><\/p>\n (5) 3\u5c64 [784]-[64]-[10] 90.8%<\/span><\/p>\n (6) 3\u5c64 [784]-[96]-[10] 90.4%<\/span><\/p>\n (7) 3\u5c64 [784]-[128]-[10] 91.5%<\/span><\/p>\n (8) 4\u5c64 [784]-[256]-[32]-[10] 91.5%<\/span><\/p>\n (9) 4\u5c64 [784]-[256]-[64]-[10] 91.9%<\/span><\/p>\n \u6b21\u56de\u300c(5) EPOCH\u6570\u3092\u5897\u3084\u3057\u3066\u6b63\u89e3\u7387\u4e0a\u6607\u300d\u3067\u306f\u3001\u5b66\u7fd2\u3092\u7e70\u308a\u8fd4\u3059\u3053\u3068\u306b\u3088\u308a\u6b63\u89e3\u7387\u3092\u5411\u4e0a\u3055\u305b\u3066\u307f\u308b\u3002<\/a><\/p>\n 1. \u30d7\u30ed\u30b0\u30e9\u30e0\u306e\u4ed5\u69d8 \u30cb\u30e5\u30fc\u30e9\u30eb\u30cd\u30c3\u30c8\u30ef\u30fc\u30af\u306e\u5165\u9580\u66f8\u306b\u8f09\u3063\u3066\u3044\u308b\u666e\u901a\u306e\u30b7\u30f3\u30d7\u30eb\u306a\u30cd\u30c3\u30c8\u30ef\u30fc\u30af\u69cb\u6210\u3067\u5b9f\u88c5\u3057\u3066\u307f\u305f\u3002 \u30d7\u30ed\u30b0\u30e9\u30e0\u306e\u4ed5\u69d8\u306f\u4ee5\u4e0b\u306e\u901a\u308a\u3002 (1) MNIST\u753b\u50cf\u30c7\u30fc\u30bf60,000\u679a\u3092\u5b66\u7fd2 (2) MNIST\u30c6\u30b9\u2026 \u7d9a\u304d\u3092\u8aad\u3080 »<\/a><\/span><\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[18,6],"tags":[],"class_list":["post-43","post","type-post","status-publish","format-standard","hentry","category-mnist","category-6"],"views":8588,"amp_enabled":true,"_links":{"self":[{"href":"https:\/\/www.dogrow.net\/nnet\/wp-json\/wp\/v2\/posts\/43","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.dogrow.net\/nnet\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.dogrow.net\/nnet\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.dogrow.net\/nnet\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.dogrow.net\/nnet\/wp-json\/wp\/v2\/comments?post=43"}],"version-history":[{"count":31,"href":"https:\/\/www.dogrow.net\/nnet\/wp-json\/wp\/v2\/posts\/43\/revisions"}],"predecessor-version":[{"id":2536,"href":"https:\/\/www.dogrow.net\/nnet\/wp-json\/wp\/v2\/posts\/43\/revisions\/2536"}],"wp:attachment":[{"href":"https:\/\/www.dogrow.net\/nnet\/wp-json\/wp\/v2\/media?parent=43"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.dogrow.net\/nnet\/wp-json\/wp\/v2\/categories?post=43"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.dogrow.net\/nnet\/wp-json\/wp\/v2\/tags?post=43"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}
