02_keras応用.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "**この講座では、Kerasの応用的な使い方をご紹介いたします**\n",
    "\n",
    "## kerasには下記のような機能があります。\n",
    "\n",
    "## ①学習の可視化\n",
    "## ②モデルの保存と読み込み\n",
    "## ③コールバック\n",
    "### 　　　1.学習過程の可視化\n",
    "### 　　　2.学習の早期終了\n",
    "## ④最適なモデルの選択"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## データ生成"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 37,
   "metadata": {},
   "outputs": [],
   "source": [
    "# 連続値データの読み込み\n",
    "from sklearn.datasets import load_iris\n",
    "iris = load_iris()\n",
    "\n",
    "# 訓練データとテストデータに分ける\n",
    "from sklearn.model_selection import train_test_split\n",
    "X_train, X_test, y_train, y_test = train_test_split(\n",
    "    iris['data'], \n",
    "    iris['target'], \n",
    "    test_size=0.3,  \n",
    "    random_state=0\n",
    ")\n",
    "\n",
    "# 正規化（Normarization）\n",
    "from sklearn.preprocessing import MinMaxScaler\n",
    "scaler_x = MinMaxScaler()\n",
    "X_train_n = scaler_x.fit_transform(X_train)\n",
    "X_test_n = scaler_x.transform(X_test) \n",
    "\n",
    "# one-hotベクトル化\n",
    "import keras\n",
    "y_train = keras.utils.to_categorical(y_train,3)\n",
    "y_test = keras.utils.to_categorical(y_test,3)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# ①学習の可視化"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 38,
   "metadata": {},
   "outputs": [],
   "source": [
    "# 必要ライブラリのインポート\n",
    "import keras\n",
    "from keras.models import Sequential           # 基本モデルのクラス\n",
    "from keras.layers import Dense                # 層\n",
    "from keras.layers import Dropout              # ドロップアウト\n",
    "from keras.layers import core                 # 活性化関数\n",
    "from keras.layers import normalization        # バッチノーマリゼーション\n",
    "from keras import losses                      # 損失関数\n",
    "from keras import optimizers                  # 重み更新法\n",
    "from keras import callbacks                   # コールバック（1epoch終了毎に呼び出される）\n",
    "\n",
    "import matplotlib.pyplot as plt               # 描画\n",
    "%matplotlib inline"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 39,
   "metadata": {
    "collapsed": true
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Epoch 1/100\n",
      "105/105 [==============================] - 1s 5ms/step - loss: 1.7083 - acc: 0.4476\n",
      "Epoch 2/100\n",
      "105/105 [==============================] - 0s 181us/step - loss: 1.1833 - acc: 0.5619\n",
      "Epoch 3/100\n",
      "105/105 [==============================] - 0s 190us/step - loss: 1.0857 - acc: 0.5619\n",
      "Epoch 4/100\n",
      "105/105 [==============================] - 0s 181us/step - loss: 1.1077 - acc: 0.5714\n",
      "Epoch 5/100\n",
      "105/105 [==============================] - 0s 181us/step - loss: 0.8115 - acc: 0.6381\n",
      "Epoch 6/100\n",
      "105/105 [==============================] - 0s 190us/step - loss: 0.9418 - acc: 0.5333\n",
      "Epoch 7/100\n",
      "105/105 [==============================] - 0s 190us/step - loss: 0.9060 - acc: 0.5238\n",
      "Epoch 8/100\n",
      "105/105 [==============================] - 0s 181us/step - loss: 0.6754 - acc: 0.6857\n",
      "Epoch 9/100\n",
      "105/105 [==============================] - 0s 181us/step - loss: 0.6928 - acc: 0.5619\n",
      "Epoch 10/100\n",
      "105/105 [==============================] - 0s 190us/step - loss: 0.6151 - acc: 0.6571\n",
      "Epoch 11/100\n",
      "105/105 [==============================] - 0s 190us/step - loss: 0.6222 - acc: 0.6762\n",
      "Epoch 12/100\n",
      "105/105 [==============================] - 0s 181us/step - loss: 0.5549 - acc: 0.6952\n",
      "Epoch 13/100\n",
      "105/105 [==============================] - 0s 181us/step - loss: 0.5920 - acc: 0.7143\n",
      "Epoch 14/100\n",
      "105/105 [==============================] - 0s 200us/step - loss: 0.5182 - acc: 0.6762\n",
      "Epoch 15/100\n",
