{"id":265,"date":"2018-07-16T11:47:41","date_gmt":"2018-07-16T02:47:41","guid":{"rendered":"https:\/\/naru.jpn.com\/wordpress\/?p=265"},"modified":"2018-07-16T12:06:10","modified_gmt":"2018-07-16T03:06:10","slug":"metal-performance-shaders-%e3%81%ae%e3%82%af%e3%83%a9%e3%82%b9%e3%83%bb%e9%96%a2%e6%95%b0%e3%81%ae%e3%82%b3%e3%83%bc%e3%83%89%e3%83%89%e3%82%ad%e3%83%a5%e3%83%a1%e3%83%b3%e3%83%88%e5%86%85%e3%81%ab","status":"publish","type":"post","link":"https:\/\/naru.jpn.com\/wordpress\/?p=265","title":{"rendered":"Metal Performance Shaders \u306e\u30af\u30e9\u30b9\u30fb\u95a2\u6570\u306e\u30b3\u30fc\u30c9\u30c9\u30ad\u30e5\u30e1\u30f3\u30c8\u5185\u306b\u51fa\u3066\u304f\u308b\u8ad6\u6587\u306e\u4e00\u89a7"},"content":{"rendered":"<p>Metal Performance Shaders \u306b\u306f\u30cb\u30e5\u30fc\u30e9\u30eb\u30cd\u30c3\u30c8\u30ef\u30fc\u30af\u306b\u95a2\u3059\u308b\u5b9f\u88c5\u304c\u542b\u307e\u308c\u3066\u3044\u307e\u3059. \u3042\u308b\u95a2\u6570\u306e\u30b3\u30fc\u30c9\u30c9\u30ad\u30e5\u30e1\u30f3\u30c8\u3092\u773a\u3081\u3066\u3044\u305f\u3068\u3053\u308d, \u30c9\u30ad\u30e5\u30e1\u30f3\u30c8\u5185\u306b\u8ad6\u6587\u3078\u306e\u53c2\u7167\u304c\u3042\u308b\u3053\u3068\u306b\u6c17\u3065\u304d\u307e\u3057\u305f.<\/p>\n<p><!--more--><\/p>\n<p>\u30b3\u30fc\u30c9\u30c9\u30ad\u30e5\u30e1\u30f3\u30c8\u304c\u66f8\u304b\u308c\u305f\u30d5\u30a1\u30a4\u30eb\u5185\u306b\u306f, \u5168\u90e8\u30678\u3064\u306e\u8ad6\u6587\u3078\u306e\u53c2\u7167\u304c\u3042\u308a\u307e\u3059.<br \/>\n\u3059\u3079\u3066\u304c\u6a5f\u68b0\u5b66\u7fd2\u306b\u95a2\u3059\u308b\u3082\u306e\u3067, \u6709\u540d\u306a\u8ad6\u6587\u3082\u3042\u308b\u3088\u3046\u3067\u3059. RNN \u306b\u3064\u3044\u3066\u306e\u8ad6\u6587\u306f\u30cd\u30c3\u30c8\u4e0a\u306b\u3082\u6bd4\u8f03\u7684\u60c5\u5831\u304c\u5c11\u306a\u3044\u3088\u3046\u3067\u3059.\u3053\u308c\u304b\u3089\u4e00\u3064\u4e00\u3064\u773a\u3081\u3066, \u3067\u304d\u308b\u3068\u3053\u308d\u307e\u3067\u6570\u5f0f\u3082\u8ffd\u3063\u3066\u3044\u3053\u3046\u3068\u601d\u3063\u3066\u3044\u307e\u3059.<\/p>\n<h2>\u8ad6\u6587\u306e\u4e00\u89a7<\/h2>\n<p>\u4ee5\u4e0b\u306b\u8ad6\u6587\u306e\u30bf\u30a4\u30c8\u30eb, \u30ea\u30f3\u30af, \u7c21\u5358\u306a\u5185\u5bb9, \u95a2\u9023\u3059\u308b\u30af\u30e9\u30b9\u3092\u305d\u308c\u305e\u308c\u8a18\u8f09\u3057\u307e\u3059.<\/p>\n<h3>1. Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift<\/h3>\n<p><a href=\"https:\/\/arxiv.org\/pdf\/1502.03167v3.pdf\">https:\/\/arxiv.org\/pdf\/1502.03167v3.pdf<\/a><\/p>\n<p>Batch Normalization \u3068\u3044\u3046\u624b\u6cd5\u3092\u7528\u3044\u305f, \u30cb\u30e5\u30fc\u30e9\u30eb\u30cd\u30c3\u30c8\u30ef\u30fc\u30af\u306e\u5b66\u7fd2\u306e\u9ad8\u901f\u5316.