05_転移学習.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# 転移学習とファインチューニング\n",
    "\n",
    "kerasは、すでに学習が済んでいる大規模なニューラルネットワークモデルを流用することができます。\n",
    "\n",
    "\n",
    "+ **fine tuning**：学習済みモデルの重みを初期値にして、重みのみ再学習する\n",
    "+ **転移学習**：学習済みモデルに、層を追加して、その層について学習する"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {
    "collapsed": true
   },
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "Using TensorFlow backend.\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "(50000, 32, 32, 3) (50000, 1) (10000, 32, 32, 3) (10000, 1)\n"
     ]
    },
    {
     "data": {
      "text/plain": [
       "<matplotlib.image.AxesImage at 0x214b9658898>"
      ]
     },
     "execution_count": 1,
     "metadata": {},
     "output_type": "execute_result"
    },
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 432x288 with 1 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    }
   ],
   "source": [
    "import keras\n",
    "from keras.datasets import cifar10\n",
    "(X_train, y_train), (X_test, y_test) = cifar10.load_data()\n",
    "\n",
    "# データの形式\n",
    "print(\n",
    "    X_train.shape,\n",
    "    y_train.shape,\n",
    "    X_test.shape,\n",
    "    y_test.shape\n",
    ")\n",
    "\n",
    "import matplotlib.pyplot as plt\n",
    "%matplotlib inline\n",
    "plt.imshow(X_train[0])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "# ベクトルからone hotに変換する\n",
    "y_train = keras.utils.to_categorical(y_train,10)\n",
    "y_test = keras.utils.to_categorical(y_test,10)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## 転移学習\n",
    "\n",
    "学習済みモデルを前半に配置して、その後ろに全結合層を配置させます。\n",
    "\n",
    "全結合層のみ学習させます。"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {
    "collapsed": true
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Epoch 1/30\n",
      "50000/50000 [==============================] - 8s 163us/step - loss: 2.4685 - accuracy: 0.1043\n",
      "Epoch 2/30\n",
      "50000/50000 [==============================] - 7s 145us/step - loss: 2.3007 - accuracy: 0.1095\n",
      "Epoch 3/30\n",
      "50000/50000 [==============================] - 7s 145us/step - loss: 2.2907 - accuracy: 0.1328\n",
      "Epoch 4/30\n",
      "50000/50000 [==============================] - 7s 144us/step - loss: 2.0447 - accuracy: 0.2739\n",
      "Epoch 5/30\n",
      "50000/50000 [==============================] - 7s 143us/step - loss: 1.5968 - accuracy: 0.4766\n",
      "Epoch 6/30\n",
      "50000/50000 [==============================] - 7s 149us/step - loss: 1.2245 - accuracy: 0.5835\n",
      "Epoch 7/30\n",
      "50000/50000 [==============================] - 8s 151us/step - loss: 1.0099 - accuracy: 0.6600\n",
      "Epoch 8/30\n",
      "50000/50000 [==============================] - 7s 149us/step - loss: 0.8983 - accuracy: 0.7040\n",
      "Epoch 9/30\n",
      "50000/50000 [==============================] - 7s 147us/step - loss: 0.8253 - accuracy: 0.7262\n",
      "Epoch 10/30\n",
      "50000/50000 [==============================] - 7s 143us/step - loss: 0.7630 - accuracy: 0.7441\n",
      "Epoch 11/30\n",
      "50000/50000 [==============================] - 7s 145us/step - loss: 0.7093 - accuracy: 0.7647\n",
      "Epoch 12/30\n",
      "50000/50000 [==============================] - 7s 148us/step - loss: 0.6498 - accuracy: 0.7824\n",
      "Epoch 13/30\n",
      "50000/50000 [==============================] - 7s 148us/step - loss: 0.6228 - accuracy: 0.7918\n",