      "105/105 [==============================] - 0s 171us/step - loss: 0.5142 - acc: 0.6476\n",
      "Epoch 16/100\n",
      "105/105 [==============================] - 0s 181us/step - loss: 0.5157 - acc: 0.6762\n",
      "Epoch 17/100\n",
      "105/105 [==============================] - 0s 190us/step - loss: 0.5181 - acc: 0.6571\n",
      "Epoch 18/100\n",
      "105/105 [==============================] - 0s 190us/step - loss: 0.4939 - acc: 0.7143\n",
      "Epoch 19/100\n",
      "105/105 [==============================] - 0s 181us/step - loss: 0.3711 - acc: 0.6952\n",
      "Epoch 20/100\n",
      "105/105 [==============================] - 0s 181us/step - loss: 0.3427 - acc: 0.6857\n",
      "Epoch 21/100\n",
      "105/105 [==============================] - 0s 181us/step - loss: 0.3963 - acc: 0.7333\n",
      "Epoch 22/100\n",
      "105/105 [==============================] - 0s 171us/step - loss: 0.3736 - acc: 0.7714\n",
      "Epoch 23/100\n",
      "105/105 [==============================] - 0s 181us/step - loss: 0.4488 - acc: 0.7143\n",
      "Epoch 24/100\n",
      "105/105 [==============================] - 0s 181us/step - loss: 0.4033 - acc: 0.7429\n",
      "Epoch 25/100\n",
      "105/105 [==============================] - 0s 181us/step - loss: 0.3054 - acc: 0.7429\n",
      "Epoch 26/100\n",
      "105/105 [==============================] - 0s 181us/step - loss: 0.4084 - acc: 0.7524\n",
      "Epoch 27/100\n",
      "105/105 [==============================] - 0s 181us/step - loss: 0.3236 - acc: 0.7714\n",
      "Epoch 28/100\n",
      "105/105 [==============================] - 0s 190us/step - loss: 0.3766 - acc: 0.7238\n",
      "Epoch 29/100\n",
      "105/105 [==============================] - 0s 181us/step - loss: 0.3038 - acc: 0.8000\n",
      "Epoch 30/100\n",
      "105/105 [==============================] - 0s 181us/step - loss: 0.3050 - acc: 0.7905\n",
      "Epoch 31/100\n",
      "105/105 [==============================] - 0s 181us/step - loss: 0.3568 - acc: 0.7524\n",
      "Epoch 32/100\n",
      "105/105 [==============================] - 0s 190us/step - loss: 0.4220 - acc: 0.7714\n",
      "Epoch 33/100\n",
      "105/105 [==============================] - 0s 190us/step - loss: 0.3265 - acc: 0.7810\n",
      "Epoch 34/100\n",
      "105/105 [==============================] - 0s 210us/step - loss: 0.2650 - acc: 0.7143\n",
      "Epoch 35/100\n",
      "105/105 [==============================] - 0s 219us/step - loss: 0.3332 - acc: 0.7238\n",
      "Epoch 36/100\n",
      "105/105 [==============================] - 0s 219us/step - loss: 0.3420 - acc: 0.7619\n",
      "Epoch 37/100\n",
      "105/105 [==============================] - 0s 210us/step - loss: 0.2961 - acc: 0.7905\n",
      "Epoch 38/100\n",
      "105/105 [==============================] - 0s 190us/step - loss: 0.2937 - acc: 0.8000\n",
      "Epoch 39/100\n",
      "105/105 [==============================] - 0s 190us/step - loss: 0.2756 - acc: 0.8190\n",
      "Epoch 40/100\n",
      "105/105 [==============================] - 0s 200us/step - loss: 0.2942 - acc: 0.8000\n",
      "Epoch 41/100\n",
      "105/105 [==============================] - 0s 200us/step - loss: 0.2796 - acc: 0.7905\n",
      "Epoch 42/100\n",
      "105/105 [==============================] - 0s 210us/step - loss: 0.2465 - acc: 0.8000\n",
      "Epoch 43/100\n",
      "105/105 [==============================] - 0s 162us/step - loss: 0.3151 - acc: 0.7810\n",
      "Epoch 44/100\n",
      "105/105 [==============================] - 0s 171us/step - loss: 0.2881 - acc: 0.8286\n",
      "Epoch 45/100\n",
      "105/105 [==============================] - 0s 181us/step - loss: 0.2488 - acc: 0.8381\n",
      "Epoch 46/100\n",
      "105/105 [==============================] - 0s 171us/step - loss: 0.2732 - acc: 0.8095\n",
      "Epoch 47/100\n",
      "105/105 [==============================] - 0s 171us/step - loss: 0.2334 - acc: 0.7810\n",
      "Epoch 48/100\n",