<\/p>\n<pre><code>class MPSCNNBatchNormalizationStatistics  \/\/ @available(iOS 11.3, *)\r\n<\/code><\/pre>\n<h3>2. Fast Guided Filter<\/h3>\n<p><a href=\"https:\/\/arxiv.org\/pdf\/1505.00996.pdf\">https:\/\/arxiv.org\/pdf\/1505.00996.pdf<\/a><\/p>\n<p>\u30a8\u30c3\u30b8\u306e\u60c5\u5831\u3092\u4fdd\u3063\u305f\u307e\u307e\u753b\u50cf\u306e\u5e73\u6ed1\u5316\u3092\u884c\u3046 Guided Filter \u3068\u3044\u3046\u624b\u6cd5\u304c\u3042\u308b. \u3053\u306e\u8a08\u7b97\u309210\u500d\u304f\u3089\u3044\u9ad8\u901f\u306b\u3057\u3066\u7d50\u679c\u3082\u76ee\u8996\u3067\u306f\u52a3\u5316\u304c\u5206\u304b\u3089\u306a\u3044\u3088\u3046\u306a\u65b9\u6cd5\u304c\u3042\u308b\u304c, \u610f\u5916\u306b\u3084\u3089\u308c\u3066\u306a\u3044. \u30b3\u30fc\u30c9\u3082\u516c\u958b\u3057\u305f.<\/p>\n<pre><code>class MPSImageGuidedFilter \/\/ @available(iOS 11.3, *)<\/code><\/pre>\n<h3>3. Delving Deep into Rectifiers: Surpassing Human-Level Performance on ImageNet Classification<\/h3>\n<p><a href=\"https:\/\/arxiv.org\/pdf\/1502.01852.pdf\">https:\/\/arxiv.org\/pdf\/1502.01852.pdf<\/a><\/p>\n<p>\u6b63\u898f\u5316\u7dda\u5f62\u95a2\u6570(ReLU)\u306f\u6700\u65b0\u306e\u30cb\u30e5\u30fc\u30e9\u30eb\u30cd\u30c3\u30c8\u30ef\u30fc\u30af\u3067\u306f\u91cd\u8981\u306a\u8981\u7d20\u3067, \u8ffd\u52a0\u306e\u8a08\u7b97\u30b3\u30b9\u30c8\u304c\u307b\u307c\u306a\u304f\u30aa\u30fc\u30d0\u30fc\u30d5\u30a3\u30c3\u30c6\u30a3\u30f3\u30b0\u306e\u30ea\u30b9\u30af\u3092\u6e1b\u3089\u3059 Parametric Rectified Linear Unit(PReLU), \u30ed\u30d0\u30b9\u30c8\u306a\u521d\u671f\u5316\u306e\u65b9\u6cd5\u306b\u3064\u3044\u3066\u7814\u7a76\u3057\u305f.<\/p>\n<pre><code>class MPSNNNeuronDescriptor\r\nclass MPSCNNNeuronPReLU\r\nclass MPSCNNConvolutionDescriptor\r\nclass MPSMatrixNeuron\r\nclass MPSMatrixNeuronGradient \/\/ @available(iOS 12.0, *)<\/code><\/pre>\n<h3>4. MULTI-SCALE CONTEXT AGGREGATION BY DILATED CONVOLUTIONS<\/h3>\n<p><a href=\"https:\/\/arxiv.org\/pdf\/1511.07122v3.pdf\">https:\/\/arxiv.org\/pdf\/1511.07122v3.pdf<\/a><\/p>\n<p>Dilated Convolutions \u3092\u7528\u3044\u305f, \u89e3\u50cf\u5ea6\u3092\u4e0b\u3052\u305f\u308a\u753b\u50cf\u306e\u518d\u30b9\u30b1\u30fc\u30ea\u30f3\u30b0\u3092\u3059\u308b\u5fc5\u8981\u306e\u306a\u3044, Dense Prediction \u306e\u70ba\u306e\u30cd\u30c3\u30c8\u30ef\u30fc\u30af\u3092\u518d\u8003\u3059\u308b.<\/p>\n<pre><code>class MPSCNNConvolutionDescriptor<\/code><\/pre>\n<h3>5. YOLO9000: Better, Faster, Stronger<\/h3>\n<p><a href=\"https:\/\/arxiv.org\/abs\/1612.08242\">https:\/\/arxiv.org\/abs\/1612.08242<\/a><\/p>\n<p>9000\u7a2e\u985e\u3092\u8d85\u3048\u308b\u30aa\u30d6\u30b8\u30a7\u30af\u30c8\u306e\u691c\u51fa\u304c\u53ef\u80fd\u306a YOLO9000 \u306e\u7d39\u4ecb.<\/p>\n<pre><code>class MPSCNNYOLOLossDescriptor \/\/ @available(iOS 12.0, *)<\/code><\/pre>\n<h3>6. Faster Training of Very Deep Networks Via p-Norm Gates<\/h3>\n<p><a href=\"https:\/\/arxiv.org\/abs\/1608.03639\">https:\/\/arxiv.org\/abs\/1608.03639<\/a><\/p>\n<p>GRU, Highway Networks,  Residual Nets \u306a\u3069\u306e\u30cd\u30c3\u30c8\u30ef\u30fc\u30af\u306b\u5bfe\u3057\u3066, \u4f59\u5206\u306a\u5b66\u7fd2\u30b3\u30b9\u30c8\u3092\u6255\u3046\u3053\u3068\u306a\u304f\u5b66\u7fd2\u901f\u5ea6\u3092\u98db\u8e8d\u7684\u306b\u5411\u4e0a\u3055\u305b\u308b p-Norm Gating \u3092\u63d0\u6848\u3059\u308b.