{"id":485,"date":"2018-08-05T05:01:06","date_gmt":"2018-08-04T20:01:06","guid":{"rendered":"http:\/\/light-of-moe.ddo.jp\/~sakura\/diary\/?p=485"},"modified":"2022-08-11T22:16:44","modified_gmt":"2022-08-11T13:16:44","slug":"keras%e3%81%ae%e3%83%a2%e3%83%87%e3%83%ab%e3%81%ae%e5%ae%9a%e7%be%a9%e6%96%b9%e6%b3%95%e3%83%a1%e3%83%a2","status":"publish","type":"post","link":"https:\/\/light-of-moe.ddo.jp\/~sakura\/diary\/?p=485","title":{"rendered":"keras\u306e\u30e2\u30c7\u30eb\u306e\u5b9a\u7fa9\u65b9\u6cd5\u30e1\u30e2"},"content":{"rendered":"\n<p>\u6700\u8fd1\u3001keras\u3067\u904a\u3076\u6a5f\u4f1a\u304c\u5897\u3048\u3066\u304d\u305f\u3002\u30e2\u30c7\u30eb\u306e\u4f5c\u308a\u65b9\u304c\u8907\u6570\u3042\u3063\u3066\u3061\u3087\u3063\u3068\u7d1b\u3089\u308f\u3057\u3044\u306e\u3067\u30e1\u30e2\u304c\u3066\u3089\u306b\u6574\u7406\u3059\u308b\u3002<br\/>\u306a\u304a\u3001\u81ea\u5206\u306ftensorflow\u306e\u4e0a\u306ekeras\u3092\u5229\u7528\u3057\u3066\u3044\u308b\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u30e2\u30c7\u30eb\u306e\u5b9a\u7fa9\u65b9\u6cd5\u306e\u7a2e\u985e\u3068\u5236\u9650<\/h2>\n\n\n\n<p>keras\u306b\u306f\u4e3b\u306b3\u7a2e\u985e\u306e\u30e2\u30c7\u30eb\u306e\u5b9a\u7fa9\u65b9\u6cd5\u304c\u3042\u308b\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\"><li>Sequence\u3092\u5229\u7528\u3059\u308b\u3082\u306e<\/li><li>Functional API\u3092\u5229\u7528\u3059\u308b\u3082\u306e<\/li><li>Model\u30af\u30e9\u30b9\u3092\u7d99\u627f\u3059\u308b\u3082\u306e<\/li><\/ul>\n\n\n\n<h3 class=\"wp-block-heading\">Model\u3092\u7d99\u627f\u3059\u308b\u5834\u5408\u306e\u6ce8\u610f\u70b9<\/h3>\n\n\n\n<p>Model\u3092\u7d99\u627f\u3059\u308b\u5834\u5408\u306b\u306f\u4ee5\u4e0b\u306e\u6ce8\u610f\u304c\u5fc5\u8981\u3067\u3042\u308b\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\"><li> model.save\u304c\u52d5\u4f5c\u3057\u306a\u3044\u3001save_weights\u306f\u52d5\u4f5c\u3059\u308b\u306e\u3067\u305d\u3061\u3089\u3092\u4f7f\u3046 <\/li><\/ul>\n\n\n\n<h3 class=\"wp-block-heading\">eager_execution\u3068\u7d44\u307f\u5408\u308f\u305b\u308b\u3068\u304d\u306e\u6ce8\u610f\u70b9<\/h3>\n\n\n\n<p>eager_execution\u3092\u5229\u7528\u3059\u308b\u3068\u304d\u306b\u4ee5\u4e0b\u306e\u5236\u9650\u304c\u3042\u308b\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\"><li> optimizer\u304ckeras\u5074\u306e\u3082\u306e\u306f\u52d5\u4f5c\u3057\u306a\u3044\u306e\u3067tensorflow\u5074\u306eoptimizer\u3092\u5229\u7528\u3059\u308b<\/li><li>tensorboard\u306ecallback\u304c\u4e0a\u624b\u304f\u52d5\u4f5c\u3057\u306a\u3044\u306e\u3067\u81ea\u524d\u3067\u3084\u308b\u5fc5\u8981\u304c\u3042\u308b <\/li><\/ul>\n\n\n\n<h2 class=\"wp-block-heading\">Sequence\u3092\u5229\u7528\u3059\u308b\u30e2\u30c7\u30eb\u5b9a\u7fa9<\/h2>\n\n\n\n<p>\u7c21\u5358\u306a\u3082\u306e\u3060\u3068\u4ee5\u4e0b\u306e\u69d8\u306aSequence\u3092\u4f7f\u3063\u305f\u5b9a\u7fa9\u304c\u3067\u304d\u308b\u3002<br\/>\u3068\u308a\u3042\u3048\u305a\u306e\u52d5\u4f5c\u306e\u78ba\u8a8d\u306b\u306f\u4fbf\u5229\u3060\u304c\u3001\u5206\u5c90\u3059\u308b\u30cd\u30c3\u30c8\u30ef\u30fc\u30af\u306a\u3069\u306e\u5b9a\u7fa9\u306f\u51fa\u6765\u306a\u3044\u3002<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>from