      "105/105 [==============================] - 0s 181us/step - loss: 0.2734 - acc: 0.7714\n",
      "Epoch 49/100\n",
      "105/105 [==============================] - 0s 181us/step - loss: 0.2468 - acc: 0.8667\n",
      "Epoch 50/100\n",
      "105/105 [==============================] - 0s 181us/step - loss: 0.2558 - acc: 0.8190\n",
      "Epoch 51/100\n",
      "105/105 [==============================] - 0s 171us/step - loss: 0.2386 - acc: 0.8476\n",
      "Epoch 52/100\n",
      "105/105 [==============================] - 0s 162us/step - loss: 0.2293 - acc: 0.8667\n",
      "Epoch 53/100\n",
      "105/105 [==============================] - 0s 171us/step - loss: 0.2278 - acc: 0.8476\n",
      "Epoch 54/100\n",
      "105/105 [==============================] - 0s 181us/step - loss: 0.2639 - acc: 0.8286\n",
      "Epoch 55/100\n",
      "105/105 [==============================] - 0s 162us/step - loss: 0.2196 - acc: 0.8286\n",
      "Epoch 56/100\n",
      "105/105 [==============================] - 0s 171us/step - loss: 0.2384 - acc: 0.8381\n",
      "Epoch 57/100\n",
      "105/105 [==============================] - 0s 181us/step - loss: 0.2576 - acc: 0.7810\n",
      "Epoch 58/100\n",
      "105/105 [==============================] - 0s 181us/step - loss: 0.2104 - acc: 0.7905\n",
      "Epoch 59/100\n",
      "105/105 [==============================] - 0s 171us/step - loss: 0.2710 - acc: 0.8000\n",
      "Epoch 60/100\n",
      "105/105 [==============================] - 0s 181us/step - loss: 0.2178 - acc: 0.7905\n",
      "Epoch 61/100\n",
      "105/105 [==============================] - 0s 190us/step - loss: 0.2133 - acc: 0.8381\n",
      "Epoch 62/100\n",
      "105/105 [==============================] - 0s 190us/step - loss: 0.2299 - acc: 0.7905\n",
      "Epoch 63/100\n",
      "105/105 [==============================] - 0s 171us/step - loss: 0.2210 - acc: 0.8190\n",
      "Epoch 64/100\n",
      "105/105 [==============================] - 0s 190us/step - loss: 0.1901 - acc: 0.9048\n",
      "Epoch 65/100\n",
      "105/105 [==============================] - 0s 181us/step - loss: 0.2214 - acc: 0.8762\n",
      "Epoch 66/100\n",
      "105/105 [==============================] - 0s 171us/step - loss: 0.1881 - acc: 0.8381\n",
      "Epoch 67/100\n",
      "105/105 [==============================] - 0s 171us/step - loss: 0.1964 - acc: 0.8476\n",
      "Epoch 68/100\n",
      "105/105 [==============================] - 0s 190us/step - loss: 0.2350 - acc: 0.8095\n",
      "Epoch 69/100\n",
      "105/105 [==============================] - ETA: 0s - loss: 0.2499 - acc: 0.937 - 0s 181us/step - loss: 0.2467 - acc: 0.8857\n",
      "Epoch 70/100\n",
      "105/105 [==============================] - 0s 181us/step - loss: 0.2039 - acc: 0.8571\n",
      "Epoch 71/100\n",
      "105/105 [==============================] - 0s 181us/step - loss: 0.1682 - acc: 0.8571\n",
      "Epoch 72/100\n",
      "105/105 [==============================] - 0s 190us/step - loss: 0.1865 - acc: 0.8381\n",
      "Epoch 73/100\n",
      "105/105 [==============================] - 0s 181us/step - loss: 0.2014 - acc: 0.8095\n",
      "Epoch 74/100\n",
      "105/105 [==============================] - 0s 181us/step - loss: 0.2025 - acc: 0.8286\n",
      "Epoch 75/100\n",
      "105/105 [==============================] - 0s 181us/step - loss: 0.1743 - acc: 0.8190\n",
      "Epoch 76/100\n",
      "105/105 [==============================] - 0s 181us/step - loss: 0.1900 - acc: 0.8476\n",
      "Epoch 77/100\n",
      "105/105 [==============================] - 0s 190us/step - loss: 0.2191 - acc: 0.8095\n",
      "Epoch 78/100\n",
      "105/105 [==============================] - 0s 171us/step - loss: 0.1737 - acc: 0.8762\n",
      "Epoch 79/100\n",
      "105/105 [==============================] - 0s 171us/step - loss: 0.1375 - acc: 0.9333\n",
      "Epoch 80/100\n",
      "105/105 [==============================] - 0s 190us/step - loss: 0.1871 - acc: 0.9143\n",
      "Epoch 81/100\n",