<\/p>\n<pre><code>class MPSGRUDescriptor<\/code><\/pre>\n<h3>7. Minimal Gated Unit for Recurrent Neural Networks<\/h3>\n<p><a href=\"https:\/\/arxiv.org\/abs\/1603.09420\">https:\/\/arxiv.org\/abs\/1603.09420<\/a><\/p>\n<p>LSTM \u3084 GRU \u3068\u3044\u3063\u305f\u8907\u96d1\u3067\u7af6\u5408\u3057\u305f\u96a0\u308c\u5c64\u304c\u305f\u304f\u3055\u3093\u3042\u308b\u3068, RNN \u3092\u7406\u89e3\u3057\u305f\u308a\u6700\u9069\u89e3\u3092\u898b\u3064\u3051\u308b\u306e\u304c\u96e3\u3057\u3044. RNN \u306e\u30df\u30cb\u30de\u30eb\u30c7\u30b6\u30a4\u30f3\u3067\u3042\u308b Minimal Gated Unit (MGU) \u3092\u63d0\u6848\u3057, \u7406\u8ad6\u7684\u30fb\u7d4c\u9a13\u7684\u306a\u7814\u7a76\u3092\u3057\u305f.<\/p>\n<pre><code>class MPSGRUDescriptor<\/code><\/pre>\n<h3>8. Adam: A Method for Stochastic Optimization<\/h3>\n<p><a href=\"https:\/\/arxiv.org\/abs\/1412.6980\">https:\/\/arxiv.org\/abs\/1412.6980<\/a><\/p>\n<p>\u78ba\u7387\u7684\u306a\u76ee\u7684\u95a2\u6570\u306e1\u6b21\u306e\u52fe\u914d\u306b\u57fa\u3065\u304f\u6700\u9069\u5316\u30a2\u30eb\u30b4\u30ea\u30ba\u30e0 Adam \u306e\u63d0\u6848.<\/p>\n<pre><code>class MPSNNOptimizerAdam \/\/ @available(iOS 12.0, *)<\/code><\/pre>\n","protected":false},"excerpt":{"rendered":"<p>Metal Performance Shaders \u306b\u306f\u30cb\u30e5\u30fc\u30e9\u30eb\u30cd\u30c3\u30c8\u30ef\u30fc\u30af\u306b\u95a2\u3059\u308b\u5b9f\u88c5\u304c\u542b\u307e\u308c\u3066\u3044\u307e\u3059. \u3042\u308b\u95a2\u6570\u306e\u30b3\u30fc\u30c9\u30c9\u30ad\u30e5\u30e1\u30f3\u30c8\u3092\u773a\u3081\u3066\u3044\u305f\u3068\u3053\u308d, \u30c9\u30ad\u30e5\u30e1\u30f3\u30c8\u5185\u306b\u8ad6\u6587\u3078\u306e\u53c2\u7167\u304c\u3042\u308b\u3053\u3068\u306b\u6c17\u3065\u304d\u307e\u3057\u305f.<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":[],"categories":[5,3],"tags":[],"_links":{"self":[{"href":"https:\/\/naru.jpn.com\/wordpress\/index.php?rest_route=\/wp\/v2\/posts\/265"}],"collection":[{"href":"https:\/\/naru.jpn.com\/wordpress\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/naru.jpn.com\/wordpress\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/naru.jpn.com\/wordpress\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/naru.jpn.com\/wordpress\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=265"}],"version-history":[{"count":6,"href":"https:\/\/naru.jpn.com\/wordpress\/index.php?rest_route=\/wp\/v2\/posts\/265\/revisions"}],"predecessor-version":[{"id":272,"href":"https:\/\/naru.jpn.com\/wordpress\/index.php?rest_route=\/wp\/v2\/posts\/265\/revisions\/272"}],"wp:attachment":[{"href":"https:\/\/naru.jpn.com\/wordpress\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=265"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/naru.jpn.com\/wordpress\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=265"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/naru.jpn.com\/wordpress\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=265"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}