      "Epoch 14/30\n",
      "50000/50000 [==============================] - 7s 149us/step - loss: 0.5693 - accuracy: 0.8090\n",
      "Epoch 15/30\n",
      "50000/50000 [==============================] - 7s 148us/step - loss: 0.5244 - accuracy: 0.8241\n",
      "Epoch 16/30\n",
      "50000/50000 [==============================] - 7s 149us/step - loss: 0.4934 - accuracy: 0.8335\n",
      "Epoch 17/30\n",
      "50000/50000 [==============================] - 7s 149us/step - loss: 0.4490 - accuracy: 0.8485\n",
      "Epoch 18/30\n",
      "50000/50000 [==============================] - 8s 151us/step - loss: 0.4054 - accuracy: 0.8642\n",
      "Epoch 19/30\n",
      "50000/50000 [==============================] - 8s 150us/step - loss: 0.3729 - accuracy: 0.8748\n",
      "Epoch 20/30\n",
      "50000/50000 [==============================] - 8s 150us/step - loss: 0.3432 - accuracy: 0.8872\n",
      "Epoch 21/30\n",
      "50000/50000 [==============================] - 8s 154us/step - loss: 0.3032 - accuracy: 0.8994\n",
      "Epoch 22/30\n",
      "50000/50000 [==============================] - 8s 150us/step - loss: 0.2722 - accuracy: 0.9104\n",
      "Epoch 23/30\n",
      "50000/50000 [==============================] - 8s 151us/step - loss: 0.2395 - accuracy: 0.9222\n",
      "Epoch 24/30\n",
      "50000/50000 [==============================] - 8s 151us/step - loss: 0.1879 - accuracy: 0.9399\n",
      "Epoch 25/30\n",
      "50000/50000 [==============================] - 8s 150us/step - loss: 0.1614 - accuracy: 0.9481\n",
      "Epoch 26/30\n",
      "50000/50000 [==============================] - 8s 150us/step - loss: 0.1871 - accuracy: 0.9442\n",
      "Epoch 27/30\n",
      "50000/50000 [==============================] - 8s 152us/step - loss: 0.1532 - accuracy: 0.9557\n",
      "Epoch 28/30\n",
      "50000/50000 [==============================] - 8s 152us/step - loss: 0.1523 - accuracy: 0.9563\n",
      "Epoch 29/30\n",
      "50000/50000 [==============================] - 8s 151us/step - loss: 0.0669 - accuracy: 0.9814\n",
      "Epoch 30/30\n",
      "50000/50000 [==============================] - 8s 151us/step - loss: 0.1365 - accuracy: 0.9605\n"
     ]
    }
   ],
   "source": [
    "#  VGG16学習済みモデルを読み込む\n",
    "from keras.applications.vgg16 import VGG16\n",
    "# ニューラルネット関係のライブラリ読み込み\n",
    "from keras.models import Sequential\n",
    "from keras.layers import Dense, Dropout, Flatten\n",
    "from keras import optimizers\n",
    "from keras import losses\n",
    "from keras import models\n",
    "from keras import layers\n",
    "\n",
    "# VGG16モデル\n",
    "img_rows = 32\n",
    "img_cols = 32\n",
    "input_tensor = layers.Input(shape=(img_rows, img_cols, 3))\n",
    "vgg16 = VGG16(\n",
    "    include_top=False,      # ネットワークの出力層側にある3つの全結合層を含むかどうか\n",
    "    weights='imagenet',      # None(ランダム初期化) 'imagenet'(ImageNetで学習した重み) \n",
    "    input_tensor=input_tensor# 入力層の形式\n",
    ")\n",
    "\n",
    "# 全結合層は自己学習モデル\n",
    "model_ = Sequential()\n",
    "model_.add(Flatten(input_shape=vgg16.output_shape[1:]))\n",
    "model_.add(Dense(256, activation='relu'))\n",
    "model_.add(Dropout(0.1))\n",
    "model_.add(Dense(256, activation='relu'))\n",
    "model_.add(Dropout(0.1))\n",
    "model_.add(Dense(256, activation='relu'))\n",
    "model_.add(Dropout(0.1))\n",
    "model_.add(Dense(10, activation='softmax'))\n",