      "105/105 [==============================] - 0s 171us/step - loss: 0.1796 - acc: 0.8571\n",
      "Epoch 82/100\n",
      "105/105 [==============================] - 0s 190us/step - loss: 0.1711 - acc: 0.8857\n",
      "Epoch 83/100\n",
      "105/105 [==============================] - 0s 190us/step - loss: 0.1643 - acc: 0.9143\n",
      "Epoch 84/100\n",
      "105/105 [==============================] - 0s 190us/step - loss: 0.1759 - acc: 0.8476\n",
      "Epoch 85/100\n",
      "105/105 [==============================] - 0s 181us/step - loss: 0.1598 - acc: 0.8667\n",
      "Epoch 86/100\n",
      "105/105 [==============================] - 0s 190us/step - loss: 0.1997 - acc: 0.8476\n",
      "Epoch 87/100\n",
      "105/105 [==============================] - 0s 181us/step - loss: 0.1522 - acc: 0.8762\n",
      "Epoch 88/100\n",
      "105/105 [==============================] - 0s 181us/step - loss: 0.1543 - acc: 0.8667\n",
      "Epoch 89/100\n",
      "105/105 [==============================] - 0s 190us/step - loss: 0.1601 - acc: 0.8667\n",
      "Epoch 90/100\n",
      "105/105 [==============================] - 0s 171us/step - loss: 0.1581 - acc: 0.8571\n",
      "Epoch 91/100\n",
      "105/105 [==============================] - 0s 181us/step - loss: 0.1649 - acc: 0.8762\n",
      "Epoch 92/100\n",
      "105/105 [==============================] - 0s 181us/step - loss: 0.1552 - acc: 0.9238\n",
      "Epoch 93/100\n",
      "105/105 [==============================] - 0s 181us/step - loss: 0.1511 - acc: 0.8762\n",
      "Epoch 94/100\n",
      "105/105 [==============================] - 0s 171us/step - loss: 0.1651 - acc: 0.8667\n",
      "Epoch 95/100\n",
      "105/105 [==============================] - 0s 181us/step - loss: 0.1537 - acc: 0.8952\n",
      "Epoch 96/100\n",
      "105/105 [==============================] - 0s 190us/step - loss: 0.1679 - acc: 0.8667\n",
      "Epoch 97/100\n",
      "105/105 [==============================] - 0s 181us/step - loss: 0.1213 - acc: 0.9143\n",
      "Epoch 98/100\n",
      "105/105 [==============================] - 0s 171us/step - loss: 0.1396 - acc: 0.8286\n",
      "Epoch 99/100\n",
      "105/105 [==============================] - 0s 181us/step - loss: 0.1612 - acc: 0.8857\n",
      "Epoch 100/100\n",
      "105/105 [==============================] - 0s 190us/step - loss: 0.1498 - acc: 0.8476\n"
     ]
    }
   ],
   "source": [
    "# モデル生成\n",
    "model = Sequential()\n",
    "\n",
    "# 層の追加\n",
    "layers=[\n",
    "    Dense(128,input_shape=(4,), activation='relu'),\n",
    "    Dropout(0.1),\n",
    "    normalization.BatchNormalization(),\n",
    "    Dense(128, activation='relu'),\n",
    "    Dropout(0.1),\n",
    "    normalization.BatchNormalization(),\n",
    "    Dense(64, activation='relu'),\n",
    "    Dropout(0.1),\n",
    "    normalization.BatchNormalization(),\n",
    "    Dense(3, activation='linear')\n",
    "]\n",
    "for layer in layers:\n",
    "    model.add(layer)\n",
    "\n",
    "# モデルの学習設定\n",
    "model.compile(\n",
    "    loss=losses.mean_squared_error,\n",
    "    optimizer=optimizers.Adam(),\n",
    "    metrics=['acc'] # モニタリング指標\n",
    ")\n",
    "\n",
    "# モデルの学習\n",
    "result = model.fit(\n",
    "    X_train_n,\n",
    "    y_train,\n",
    "    batch_size=32,\n",
    "    epochs=100\n",
    ")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 40,
   "metadata": {
    "collapsed": true
   },
   "outputs": [
    {
     "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": [
    "# 描画データ\n",
    "acc = result.history['acc']    # 学習データの正答率\n",
    "loss = result.history['loss']  # 学習データの損失\n",
    "\n",
    "# x軸\n",
    "epochs = range(len(acc))\n",
    "\n",
    "# 図示\n",
    "plt.plot(epochs, acc, label='Train Accuracy')\n",
    "plt.plot(epochs, loss, label='Train Loss')\n",
    "plt.legend()\n",
    "plt.show()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# ②モデルの保存と読み込み\n",
    "\n",
    "学習したモデルは、保存できます。\n",