keras.models import Sequential\nfrom keras.layers import Dense, Activation\n\nmodel = Sequential([\n    Dense(32, input_shape=(784,)),\n    Activation('relu'),\n    Dense(10),\n    Activation('softmax'),\n])<\/code><\/pre>\n\n\n\n<h2 class=\"wp-block-heading\">Functional API\u3092\u5229\u7528\u3059\u308b\u30e2\u30c7\u30eb\u5b9a\u7fa9<\/h2>\n\n\n\n<p>Functional API\u3092\u5229\u7528\u3057\u3066\u4ee5\u4e0b\u306e\u3088\u3046\u306b\u3082\u5b9a\u7fa9\u3067\u304d\u308b\u3002<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>from keras.layers import Input, Dense\nfrom keras.models import Model\n\n# This returns a tensor\ninputs = Input(shape=(784,))\n\n# a layer instance is callable on a tensor, and returns a tensor\nx = Dense(64, activation='relu')(inputs)\nx = Dense(64, activation='relu')(x)\npredictions = Dense(10, activation='softmax')(x)\n\n# This creates a model that includes\n# the Input layer and three Dense layers\nmodel = Model(inputs=inputs, outputs=predictions)\nmodel.compile(optimizer='rmsprop',\n              loss='categorical_crossentropy',\n              metrics=['accuracy'])\nmodel.fit(data, labels)  # starts training<\/code><\/pre>\n\n\n\n<h2 class=\"wp-block-heading\">Model\u30af\u30e9\u30b9\u3092\u7d99\u627f\u3059\u308b\u5b9a\u7fa9<\/h2>\n\n\n\n<p>Model\u30af\u30e9\u30b9\u3092\u7d99\u627f\u3057\u3066\u72ec\u81ea\u306e\u30e2\u30c7\u30eb\u30af\u30e9\u30b9\u3092\u4f5c\u6210\u3059\u308b\u3053\u3068\u3082\u3067\u304d\u308b\u3002<br\/>eager_execution\u3068\u7d44\u307f\u5408\u308f\u305b\u308b\u3068\u5f37\u529b\u3002<\/p>\n\n\n\n<p>keras\u958b\u767a\u8005\u306e<a href=\"https:\/\/twitter.com\/fchollet\">Fran\u00e7ois Chollet<\/a>\u00a0\u3055\u3093\u306e\u4ee5\u4e0b\u306etweet\u306e\u3088\u3046\u306b\u3001\u7c21\u6f54\u306bRNN\u306a\u3069\u306e\u8a18\u8ff0\u304c\u3067\u304d\u3066\u3001\u30c7\u30d0\u30c3\u30b0\u306a\u3069\u3082call\u306b\u9069\u5f53\u306bbreak\u5165\u308c\u305f\u308a\u3001print\u3057\u3066\u30c7\u30d0\u30c3\u30b0\u3059\u308b\u3053\u3068\u3082\u51fa\u6765\u307e\u3059\u3002<\/p>\n\n\n\n<figure class=\"wp-block-embed-twitter wp-block-embed is-type-rich is-provider-twitter\">\n<blockquote class=\"twitter-tweet\" data-width=\"550\" data-dnt=\"true\"><p lang=\"en\" dir=\"ltr\">As of TensorFlow 1.7, with eager execution, you can use a Chainer-like API in tf.keras to create models with dynamic control flow. This should be especially useful for NLP researchers working with recursive models! <a href=\"https:\/\/t.co\/ZfAs3u0HJ0\">pic.twitter.com\/ZfAs3u0HJ0<\/a><\/p>&mdash; Fran\u00e7ois Chollet (@fchollet) <a href=\"https:\/\/twitter.com\/fchollet\/status\/979893207965171712?ref_src=twsrc%5Etfw\">March 31, 2018<\/a><\/blockquote><script async src=\"https:\/\/platform.twitter.com\/widgets.js\" charset=\"utf-8\"><\/script>\n<\/figure>\n\n\n\n<p>\u3082\u3063\u3068\u5358\u7d14\u306a\u4f8b\u3060\u3068\u3001\u4ee5\u4e0b\u306e\u3088\u3046\u306a\u611f\u3058\u3067\u3059\u3002<br\/><\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>import keras\n\nclass SimpleMLP(keras.Model):\n\n    def __init__(self, use_bn=False, use_dp=False, num_classes=10):\n        super(SimpleMLP, self).