    "\n",
    "# VGG16モデルと自己学習モデルの結合\n",
    "model = models.Model(\n",
    "    inputs=vgg16.input, \n",
    "    outputs=model_(vgg16.output)\n",
    ")\n",
    "\n",
    "# modelの14層目までのモデル重みは学習させない\n",
    "for layer in model.layers[:15]:\n",
    "    layer.trainable = False\n",
    "\n",
    "# モデルの学習設定\n",
    "model.compile(\n",
    "    loss=losses.categorical_crossentropy,\n",
    "    optimizer=optimizers.SGD(lr=0.01),\n",
    "    metrics=['accuracy']\n",
    ")\n",
    "\n",
    "# モデルの学習\n",
    "result = model.fit(\n",
    "    X_train,\n",
    "    y_train,\n",
    "    batch_size=512,\n",
    "    epochs=30,\n",
    ")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "# モデルの保存\n",
    "model.save('./model/model_cnn.h5')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "# テストデータへの適用\n",
    "import numpy as np\n",
    "y_predict = np.argmax(model.predict(X_test),axis=1)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {
    "collapsed": true
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "0.7585\n"
     ]
    }
   ],
   "source": [
    "# 正答率の算出\n",
    "y_test = np.argmax(y_test,axis=1)\n",
    "acc=np.sum(y_test==y_predict)/len(y_test)\n",
    "print(acc)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# 転移学習を行わなかった場合"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "# データ読み込み\n",
    "from keras.datasets import cifar10\n",
    "(X_train, y_train), (X_test, y_test) = cifar10.load_data()\n",
    "\n",
    "# ベクトルからone hotに変換する\n",
    "y_train = keras.utils.to_categorical(y_train,10)\n",
    "y_test = keras.utils.to_categorical(y_test,10)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "from keras.models import Sequential\n",
    "from keras.layers import Dense, Dropout, Flatten, Conv2D, MaxPooling2D, ReLU, Softmax\n",
    "from keras import optimizers\n",
    "from keras import losses\n",
    "\n",
    "# 基本モデル作成\n",
    "model = Sequential()\n",
    "\n",
    "# 3×3の畳み込み層の追加\n",
    "model.add(\n",
    "    Conv2D(\n",
    "        filters=32,\n",
    "        kernel_size=(3, 3),\n",
    "        activation=ReLU(),\n",
    "    )\n",
    ")\n",
    "\n",
    "# 2×2のmaxプーリング層追加\n",
    "model.add(\n",
    "    MaxPooling2D(\n",
    "        pool_size=(2, 2)\n",
    "    )\n",
    ")\n",
    "\n",
    "# ドロップアウト層の追加\n",
    "model.add(\n",
    "    Dropout(0.2)\n",
    ")\n",
    "\n",
    "# フラット層の追加\n",
    "model.add(\n",
    "    Flatten()\n",
    ")\n",
    "\n",
    "# 全結合層の追加\n",
    "model.add(\n",
    "    Dense(\n",
    "        128, \n",
    "        activation=ReLU()\n",
    "    )\n",
    ")\n",
    "\n",
    "# ドロップアウト層の追加\n",
    "model.add(\n",
    "    Dropout(0.2)\n",
    ")\n",
    "\n",
    "# 出力層\n",
    "model.add(\n",
    "    Dense(\n",
    "        10, \n",
    "        activation=Softmax()\n",
    "    )\n",
    ")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "# モデルの学習設定\n",
    "model.compile(\n",
    "    loss=losses.categorical_crossentropy,\n",
    "    optimizer=optimizers.SGD(lr=0.01),\n",
    "    metrics=['accuracy']\n",
    ")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "metadata": {
    "collapsed": true
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Epoch 1/100\n",
      "50000/50000 [==============================] - 2s 39us/step - loss: 14.5063 - accuracy: 0.1000\n",
      "Epoch 2/100\n",
      "50000/50000 [==============================] - 2s 38us/step - loss: 14.5063 - accuracy: 0.1000\n",