    "\n",
    "保存したモデルは、使用したいときにプログラムから読み込み利用できます。"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 41,
   "metadata": {},
   "outputs": [],
   "source": [
    "# モデルの保存\n",
    "model.save('./model/model.h5')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 42,
   "metadata": {},
   "outputs": [],
   "source": [
    "# モデルの読み込み\n",
    "model = keras.models.load_model('./model/model.h5')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 43,
   "metadata": {
    "collapsed": true
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "[2 1 0 1 0 2 0 1 1 1 1 1 1 1 1 0 1 1 1 1 2 1 1 0 1 0 1 1 1 1 1 1 1 1 2 1 1\n",
      " 1 1 1 2 0 2 1 1]\n"
     ]
    }
   ],
   "source": [
    "# 予測値の算出\n",
    "import numpy as np\n",
    "y_predict = np.argmax(model.predict(X_test_n),axis=1)\n",
    "print(y_predict)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# ③コールバック\n",
    "\n",
    "コールバックは訓練中で適用される関数のことです。訓練中にモデル内部の状態と統計量を可視化する際に、コールバックを使います。その他にも、各エポック毎に「何かしらの処理」を行いたい場合、ここに記述します。\n",
    "\n",
    "\n",
    "> コールバックについて紹介する前に、データセットの種類について紹介します。\n",
    "> \n",
    "> + **訓練データ**：重み更新に使用する。\n",
    "> + **テストデータ**：モデル性能の検証に利用する。\n",
    "> + **検証データ(バリデーションデータ)**：学習の早期終了（EarlyStopping）やハイパーパラメータの調整に利用する\n",
    "> \n",
    "> ＊ ほとんどの場合、ハイパーパラメータの調整は、訓練データを用いて行うため、気にしなくてよいが、学術的なデータ解析の際には、検証用データも使用する。ほとんどの実務的な場合においては、検証データは、学習の早期終了などに用いられると考えてよい。\n",
    "\n",
    "\n",
    "### コールバックで出来ること\n",
    "\n",
    "**1. 学習の早期終了（EarlyStopping）**：バリデーションデータへの精度に基づいて、改善しなかった場合は、学習を早期に打ち切り、過学習の防止に務める。\n",
    "\n",
    "**2. 学習過程の保存（ModelCheckpoint）**：学習の途中経過を記録しておく。今まで見てきた手法では、学習終了時点のモデルを採用し、予測値の算出を行ったが、学習途中の精度が良い場合がある。そのような場合は、このModelCheckpointで記録しておいたモデルの中から、最も精度のいいモデルを採用する。"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### 1.学習の早期終了（EarlyStopping）"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 44,
   "metadata": {
    "collapsed": true
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Train on 105 samples, validate on 45 samples\n",
      "Epoch 1/300\n",
      "105/105 [==============================] - 1s 8ms/step - loss: 1.4757 - acc: 0.5048 - val_loss: 0.3340 - val_acc: 0.2000\n",
      "Epoch 2/300\n",
      "105/105 [==============================] - 0s 238us/step - loss: 1.2322 - acc: 0.5905 - val_loss: 0.3106 - val_acc: 0.0667\n",
      "Epoch 3/300\n",
      "105/105 [==============================] - 0s 219us/step - loss: 0.9269 - acc: 0.5905 - val_loss: 0.2922 - val_acc: 0.2889\n",
      "Epoch 4/300\n",
      "105/105 [==============================] - 0s 238us/step - loss: 1.0448 - acc: 0.5714 - val_loss: 0.2770 - val_acc: 0.2889\n",
      "Epoch 5/300\n",
      "105/105 [==============================] - 0s 247us/step - loss: 0.8049 - acc: 0.6095 - val_loss: 0.2646 - val_acc: 0.2667\n",
      "Epoch 6/300\n",
      "105/105 [==============================] - 0s 219us/step - loss: 0.7839 - acc: 0.6571 - val_loss: 0.2515 - val_acc: 0.4667\n",
      "Epoch 7/300\n",
      "105/105 [==============================] - 0s 229us/step - loss: 0.6561 - acc: 0.7333 - val_loss: 0.2402 - val_acc: 0.5556\n",
      "Epoch 8/300\n",
      "105/105 [==============================] - 0s 219us/step - loss: 0.7040 - acc: 0.6095 - val_loss: 0.2313 - val_acc: 0.6000\n",
      "Epoch 9/300\n",
      "105/105 [==============================] - 0s 248us/step - loss: 0.5610 - acc: 0.7429 - val_loss: 0.2242 - val_acc: 0.5333\n",
      "Epoch 10/300\n",
      "105/105 [==============================] - 0s 229us/step - loss: 0.4953 - acc: 0.6476 - val_loss: 0.2192 - val_acc: 0.5333\n",
      "Epoch 11/300\n",
      "105/105 [==============================] - 0s 248us/step - loss: 0.6579 - acc: 0.6381 - val_loss: 0.2141 - val_acc: 0.4667\n",
      "Epoch 12/300\n",
      "105/105 [==============================] - 0s 238us/step - loss: 0.4717 - acc: 0.7333 - val_loss: 0.2111 - val_acc: 0.4222\n",
      "Epoch 13/300\n",
      "105/105 [==============================] - 0s 238us/step - loss: 0.4183 - acc: 0.7905 - val_loss: 0.2098 - val_acc: 0.4000\n",