__init__(name='mlp')\n        self.use_bn = use_bn\n        self.use_dp = use_dp\n        self.num_classes = num_classes\n\n        self.dense1 = keras.layers.Dense(32, activation='relu')\n        self.dense2 = keras.layers.Dense(num_classes, activation='softmax')\n        if self.use_dp:\n            self.dp = keras.layers.Dropout(0.5)\n        if self.use_bn:\n            self.bn = keras.layers.BatchNormalization(axis=-1)\n\n    def call(self, inputs):\n        x = self.dense1(inputs)\n        if self.use_dp:\n            x = self.dp(x)\n        if self.use_bn:\n            x = self.bn(x)\n        return self.dense2(x)\n\nmodel = SimpleMLP()\nmodel.compile(...)\nmodel.fit(...)<\/code><\/pre>\n\n\n\n<p>\u6ce8\u610f\u70b9\u3068\u3057\u3066\u3001Model\u30af\u30e9\u30b9\u306f__setattr__\u3068\u304b\u3067\u5404\u30ec\u30a4\u30e4\u306e\u7ba1\u7406\u3092\u884c\u3063\u3066\u304a\u308a\u3001\u76f4\u63a5 self.layer_name = SomeLayer()\u00a0 \u3068\u3044\u3046\u30b3\u30fc\u30c9\u3092\u66f8\u304b\u306a\u3044\u3068optimizer\u304b\u3089\u5909\u6570\u304c\u898b\u3048\u306a\u304f\u3066\u5168\u304f\u5b66\u7fd2\u304c\u3055\u308c\u3066\u3044\u306a\u3044\u306a\u3069\u306e\u7f60\u304c\u3042\u308b\u3002<\/p>\n\n\n\n<p>\u8907\u6570\u306e\u30ec\u30a4\u30e4\u3092for\u6587\u306a\u3069\u3092\u5229\u7528\u3057\u3066\u4e0a\u624b\u304f\u5b9a\u7fa9\u3067\u304d\u305f\u3068\u601d\u3063\u305f\u3089\u4e0a\u624b\u304f\u52d5\u3044\u3066\u3044\u306a\u3044\u3068\u3044\u3046\u3053\u3068\u304c\u8d77\u3053\u308a\u5f97\u308b\u306e\u3067\u6ce8\u610f\u304c\u5fc5\u8981\u3002<\/p>\n\n\n\n<p>\u307e\u305f\u3001\u73fe\u5728\u306e\u5b9f\u88c5\u3060\u3068fit\u306e\u969b\u306b\u30e2\u30c7\u30eb\u304c\u5229\u7528\u3059\u308b\u30e1\u30e2\u30ea\u306e\u8a08\u7b97\u306b\u30d0\u30b0\u304c\u3042\u308b\u3088\u3046\u3067\u3001\u4ee5\u4e0b\u306e\u69d8\u306b\u30b5\u30a4\u30ba\u3092\u660e\u793a\u3057\u3066\u304b\u3089fit\u3057\u306a\u3044\u3068GPU\u306e\u30e1\u30e2\u30ea\u304c\u67af\u6e07\u3057\u3066\u52d5\u4f5c\u3057\u306a\u3044\u3002<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code># batch size\u304c1\u76f8\u5f53\u306e\u30c7\u30fc\u30bf\u3092\u6307\u5b9a\ndummy_x = np.zeros((1, 28, 28, 1))\nmodel._set_inputs(dummy_x)\n<\/code><\/pre>\n\n\n\n<p>\u95a2\u9023\u60c5\u5831:\u00a0<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/19241\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/19241<\/a><br\/><\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u307e\u3068\u3081<\/h2>\n\n\n\n<p>\u3068\u308a\u3042\u3048\u305a\u8a18\u6cd5\u7684\u306b\u306fModel\u3092\u7d99\u627f\u3057\u3066eager_execution\u3092\u5229\u7528\u3057\u3066\u3044\u308b\u306e\u304c\u5e78\u305b\u305d\u3046\u3002\u305f\u3060\u3057\u3001\u5192\u982d\u306b\u66f8\u3044\u305f\u5236\u9650\u3084\u30d0\u30b0\u