\n\u3053\u306e\u4e2d\u3067\u306f 4\u5c64 [784]-[256]-[64]-[10] \u306e\u6b63\u89e3\u7387 91.9%<\/span> \u304c\u6700\u3082\u9ad8\u304b\u3063\u305f\u3002<\/p>\n\r\noctave:10> tic;NNET_control([784 16 10]);toc;\r\n[ 0] 938 \/ 980 ( 95.7%)\r\n[ 1] 1109 \/ 1135 ( 97.7%)\r\n[ 2] 888 \/ 1032 ( 86.0%)\r\n[ 3] 892 \/ 1010 ( 88.3%)\r\n[ 4] 878 \/ 982 ( 89.4%)\r\n[ 5] 719 \/ 892 ( 80.6%)\r\n[ 6] 889 \/ 958 ( 92.8%)\r\n[ 7] 896 \/ 1028 ( 87.2%)\r\n[ 8] 803 \/ 974 ( 82.4%)\r\n[ 9] 890 \/ 1009 ( 88.2%)\r\nTotal 8902 \/ 10000 ( 89.0%)\r\nElapsed time is 63.3441 seconds.<\/pre>\n
\r\noctave:7> tic;NNET_control([784 24 10]);toc;\r\n[ 0] 949 \/ 980 ( 96.8%)\r\n[ 1] 1109 \/ 1135 ( 97.7%)\r\n[ 2] 907 \/ 1032 ( 87.9%)\r\n[ 3] 907 \/ 1010 ( 89.8%)\r\n[ 4] 887 \/ 982 ( 90.3%)\r\n[ 5] 708 \/ 892 ( 79.4%)\r\n[ 6] 882 \/ 958 ( 92.1%)\r\n[ 7] 925 \/ 1028 ( 90.0%)\r\n[ 8] 858 \/ 974 ( 88.1%)\r\n[ 9] 856 \/ 1009 ( 84.8%)\r\nTotal 8988 \/ 10000 ( 89.9%)\r\nElapsed time is 65.8959 seconds.<\/pre>\n
\r\noctave:6> tic;NNET_control([784 32 10]);toc;\r\n[ 0] 950 \/ 980 ( 96.9%)\r\n[ 1] 1108 \/ 1135 ( 97.6%)\r\n[ 2] 885 \/ 1032 ( 85.8%)\r\n[ 3] 905 \/ 1010 ( 89.6%)\r\n[ 4] 894 \/ 982 ( 91.0%)\r\n[ 5] 718 \/ 892 ( 80.5%)\r\n[ 6] 895 \/ 958 ( 93.4%)\r\n[ 7] 945 \/ 1028 ( 91.9%)\r\n[ 8] 828 \/ 974 ( 85.0%)\r\n[ 9] 869 \/ 1009 ( 86.1%)\r\nTotal 8997 \/ 10000 ( 90.0%)\r\nElapsed time is 71.066 seconds.<\/pre>\n
\r\noctave:9> tic;NNET_control([784 48 10]);toc;\r\n[ 0] 955 \/ 980 ( 97.4%)\r\n[ 1] 1117 \/ 1135 ( 98.4%)\r\n[ 2] 909 \/ 1032 ( 88.1%)\r\n[ 3] 931 \/ 1010 ( 92.2%)\r\n[ 4] 920 \/ 982 ( 93.7%)\r\n[ 5] 746 \/ 892 ( 83.6%)\r\n[ 6] 892 \/ 958 ( 93.1%)\r\n[ 7] 917 \/ 1028 ( 89.2%)\r\n[ 8] 824 \/ 974 ( 84.6%)\r\n[ 9] 847 \/ 1009 ( 83.9%)\r\nTotal 9058 \/ 10000 ( 90.6%)\r\nElapsed time is 74.398 seconds.<\/pre>\n
\r\noctave:1> tic;NNET_control([784 64 10]);toc;\r\n[ 0] 949 \/ 980 ( 96.8%)\r\n[ 1] 1109 \/ 1135 ( 97.7%)\r\n[ 2] 907 \/ 1032 ( 87.9%)\r\n[ 3] 901 \/ 1010 ( 89.2%)\r\n[ 4] 898 \/ 982 ( 91.4%)\r\n[ 5] 745 \/ 892 ( 83.5%)\r\n[ 6] 898 \/ 958 ( 93.7%)\r\n[ 7] 929 \/ 1028 ( 90.4%)\r\n[ 8] 859 \/ 974 ( 88.2%)\r\n[ 9] 888 \/ 1009 ( 88.0%)\r\nTotal 9083 \/ 10000 ( 90.8%)\r\nElapsed time is 81.1507 seconds.<\/pre>\n
\r\noctave:4> tic;NNET_control([784 96 10]);toc;\r\n[ 0] 953 \/ 980 ( 97.2%)\r\n[ 1] 1114 \/ 1135 ( 98.1%)\r\n[ 2] 910 \/ 1032 ( 88.2%)\r\n[ 3] 915 \/ 1010 ( 90.6%)\r\n[ 4] 851 \/ 982 ( 86.7%)\r\n[ 5] 720 \/ 892 ( 80.7%)\r\n[ 6] 888 \/ 958 ( 92.7%)\r\n[ 7] 937 \/ 1028 ( 91.1%)\r\n[ 8] 865 \/ 974 ( 88.8%)\r\n[ 9] 888 \/ 1009 ( 88.0%)\r\nTotal 9041 \/ 10000 ( 90.4%)\r\nElapsed time is 112.633 seconds.<\/pre>\n