      "Epoch 3/100\n",
      "50000/50000 [==============================] - 2s 38us/step - loss: 14.5063 - accuracy: 0.1000\n",
      "Epoch 4/100\n",
      "50000/50000 [==============================] - 2s 39us/step - loss: 14.5063 - accuracy: 0.1000\n",
      "Epoch 5/100\n",
      "50000/50000 [==============================] - 2s 39us/step - loss: 14.5063 - accuracy: 0.1000\n",
      "Epoch 6/100\n",
      "50000/50000 [==============================] - 2s 39us/step - loss: 14.5063 - accuracy: 0.1000\n",
      "Epoch 7/100\n",
      "50000/50000 [==============================] - 2s 39us/step - loss: 14.5063 - accuracy: 0.1000\n",
      "Epoch 8/100\n",
      "50000/50000 [==============================] - 2s 39us/step - loss: 14.5063 - accuracy: 0.1000: 1s - loss: 1\n",
      "Epoch 9/100\n",
      "50000/50000 [==============================] - 2s 39us/step - loss: 14.5060 - accuracy: 0.1000\n",
      "Epoch 10/100\n",
      "50000/50000 [==============================] - 2s 39us/step - loss: 14.5063 - accuracy: 0.1000\n",
      "Epoch 11/100\n",
      "50000/50000 [==============================] - 2s 39us/step - loss: 14.5063 - accuracy: 0.1000\n",
      "Epoch 12/100\n",
      "50000/50000 [==============================] - 2s 38us/step - loss: 14.5063 - accuracy: 0.1000\n",
      "Epoch 13/100\n",
      "50000/50000 [==============================] - 2s 38us/step - loss: 14.5047 - accuracy: 0.1001\n",
      "Epoch 14/100\n",
      "50000/50000 [==============================] - 2s 39us/step - loss: 14.4958 - accuracy: 0.1006\n",
      "Epoch 15/100\n",
      "50000/50000 [==============================] - 2s 38us/step - loss: 14.4677 - accuracy: 0.1022\n",
      "Epoch 16/100\n",
      "50000/50000 [==============================] - 2s 38us/step - loss: 14.2059 - accuracy: 0.1182\n",
      "Epoch 17/100\n",
      "50000/50000 [==============================] - 2s 38us/step - loss: 14.1089 - accuracy: 0.1242\n",
      "Epoch 18/100\n",
      "50000/50000 [==============================] - 2s 38us/step - loss: 13.7593 - accuracy: 0.1460\n",
      "Epoch 19/100\n",
      "50000/50000 [==============================] - 2s 38us/step - loss: 14.0478 - accuracy: 0.1283\n",
      "Epoch 20/100\n",
      "50000/50000 [==============================] - 2s 39us/step - loss: 13.7063 - accuracy: 0.1494\n",
      "Epoch 21/100\n",
      "50000/50000 [==============================] - 2s 39us/step - loss: 13.6017 - accuracy: 0.1558\n",
      "Epoch 22/100\n",
      "50000/50000 [==============================] - 2s 39us/step - loss: 13.7079 - accuracy: 0.1493\n",
      "Epoch 23/100\n",
      "50000/50000 [==============================] - 2s 39us/step - loss: 13.7043 - accuracy: 0.1494\n",
      "Epoch 24/100\n",
      "50000/50000 [==============================] - 2s 39us/step - loss: 13.6461 - accuracy: 0.1532\n",
      "Epoch 25/100\n",
      "50000/50000 [==============================] - 2s 38us/step - loss: 14.3334 - accuracy: 0.1107\n",
      "Epoch 26/100\n",
      "50000/50000 [==============================] - 2s 37us/step - loss: 14.1788 - accuracy: 0.1202\n",
      "Epoch 27/100\n",
      "50000/50000 [==============================] - 2s 38us/step - loss: 13.7856 - accuracy: 0.1446\n",
      "Epoch 28/100\n",
      "50000/50000 [==============================] - 2s 37us/step - loss: 13.7013 - accuracy: 0.1498\n",