      "Epoch 14/300\n",
      "105/105 [==============================] - 0s 238us/step - loss: 0.4942 - acc: 0.7238 - val_loss: 0.2098 - val_acc: 0.4000\n",
      "Epoch 15/300\n",
      "105/105 [==============================] - 0s 238us/step - loss: 0.4294 - acc: 0.7143 - val_loss: 0.2105 - val_acc: 0.4000\n",
      "Epoch 16/300\n",
      "105/105 [==============================] - 0s 219us/step - loss: 0.4642 - acc: 0.7429 - val_loss: 0.2119 - val_acc: 0.4000\n",
      "Epoch 17/300\n",
      "105/105 [==============================] - 0s 238us/step - loss: 0.4961 - acc: 0.7333 - val_loss: 0.2152 - val_acc: 0.4000\n",
      "Epoch 18/300\n",
      "105/105 [==============================] - 0s 229us/step - loss: 0.3765 - acc: 0.6952 - val_loss: 0.2156 - val_acc: 0.4000\n",
      "Epoch 19/300\n",
      "105/105 [==============================] - 0s 257us/step - loss: 0.4034 - acc: 0.7714 - val_loss: 0.2147 - val_acc: 0.4000\n",
      "Epoch 20/300\n",
      "105/105 [==============================] - 0s 248us/step - loss: 0.3955 - acc: 0.7619 - val_loss: 0.2132 - val_acc: 0.4000\n",
      "Epoch 21/300\n",
      "105/105 [==============================] - 0s 238us/step - loss: 0.3664 - acc: 0.7810 - val_loss: 0.2116 - val_acc: 0.4000\n",
      "Epoch 22/300\n",
      "105/105 [==============================] - 0s 238us/step - loss: 0.3201 - acc: 0.7619 - val_loss: 0.2111 - val_acc: 0.4000\n",
      "Epoch 23/300\n",
      "105/105 [==============================] - 0s 248us/step - loss: 0.3335 - acc: 0.8095 - val_loss: 0.2106 - val_acc: 0.4000\n",
      "Epoch 24/300\n",
      "105/105 [==============================] - 0s 229us/step - loss: 0.4176 - acc: 0.7238 - val_loss: 0.2092 - val_acc: 0.4000\n",
      "Epoch 25/300\n",
      "105/105 [==============================] - 0s 248us/step - loss: 0.3581 - acc: 0.7524 - val_loss: 0.2091 - val_acc: 0.4000\n",
      "Epoch 26/300\n",
      "105/105 [==============================] - 0s 238us/step - loss: 0.3097 - acc: 0.8095 - val_loss: 0.2080 - val_acc: 0.4000\n",
      "Epoch 27/300\n",
      "105/105 [==============================] - 0s 248us/step - loss: 0.2848 - acc: 0.7619 - val_loss: 0.2084 - val_acc: 0.4000\n",
      "Epoch 28/300\n",
      "105/105 [==============================] - 0s 229us/step - loss: 0.3790 - acc: 0.7238 - val_loss: 0.2105 - val_acc: 0.4000\n",
      "Epoch 29/300\n",
      "105/105 [==============================] - 0s 238us/step - loss: 0.3112 - acc: 0.8286 - val_loss: 0.2116 - val_acc: 0.4000\n",
      "Epoch 30/300\n",
      "105/105 [==============================] - 0s 229us/step - loss: 0.3109 - acc: 0.8000 - val_loss: 0.2119 - val_acc: 0.4000\n",
      "Epoch 31/300\n",
      "105/105 [==============================] - 0s 248us/step - loss: 0.3105 - acc: 0.7619 - val_loss: 0.2128 - val_acc: 0.4000\n",
      "Epoch 32/300\n",
      "105/105 [==============================] - 0s 238us/step - loss: 0.3070 - acc: 0.7429 - val_loss: 0.2125 - val_acc: 0.4000\n",
      "Epoch 33/300\n",
      "105/105 [==============================] - 0s 238us/step - loss: 0.2857 - acc: 0.7810 - val_loss: 0.2156 - val_acc: 0.4000\n",
      "Epoch 34/300\n",
      "105/105 [==============================] - 0s 238us/step - loss: 0.2888 - acc: 0.7524 - val_loss: 0.2161 - val_acc: 0.4000\n",
      "Epoch 35/300\n",
      "105/105 [==============================] - 0s 238us/step - loss: 0.2783 - acc: 0.8000 - val_loss: 0.2159 - val_acc: 0.4000\n",
      "Epoch 36/300\n",
      "105/105 [==============================] - 0s 248us/step - loss: 0.2708 - acc: 0.8095 - val_loss: 0.2131 - val_acc: 0.4000\n"
     ]
    }
   ],
   "source": [
    "# モデル生成\n",
    "model = Sequential()\n",
    "\n",
    "# 層の追加\n",
    "layers=[\n",
    "    Dense(128, activation='relu'),\n",
    "    Dropout(0.1),\n",
    "    normalization.BatchNormalization(),\n",
    "    Dense(128, activation='relu'),\n",
    "    Dropout(0.1),\n",