\r\noctave:5> tic;NNET_control([784 128 10]);toc;\r\n[ 0] 951 \/ 980 ( 97.0%)\r\n[ 1] 1112 \/ 1135 ( 98.0%)\r\n[ 2] 931 \/ 1032 ( 90.2%)\r\n[ 3] 897 \/ 1010 ( 88.8%)\r\n[ 4] 879 \/ 982 ( 89.5%)\r\n[ 5] 784 \/ 892 ( 87.9%)\r\n[ 6] 892 \/ 958 ( 93.1%)\r\n[ 7] 927 \/ 1028 ( 90.2%)\r\n[ 8] 868 \/ 974 ( 89.1%)\r\n[ 9] 904 \/ 1009 ( 89.6%)\r\nTotal 9145 \/ 10000 ( 91.5%)\r\nElapsed time is 152.437 seconds.<\/pre>\n
\r\noctave:12> tic;NNET_control([784 256 32 10]);toc;\r\n[ 0] 957 \/ 980 ( 97.7%)\r\n[ 1] 1098 \/ 1135 ( 96.7%)\r\n[ 2] 929 \/ 1032 ( 90.0%)\r\n[ 3] 901 \/ 1010 ( 89.2%)\r\n[ 4] 876 \/ 982 ( 89.2%)\r\n[ 5] 781 \/ 892 ( 87.6%)\r\n[ 6] 899 \/ 958 ( 93.8%)\r\n[ 7] 942 \/ 1028 ( 91.6%)\r\n[ 8] 862 \/ 974 ( 88.5%)\r\n[ 9] 902 \/ 1009 ( 89.4%)\r\nTotal 9147 \/ 10000 ( 91.5%)\r\nElapsed time is 301.459 seconds.<\/pre>\n
\r\noctave:11> tic;NNET_control([784 256 64 10]);toc;\r\n[ 0] 960 \/ 980 ( 98.0%)\r\n[ 1] 1109 \/ 1135 ( 97.7%)\r\n[ 2] 904 \/ 1032 ( 87.6%)\r\n[ 3] 882 \/ 1010 ( 87.3%)\r\n[ 4] 909 \/ 982 ( 92.6%)\r\n[ 5] 786 \/ 892 ( 88.1%)\r\n[ 6] 910 \/ 958 ( 95.0%)\r\n[ 7] 936 \/ 1028 ( 91.1%)\r\n[ 8] 874 \/ 974 ( 89.7%)\r\n[ 9] 921 \/ 1009 ( 91.3%)\r\nTotal 9191 \/ 10000 ( 91.9%)\r\nElapsed time is 309.785 seconds.<\/pre>\n
3. \u30d7\u30ed\u30b0\u30e9\u30e0\u306e\u30bd\u30fc\u30b9\u30b3\u30fc\u30c9<\/h1>\n
(1) NNET_control.m<\/h3>\n
\r\nfunction NNET_control( num_unit_of_each_layer )\r\n\r\n <span class="my_fc_green">% \u5b66\u7fd2\u753b\u50cf\u30fb\u30e9\u30d9\u30eb\u3001\u30c6\u30b9\u30c8\u753b\u50cf\u30fb\u30e9\u30d9\u30eb\u3092\u30d5\u30a1\u30a4\u30eb\u304b\u3089\u8aad\u307f\u8fbc\u307f<\/span>\r\n [train_img, train_lbl] = load_MNIST( 'train-images-idx3-ubyte', 'train-labels-idx1-ubyte' );\r\n [test_img, test_lbl ] = load_MNIST( 't10k-images-idx3-ubyte', 't10k-labels-idx1-ubyte' );\r\n\r\n <span class="my_fc_green">% \u5404\u753b\u50cf\u30c7\u30fc\u30bf\u3092 0.0\uff5e1.0\u306e\u7bc4\u56f2\u306b\u6b63\u898f\u5316<\/span>\r\n train_img = train_img \/ 255;\r\n test_img = test_img \/ 255;\r\n\r\n <span class="my_fc_green">% \u6307\u5b9a\u3055\u308c\u305f\u5c64\u6570\u3001\u30e6\u30cb\u30c3\u30c8\u6570\u3067\u30cb\u30e5\u30fc\u30e9\u30eb\u30cd\u30c3\u30c8\u30ef\u30fc\u30af\u3092\u4f5c\u6210<\/span>\r\n nn = NNET_setup( num_unit_of_each_layer );\r\n\r\n <span class="my_fc_green">% \u5b66\u7fd2\u5b9f\u884c<\/span>\r\n nn = NNET_learn( nn, train_img, train_lbl );\r\n\r\n <span class="my_fc_green">% \u30c6\u30b9\u30c8\u5b9f\u884c<\/span>\r\n result = NNET_test( nn, test_img, test_lbl );\r\n\r\n <span class="my_fc_green">% \u30c6\u30b9\u30c8\u7d50\u679c\u3092\u8868\u793a<\/span>\r\n for i=1: 10\r\n printf('[%2d] %4d \/ %4d (%5.1f%%) \\n', i-1, result(i,2), result(i,1), result(i,2)\/result(i,1)*100);\r\n end\r\n sum_result = sum(result,1);\r\n printf('Total %5d \/ %5d (%5.1f%%) \\n', sum_result(2), sum_result(1), sum_result(2)\/sum_result(1)*100);\r\nend\r\n<\/pre>\n(2) load_MNIST.m<\/h3>\n