      "Epoch 29/100\n",
      "50000/50000 [==============================] - 2s 37us/step - loss: 13.7600 - accuracy: 0.1461\n",
      "Epoch 30/100\n",
      "50000/50000 [==============================] - 2s 37us/step - loss: 13.6802 - accuracy: 0.1511\n",
      "Epoch 31/100\n",
      "50000/50000 [==============================] - 2s 37us/step - loss: 13.8910 - accuracy: 0.1381\n",
      "Epoch 32/100\n",
      "50000/50000 [==============================] - 2s 38us/step - loss: 13.9241 - accuracy: 0.1360\n",
      "Epoch 33/100\n",
      "50000/50000 [==============================] - 2s 37us/step - loss: 13.8420 - accuracy: 0.1411\n",
      "Epoch 34/100\n",
      "50000/50000 [==============================] - 2s 37us/step - loss: 14.1659 - accuracy: 0.1211\n",
      "Epoch 35/100\n",
      "50000/50000 [==============================] - 2s 38us/step - loss: 13.8402 - accuracy: 0.1413\n",
      "Epoch 36/100\n",
      "50000/50000 [==============================] - 2s 37us/step - loss: 13.9109 - accuracy: 0.1368\n",
      "Epoch 37/100\n",
      "50000/50000 [==============================] - 2s 37us/step - loss: 13.7319 - accuracy: 0.1480\n",
      "Epoch 38/100\n",
      "50000/50000 [==============================] - 2s 38us/step - loss: 13.7588 - accuracy: 0.1461\n",
      "Epoch 39/100\n",
      "50000/50000 [==============================] - 2s 37us/step - loss: 14.2317 - accuracy: 0.1169\n",
      "Epoch 40/100\n",
      "50000/50000 [==============================] - 2s 38us/step - loss: 14.4887 - accuracy: 0.1011\n",
      "Epoch 41/100\n",
      "50000/50000 [==============================] - 2s 37us/step - loss: 14.4872 - accuracy: 0.1011\n",
      "Epoch 42/100\n",
      "50000/50000 [==============================] - 2s 37us/step - loss: 14.3486 - accuracy: 0.1097\n",
      "Epoch 43/100\n",
      "50000/50000 [==============================] - 2s 38us/step - loss: 14.4938 - accuracy: 0.1008\n",
      "Epoch 44/100\n",
      "50000/50000 [==============================] - 2s 38us/step - loss: 14.5005 - accuracy: 0.1004\n",
      "Epoch 45/100\n",
      "50000/50000 [==============================] - 2s 37us/step - loss: 14.4960 - accuracy: 0.1006\n",
      "Epoch 46/100\n",
      "50000/50000 [==============================] - 2s 38us/step - loss: 14.4836 - accuracy: 0.1014\n",
      "Epoch 47/100\n",
      "50000/50000 [==============================] - 2s 37us/step - loss: 14.3598 - accuracy: 0.1090\n",
      "Epoch 48/100\n",
      "50000/50000 [==============================] - 2s 37us/step - loss: 13.8153 - accuracy: 0.1428\n",
      "Epoch 49/100\n",
      "50000/50000 [==============================] - 2s 38us/step - loss: 14.2655 - accuracy: 0.1149\n",
      "Epoch 50/100\n",
      "50000/50000 [==============================] - 2s 37us/step - loss: 14.0091 - accuracy: 0.1308\n",
      "Epoch 51/100\n",
      "50000/50000 [==============================] - 2s 37us/step - loss: 14.3254 - accuracy: 0.1112\n",
      "Epoch 52/100\n",
      "50000/50000 [==============================] - 2s 37us/step - loss: 14.4644 - accuracy: 0.1026\n",
      "Epoch 53/100\n",
      "50000/50000 [==============================] - 2s 37us/step - loss: 14.4631 - accuracy: 0.1027\n",
      "Epoch 54/100\n",
      "50000/50000 [==============================] - 2s 37us/step - loss: 14.4637 - accuracy: 0.1026\n",
      "Epoch 55/100\n",