3082\u3061\u3089\u307b\u3089\u3042\u308b\u3088\u3046\u306a\u306e\u3067\u6ce8\u610f\u306f\u5fc5\u8981\u3002<\/p>\n\n\n\n<p><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u6700\u8fd1\u3001keras\u3067\u904a\u3076\u6a5f\u4f1a\u304c\u5897\u3048\u3066\u304d\u305f\u3002\u30e2\u30c7\u30eb\u306e\u4f5c\u308a\u65b9\u304c\u8907\u6570\u3042\u3063\u3066\u3061\u3087\u3063\u3068\u7d1b\u3089\u308f\u3057\u3044\u306e\u3067\u30e1\u30e2\u304c\u3066\u3089\u306b\u6574\u7406\u3059\u308b\u3002\u306a\u304a\u3001\u81ea\u5206\u306ftensorflow\u306e\u4e0a\u306ekeras\u3092\u5229\u7528\u3057\u3066\u3044\u308b\u3002 \u30e2\u30c7\u30eb\u306e\u5b9a\u7fa9\u65b9\u6cd5\u306e\u7a2e\u985e\u3068\u5236\u9650 keras\u306b\u306f\u2026 <span class=\"read-more\"><a href=\"https:\/\/light-of-moe.ddo.jp\/~sakura\/diary\/?p=485\">\u7d9a\u304d\u3092\u8aad\u3080 &raquo;<\/a><\/span><\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"jetpack_post_was_ever_published":false,"_jetpack_newsletter_access":"","_jetpack_dont_email_post_to_subs":false,"_jetpack_newsletter_tier_id":0,"_jetpack_memberships_contains_paywalled_content":false,"_jetpack_memberships_contains_paid_content":false,"footnotes":"","jetpack_publicize_message":"","jetpack_publicize_feature_enabled":true,"jetpack_social_post_already_shared":true,"jetpack_social_options":{"image_generator_settings":{"template":"highway","default_image_id":0,"font":"","enabled":false},"version":2}},"categories":[2,7],"tags":[],"class_list":["post-485","post","type-post","status-publish","format-standard","hentry","category-tensorflow","category-7"],"jetpack_publicize_connections":[],"jetpack_featured_media_url":"","jetpack_sharing_enabled":true,"jetpack_shortlink":"https:\/\/wp.me\/p4wwqG-7P","_links":{"self":[{"href":"https:\/\/light-of-moe.ddo.jp\/~sakura\/diary\/index.php?rest_route=\/wp\/v2\/posts\/485","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/light-of-moe.ddo.jp\/~sakura\/diary\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/light-of-moe.ddo.jp\/~sakura\/diary\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/light-of-moe.ddo.jp\/~sakura\/diary\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/light-of-moe.ddo.jp\/~sakura\/diary\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=485"}],"version-history":[{"count":23,"href":"https:\/\/light-of-moe.ddo.jp\/~sakura\/diary\/index.php?rest_route=\/wp\/v2\/posts\/485\/revisions"}],"predecessor-version":[{"id":1033,"href":"https:\/\/light-of-moe.ddo.jp\/~sakura\/diary\/index.php?rest_route=\/wp\/v2\/posts\/485\/revisions\/1033"}],"wp:attachment":[{"href":"https:\/\/light-of-moe.ddo.jp\/~sakura\/diary\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=485"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/light-of-moe.ddo.jp\/~sakura\/diary\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=485"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/light-of-moe.ddo.jp\/~sakura\/diary\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=485"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}