\r\nfunction [image label] = load_MNIST( image_file, label_file )\r\n <span class="my_fc_green">%\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\r\n % \u753b\u50cf\u3092\u30ed\u30fc\u30c9\r\n % \u30d5\u30a1\u30a4\u30eb\u304b\u3089\u30ed\u30fc\u30c9\u3057\u305f\u753b\u50cf\u3092 image[28][28][60000]\u306e\u914d\u5217\u3067\u51fa\u529b<\/span>\r\n fid = fopen(image_file,'r','b');\r\n magic_number = fread(fid, 1, 'int32');\r\n number_of_items = fread(fid, 1, 'int32');\r\n number_of_rows = fread(fid, 1, 'int32');\r\n number_of_columns = fread(fid, 1, 'int32');\r\n img = fread(fid, [number_of_rows*number_of_columns number_of_items],'uint8');\r\n image = permute(reshape(img, number_of_rows, number_of_columns,number_of_items),[2 1 3]);\r\n fclose(fid);\r\n\r\n <span class="my_fc_green">%\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\r\n % \u30e9\u30d9\u30eb\u3092\u30ed\u30fc\u30c9\r\n % \u30d5\u30a1\u30a4\u30eb\u304b\u3089\u30ed\u30fc\u30c9\u3057\u305f\u30e9\u30d9\u30eb\u3092 label[10][60000]\u306e\u914d\u5217\u3067\u51fa\u529b<\/span>\r\n fid = fopen(label_file,'r','b');\r\n magic_number = fread(fid, 1, 'int32');\r\n number_of_items = fread(fid, 1, 'int32');\r\n lbl = fread(fid, number_of_items, 'uint8');\r\n idx = [1:number_of_items]';\r\n lblidx = lbl * number_of_items + idx;\r\n label = zeros(number_of_items,10);\r\n label(lblidx) = 1;\r\n label = label';\r\n fclose(fid);\r\nend\r\n<\/pre>\n
(3) NNET_setup.m<\/h3>\n
\r\nfunction nn = NNET_setup( num_unit_of_each_layer )\r\n\r\n % \u4e71\u6570\u751f\u6210\u5668\u3092\u521d\u671f\u5316\r\n rand('seed', 0);\r\n\r\n % \u6307\u5b9a\u3055\u308c\u305f\u5c64\u6570\u3092\u53d6\u5f97\r\n num_layer = numel( num_unit_of_each_layer );\r\n\r\n % \u5168\u5c64\u3092\u521d\u671f\u5316\r\n for i=2 : num_layer\r\n\r\n % \u73fe\u5c64\u3068\u524d\u5c64\u306e\u30e6\u30cb\u30c3\u30c8\u6570\u3092\u53d6\u5f97\r\n num_unit_pre = num_unit_of_each_layer( i - 1 );\r\n num_unit = num_unit_of_each_layer( i );\r\n\r\n % \u5404\u7d50\u5408\u7dda\u306e\u8377\u91cd\u3092 -1\uff5e1\u306e\u4e00\u69d8\u5206\u5e03\u4e71\u6570\u3067\u521d\u671f\u5316\r\n nn.layer{i}.weight = -1 + rand( num_unit, num_unit_pre ) * 2;\r\n\r\n % \u30d0\u30a4\u30a2\u30b9\u3092\u521d\u671f\u5316\r\n nn.layer{i}.bias = zeros( num_unit, 1 );\r\n end\r\nend\r\n<\/pre>\n(4) NNET_learn.m<\/h3>\n