      "50000/50000 [==============================] - 2s 38us/step - loss: 14.4660 - accuracy: 0.1025\n",
      "Epoch 56/100\n",
      "50000/50000 [==============================] - 2s 38us/step - loss: 14.4705 - accuracy: 0.1022: 0s - loss: 14.4674 - a\n",
      "Epoch 57/100\n",
      "50000/50000 [==============================] - 2s 38us/step - loss: 14.3801 - accuracy: 0.1078\n",
      "Epoch 58/100\n",
      "50000/50000 [==============================] - 2s 37us/step - loss: 14.4402 - accuracy: 0.1041\n",
      "Epoch 59/100\n",
      "50000/50000 [==============================] - 2s 38us/step - loss: 14.4299 - accuracy: 0.1047\n",
      "Epoch 60/100\n",
      "50000/50000 [==============================] - 2s 38us/step - loss: 14.4331 - accuracy: 0.1045\n",
      "Epoch 61/100\n",
      "50000/50000 [==============================] - 2s 38us/step - loss: 14.3978 - accuracy: 0.1067\n",
      "Epoch 62/100\n",
      "50000/50000 [==============================] - 2s 38us/step - loss: 14.3466 - accuracy: 0.1099\n",
      "Epoch 63/100\n",
      "50000/50000 [==============================] - 2s 38us/step - loss: 14.2033 - accuracy: 0.1188\n",
      "Epoch 64/100\n",
      "50000/50000 [==============================] - 2s 37us/step - loss: 14.2550 - accuracy: 0.1156\n",
      "Epoch 65/100\n",
      "50000/50000 [==============================] - 2s 38us/step - loss: 13.7907 - accuracy: 0.1444\n",
      "Epoch 66/100\n",
      "50000/50000 [==============================] - 2s 38us/step - loss: 13.8947 - accuracy: 0.1379\n",
      "Epoch 67/100\n",
      "50000/50000 [==============================] - 2s 38us/step - loss: 14.0180 - accuracy: 0.1303\n",
      "Epoch 68/100\n",
      "50000/50000 [==============================] - 2s 38us/step - loss: 13.8599 - accuracy: 0.1400\n",
      "Epoch 69/100\n",
      "50000/50000 [==============================] - 2s 38us/step - loss: 14.2219 - accuracy: 0.1176\n",
      "Epoch 70/100\n",
      "50000/50000 [==============================] - 2s 38us/step - loss: 14.0685 - accuracy: 0.1271\n",
      "Epoch 71/100\n",
      "50000/50000 [==============================] - 2s 38us/step - loss: 13.7604 - accuracy: 0.1462\n",
      "Epoch 72/100\n",
      "50000/50000 [==============================] - 2s 37us/step - loss: 13.8977 - accuracy: 0.1377\n",
      "Epoch 73/100\n",
      "50000/50000 [==============================] - 2s 37us/step - loss: 13.7004 - accuracy: 0.1500\n",
      "Epoch 74/100\n",
      "50000/50000 [==============================] - 2s 38us/step - loss: 13.9496 - accuracy: 0.1345\n",
      "Epoch 75/100\n",
      "50000/50000 [==============================] - 2s 37us/step - loss: 14.3599 - accuracy: 0.1091\n",
      "Epoch 76/100\n",
      "50000/50000 [==============================] - 2s 38us/step - loss: 14.1182 - accuracy: 0.1241\n",
      "Epoch 77/100\n",
      "50000/50000 [==============================] - 2s 38us/step - loss: 13.8943 - accuracy: 0.1379\n",
      "Epoch 78/100\n",
      "50000/50000 [==============================] - 2s 38us/step - loss: 14.1600 - accuracy: 0.1215\n",
      "Epoch 79/100\n",
      "50000/50000 [==============================] - 2s 38us/step - loss: 13.6464 - accuracy: 0.1532\n",
      "Epoch 80/100\n",
      "50000/50000 [==============================] - 2s 38us/step - loss: 13.5471 - accuracy: 0.1593\n",
      "Epoch 81/100\n",