    "    normalization.BatchNormalization(),\n",
    "    Dense(64, activation='relu'),\n",
    "    Dropout(0.1),\n",
    "    normalization.BatchNormalization(),\n",
    "    Dense(3, activation='linear')\n",
    "]\n",
    "for layer in layers:\n",
    "    model.add(layer)\n",
    "\n",
    "# モデルの学習設定\n",
    "model.compile(\n",
    "    loss=losses.mean_squared_error,\n",
    "    optimizer=optimizers.Adam(),\n",
    "    metrics=['acc']\n",
    ")\n",
    "\n",
    "# モデルの学習\n",
    "result = model.fit(\n",
    "    X_train_n,\n",
    "    y_train,\n",
    "    batch_size=32,\n",
    "    epochs=300, \n",
    "    validation_data=(X_test_n,y_test),\n",
    "    callbacks=[\n",
    "        callbacks.EarlyStopping(\n",
    "            monitor='val_loss',\n",
    "            patience=10         # 様子見するエポック数\n",
    "        )\n",
    "    ]\n",
    ")"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### 2.学習過程の保存（ModelCheckpoint）"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 49,
   "metadata": {
    "collapsed": true
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Train on 105 samples, validate on 45 samples\n",
      "Epoch 1/10\n",
      "105/105 [==============================] - 1s 9ms/step - loss: 1.7516 - acc: 0.4381 - val_loss: 0.3284 - val_acc: 0.5778\n",
      "Epoch 2/10\n",
      "105/105 [==============================] - 0s 276us/step - loss: 1.4786 - acc: 0.5333 - val_loss: 0.3016 - val_acc: 0.3556\n",
      "Epoch 3/10\n",
      "105/105 [==============================] - 0s 247us/step - loss: 1.2214 - acc: 0.6952 - val_loss: 0.2783 - val_acc: 0.5333\n",
      "Epoch 4/10\n",
      "105/105 [==============================] - 0s 238us/step - loss: 1.3740 - acc: 0.4952 - val_loss: 0.2598 - val_acc: 0.4222\n",
      "Epoch 5/10\n",
      "105/105 [==============================] - 0s 238us/step - loss: 1.1571 - acc: 0.4952 - val_loss: 0.2471 - val_acc: 0.4000\n",
      "Epoch 6/10\n",
      "105/105 [==============================] - 0s 248us/step - loss: 1.0838 - acc: 0.6286 - val_loss: 0.2399 - val_acc: 0.4000\n",
      "Epoch 7/10\n",
      "105/105 [==============================] - 0s 276us/step - loss: 1.0153 - acc: 0.6952 - val_loss: 0.2378 - val_acc: 0.4000\n",
      "Epoch 8/10\n",
      "105/105 [==============================] - 0s 257us/step - loss: 0.7461 - acc: 0.6571 - val_loss: 0.2395 - val_acc: 0.4000\n",
      "Epoch 9/10\n",
      "105/105 [==============================] - 0s 267us/step - loss: 0.8571 - acc: 0.6571 - val_loss: 0.2428 - val_acc: 0.4000\n",
      "Epoch 10/10\n",
      "105/105 [==============================] - 0s 267us/step - loss: 0.7573 - acc: 0.6571 - val_loss: 0.2475 - val_acc: 0.4000\n"
     ]
    }
   ],
   "source": [
    "# モデル生成\n",
    "model = Sequential()\n",
    "\n",
    "# 層の追加\n",
    "layers=[\n",
    "    Dense(128, activation='relu'),\n",
    "    Dropout(0.1),\n",
    "    normalization.BatchNormalization(),\n",
    "    Dense(128, activation='relu'),\n",
    "    Dropout(0.1),\n",
    "    normalization.BatchNormalization(),\n",
    "    Dense(64, activation='relu'),\n",
    "    Dropout(0.1),\n",
    "    normalization.BatchNormalization(),\n",
    "    Dense(3, activation='linear')\n",
    "]\n",
    "for layer in layers:\n",
    "    model.add(layer)\n",
    "\n",
    "# モデルの学習設定\n",
    "model.compile(\n",
    "    loss=losses.mean_squared_error,\n",
    "    optimizer=optimizers.Adam(),\n",
    "    metrics=['acc']\n",
    ")\n",
    "\n",
    "# モデルの学習\n",
    "result = model.fit(\n",
    "    X_train_n,\n",
    "    y_train,\n",
    "    batch_size=32,\n",
    "    epochs=10,\n",
    "    validation_data=(X_test_n,y_test),\n",
    "    callbacks=[\n",
    "        callbacks.ModelCheckpoint(\n",
    "            filepath = './model/best_model.h5', # 最適なモデルの保存先\n",
    "            monitor='val_loss', \n",
    "            save_best_only=True\n",
    "        )\n",
    "    ]\n",
    ")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 50,