\r\nfunction nn = NNET_learn( nn, train_img, train_lbl )\r\n\r\n % \u5b66\u7fd2\u30c7\u30fc\u30bf\u6570\u3092\u53d6\u5f97\r\n num_data = size( train_img, 3 );\r\n\r\n % \u5b66\u7fd2\u30c7\u30fc\u30bf\u3092\u30b7\u30e3\u30c3\u30d5\u30eb\r\n randvector = randperm( num_data );\r\n train_img = train_img(:,:,randvector);\r\n train_lbl = train_lbl(:,randvector);\r\n\r\n % \u5168\u5b66\u7fd2\u753b\u50cf\u30921\u500b\u305a\u3064\u5b66\u7fd2\u3055\u305b\u308b\r\n for i=1 : num_data\r\n % \u9806\u4f1d\u64ad\r\n nn = NNET_propagation_forward( nn, train_img(:,:,i) );\r\n\r\n % \u8aa4\u5dee\u9006\u4f1d\u64ad\r\n nn = NNET_propagation_back( nn, train_lbl(:,i) );\r\n\r\n % \u30d1\u30e9\u30e1\u30fc\u30bf\u30fc\u66f4\u65b0\r\n nn = NNET_update( nn, 0.05 );\r\n end\r\nend\r\n<\/pre>\n
(5) NNET_propagation_forward.m<\/h3>\n
\r\nfunction nn = NNET_propagation_forward( nn, train_img )\r\n\r\n % \u5165\u529b\u5c64\u306e\u51fa\u529b\u5024\u3092\u8a18\u61b6\r\n nn.layer{1}.out = train_img(:); % [n0][1]\r\n\r\n % \u5168\u5c64\u306b\u3064\u3044\u3066\u9806\u4f1d\u64ad\u3092\u5b9f\u884c\r\n for i=2 : numel(nn.layer)\r\n\r\n % a = \u03a3wz + bias w[n1][n0] z[n0][1] bias[n1][1]\r\n nn.layer{i}.actprm = nn.layer{i}.weight * nn.layer{i-1}.out + nn.layer{i}.bias;\r\n\r\n % out = sigmoid(a) out[n1][1]\r\n nn.layer{i}.out = sigmoid( nn.layer{i}.actprm );\r\n end\r\nend\r\n<\/pre>\n(6) NNET_propagation_back.m<\/h3>\n
\r\nfunction nn = NNET_propagation_back( nn, train_lbl )\r\n\r\n % \u5c64\u6570\u3092\u53d6\u5f97\r\n num_layer = numel(nn.layer);\r\n\r\n % \u51fa\u529b\u5c64\u3067\u691c\u51fa\u3055\u308c\u305f\u8aa4\u5dee\u91cf\u3068\u9006\u4f1d\u64ad\u3059\u308b\u52fe\u914d\u306e\u521d\u671f\u5024\u3092\u7b97\u51fa\r\n [err, nn.layer{num_layer}.grad] = lossfunc(nn.layer{num_layer}.out, train_lbl);\r\n\r\n % \u5168\u5c64\u306b\u3064\u3044\u3066\u8aa4\u5dee\u9006\u4f1d\u64ad\u3092\u5b9f\u884c\r\n for i=num_layer : -1 : 2\r\n\r\n % \u76f4\u524d\u5c64\u306e\u5404\u30cb\u30e5\u30fc\u30ed\u30f3\u306b\u4f1d\u64ad\u3059\u308b\u52fe\u914d\u3092\u7b97\u51fa\r\n % \u03b4out \r\n % ----- = w x h'(a) \r\n % \u03b4in \r\n % | \u03b4out| \r\n % grad = \u03a3|gout x -----| \r\n % | \u03b4in | \r\n\r\n % \u914d\u5217\u8981\u7d20\u6570\u306e\u540c\u30582\u30d1\u30e9\u30e1\u30fc\u30bf\u3092\u5148\u306b\u8a08\u7b97 grad[n1][1] out[n1][1]\r\n derr = nn.layer{i}.grad .