      "50000/50000 [==============================] - 2s 37us/step - loss: 13.4399 - accuracy: 0.1660\n",
      "Epoch 82/100\n",
      "50000/50000 [==============================] - 2s 38us/step - loss: 13.7300 - accuracy: 0.1481\n",
      "Epoch 83/100\n",
      "50000/50000 [==============================] - 2s 38us/step - loss: 13.5933 - accuracy: 0.1565\n",
      "Epoch 84/100\n",
      "50000/50000 [==============================] - 2s 38us/step - loss: 13.8726 - accuracy: 0.1392\n",
      "Epoch 85/100\n",
      "50000/50000 [==============================] - 2s 38us/step - loss: 13.4759 - accuracy: 0.1638\n",
      "Epoch 86/100\n",
      "50000/50000 [==============================] - 2s 38us/step - loss: 13.5842 - accuracy: 0.1571\n",
      "Epoch 87/100\n",
      "50000/50000 [==============================] - 2s 38us/step - loss: 13.7459 - accuracy: 0.1472\n",
      "Epoch 88/100\n",
      "50000/50000 [==============================] - 2s 38us/step - loss: 14.0742 - accuracy: 0.1268\n",
      "Epoch 89/100\n",
      "50000/50000 [==============================] - 2s 37us/step - loss: 14.4160 - accuracy: 0.1056\n",
      "Epoch 90/100\n",
      "50000/50000 [==============================] - 2s 38us/step - loss: 14.4054 - accuracy: 0.1063\n",
      "Epoch 91/100\n",
      "50000/50000 [==============================] - 2s 38us/step - loss: 14.4238 - accuracy: 0.1051\n",
      "Epoch 92/100\n",
      "50000/50000 [==============================] - 2s 38us/step - loss: 14.4405 - accuracy: 0.1041\n",
      "Epoch 93/100\n",
      "50000/50000 [==============================] - 2s 38us/step - loss: 14.4445 - accuracy: 0.1038\n",
      "Epoch 94/100\n",
      "50000/50000 [==============================] - 2s 38us/step - loss: 14.4570 - accuracy: 0.1031\n",
      "Epoch 95/100\n",
      "50000/50000 [==============================] - 2s 38us/step - loss: 14.4521 - accuracy: 0.1034\n",
      "Epoch 96/100\n",
      "50000/50000 [==============================] - 2s 38us/step - loss: 14.4476 - accuracy: 0.1036\n",
      "Epoch 97/100\n",
      "50000/50000 [==============================] - ETA: 0s - loss: 14.4480 - accuracy: 0.103 - 2s 38us/step - loss: 14.4505 - accuracy: 0.1035\n",
      "Epoch 98/100\n",
      "50000/50000 [==============================] - 2s 37us/step - loss: 14.2254 - accuracy: 0.1174\n",
      "Epoch 99/100\n",
      "50000/50000 [==============================] - 2s 38us/step - loss: 14.0166 - accuracy: 0.1304\n",
      "Epoch 100/100\n",
      "50000/50000 [==============================] - 2s 38us/step - loss: 14.0937 - accuracy: 0.1256\n"
     ]
    }
   ],
   "source": [
    "# モデルの学習\n",
    "result = model.fit(\n",
    "    X_train,\n",
    "    y_train,\n",
    "    batch_size=512,\n",
    "    epochs=100,\n",
    ")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "metadata": {
    "collapsed": true
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "[14.10816884765625, 0.12470000237226486]\n"
     ]
    }
   ],
   "source": [
    "# モデル結果の出力\n",
    "score = model.evaluate(X_test, y_test, verbose=0)\n",
    "print(score)"
   ]
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 3",
   "language": "python",
   "name": "python3"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 3
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython3",
   "version": "3.6.8"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}