* dsigmoid(nn.layer{i}.out);\r\n\r\n % \u03a3(w\u30fbderr) w[n1][n0] derr[n1][1] grad[n0][1]\r\n nn.layer{i-1}.grad = nn.layer{i}.weight' * derr;\r\n\r\n % \u7d50\u5408\u8377\u91cd\u306e\u4fee\u6b63\u91cf\u3092\u7b97\u51fa\r\n % \u03b4E \r\n % ---- = grad\u30fbh'(a)\u30fbout \r\n % \u03b4w \r\n % IN\u5074\u30e6\u30cb\u30c3\u30c8\uff0dOUT\u5074\u30e6\u30cb\u30c3\u30c8\u306e\u7d44\u307f\u5408\u308f\u305b\u3054\u3068\u306b\u7b97\u51fa out[n0][1] derr[n1][1] dw[n1][n0]\r\n nn.layer{i}.dweight = derr * nn.layer{i-1}.out';\r\n\r\n % \u30d0\u30a4\u30a2\u30b9\u306e\u4fee\u6b63\u91cf\u3092\u7b97\u51fa\r\n nn.layer{i}.dbias = derr;\r\n end\r\nend\r\n<\/pre>\n(7) NNET_update.m<\/h3>\n
\r\nfunction nn = NNET_update( nn, ratio )\r\n\r\n % \u5168\u5c64\u306b\u3064\u3044\u3066\u7d50\u5408\u8377\u91cd\u3068\u30d0\u30a4\u30a2\u30b9\u3092\u66f4\u65b0\r\n for i=2: numel(nn.layer)\r\n nn.layer{i}.weight = nn.layer{i}.weight - ratio * nn.layer{i}.dweight;\r\n nn.layer{i}.bias = nn.layer{i}.bias - ratio * nn.layer{i}.dbias;\r\n end\r\nend\r\n<\/pre>\n(8) NNET_test.m<\/h3>\n
\r\nfunction result = NNET_test( nn, test_img, test_lbl )\r\n\r\n % \u30c6\u30b9\u30c8\u30c7\u30fc\u30bf\u6570\u3092\u53d6\u5f97\r\n num_data = size(test_img, 3);\r\n\r\n % \u30c6\u30b9\u30c8\u7d50\u679c\u306e\u8a18\u9332\u9818\u57df\u3092\u521d\u671f\u5316\r\n result = zeros(size(test_lbl,1),2);\r\n\r\n % \u5168\u5b66\u7fd2\u753b\u50cf\u3092\u30c6\u30b9\u30c8\u5b9f\u884c\r\n for i=1 : num_data\r\n % \u9806\u4f1d\u64ad\r\n nn = NNET_propagation_forward( nn, test_img(:,:,i) );\r\n\r\n % \u81ea\u52d5\u8b58\u5225\u7d50\u679c(\u51fa\u529b\u5c64\u306e\u51fa\u529b\u5024)\u3092\u53d6\u5f97\r\n result_lbl = nn.layer{numel(nn.layer)}.out;\r\n [~, idx_cor] = max(test_lbl(:,i)); % \u671f\u5f85\u5024\u3092\u53d6\u5f97\r\n [~, idx_res] = max(result_lbl); % \u5224\u5b9a\u7d50\u679c\u3092\u53d6\u5f97\r\n\r\n % \u30c6\u30b9\u30c8\u7d50\u679c\u3092\u8a18\u61b6\r\n result(idx_cor ,1) = result(idx_cor ,1) + 1; % \u6570\u5b57\u5225\u306e\u30c6\u30b9\u30c8\u6570+1\r\n if idx_cor == idx_res\r\n result(idx_cor ,2) = result(idx_cor ,2) + 1; % \u6570\u5b57\u5225\u306e\u6b63\u89e3\u6570+1\r\n end\r\n end\r\nend\r\n<\/pre>\n(9) sigmoid.m<\/h3>\n
\r\nfunction y = sigmoid( x )\r\n y = 1 .\/ (1 + exp(-x));\r\nend\r\n<\/pre>\n
(10) dsigmoid.m<\/h3>\n
\r\nfunction y = dsigmoid( x )\r\n y = x .* (1 - x);\r\nend\r\n<\/pre>\n
(11) lossfunc.m<\/h3>\n
\r\nfunction [err grad] = lossfunc( out, lbl )\r\n\r\n % \u4f1d\u64ad\u3059\u308b\u8aa4\u5dee\u306e\u521d\u671f\u5024\u3092\u7b97\u51fa\r\n grad = out - lbl; % out[n1][1] lbl[n1][1]\r\n\r\n % \u8aa4\u5dee\u95a2\u6570\u7a2e\u5225\u306f\u4e8c\u4e57\u548c\u8aa4\u5dee\u3068\u3059\u308b\u3002\r\n % 1 2 \r\n % err = ---\u03a3(out - lbl) \r\n % 2 \r\n err = grad' * grad \/ 2;\r\nend\r\n<\/pre>\n
\n","protected":false},"excerpt":{"rendered":"