   "metadata": {
    "collapsed": true
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "[[0.05930027 0.3430214  0.22959554]\n",
      " [0.05934943 0.43295258 0.21278289]\n",
      " [0.1278674  0.48997056 0.15808052]\n",
      " [0.03592425 0.33646876 0.24036348]\n",
      " [0.10205448 0.49092472 0.17329884]\n",
      " [0.04257865 0.30185586 0.2487933 ]\n",
      " [0.10584481 0.49729443 0.17316812]\n",
      " [0.06102544 0.3878237  0.21672049]\n",
      " [0.05564784 0.3878922  0.2104899 ]\n",
      " [0.05994523 0.4182344  0.21257496]\n",
      " [0.03730516 0.39088362 0.21516088]\n",
      " [0.06378451 0.39696425 0.21781287]\n",
      " [0.0553444  0.40874594 0.20600522]\n",
      " [0.05287849 0.3964197  0.21759549]\n",
      " [0.05566785 0.40245712 0.21626315]\n",
      " [0.10221743 0.49720657 0.17267498]\n",
      " [0.05796535 0.40223843 0.22066969]\n",
      " [0.05724071 0.4356709  0.21062246]\n",
      " [0.08896141 0.478706   0.17579108]\n",
      " [0.12268116 0.49760956 0.16269466]\n",
      " [0.05553887 0.37737054 0.22570205]\n",
      " [0.0625523  0.4114194  0.2204031 ]\n",
      " [0.08889423 0.4851554  0.17500228]\n",
      " [0.08028241 0.4741372  0.17432153]\n",
      " [0.05154018 0.37930095 0.2317046 ]\n",
      " [0.0906919  0.5040612  0.17199141]\n",
      " [0.10997855 0.48887646 0.17437798]\n",
      " [0.0606271  0.41060257 0.20955423]\n",
      " [0.07816874 0.4571175  0.19645381]\n",
      " [0.10131835 0.49102598 0.1776253 ]\n",
      " [0.05073795 0.36286473 0.23643696]\n",
      " [0.06497294 0.415586   0.21885204]\n",
      " [0.11252225 0.49603707 0.17042917]\n",
      " [0.05438586 0.37510175 0.23193565]\n",
      " [0.03987936 0.34057903 0.23967764]\n",
      " [0.06999377 0.44146514 0.21453896]\n",
      " [0.13294895 0.4885183  0.16240469]\n",
      " [0.04575501 0.39023536 0.22207662]\n",
      " [0.06201271 0.40931344 0.2188488 ]\n",
      " [0.06348844 0.42981356 0.21218398]\n",
      " [0.04334149 0.34382623 0.24486858]\n",
      " [0.08576224 0.48942536 0.17581871]\n",
      " [0.03891354 0.34099913 0.24150166]\n",
      " [0.10259671 0.4920513  0.18425107]\n",
      " [0.1242815  0.49641752 0.16546082]]\n"
     ]
    }
   ],
   "source": [
    "# 最適モデルの読み込み\n",
    "model = keras.models.load_model('./model/best_model.h5')\n",
    "y_predict = model.predict(X_test_n)\n",
    "print(y_predict)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# 自作のコールバック関数\n",
    "\n",
    "コールバック関数は、自作したものを使用することもできる。\n",
    "\n",
    "今回は、EPOCH毎に、精度を描画するコールバック関数を実装する。"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 51,
   "metadata": {},
   "outputs": [],
   "source": [
    "class LossHistory(keras.callbacks.Callback):\n",
    "    def __init__(self):\n",
    "        \"\"\"学習全体を通して永続的に使用したい変数を定義する\n",
    "        \"\"\"\n",
    "        # 各epoch毎にlossを格納していく\n",
    "        self.loss = []\n",
    "\n",
    "    def on_epoch_end(self, epoch, logs={}):\n",
    "        \"\"\"各epoch終了ごとに呼び出される\n",
    "        \"\"\"\n",
    "        # lossを格納\n",
    "        self.loss.append(logs['loss'])\n",
    "        \n",
    "        # グラフ初期化\n",
    "        clear_output(wait = True)\n",
    "        \n",
    "        # グラフ描画部\n",
    "        plt.figure(num=1, clear=True)\n",
    "        plt.title('loss')\n",
    "        plt.xlabel('epoch')\n",
    "        plt.ylabel('loss')\n",
    "        plt.plot(self.loss, label='train')\n",
    "        plt.pause(0.1)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 52,
   "metadata": {
    "collapsed": true
   },
   "outputs": [
    {
     "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": [
    "# 必要ライブラリのインポート\n",
    "from keras.models import Sequential\n",
    "from keras.layers import Dense,Dropout\n",
    "from keras import losses\n",
    "from keras import optimizers\n",
    "from keras import callbacks\n",
    "import os\n",
    "from IPython.display import clear_output\n",
    "\n",
    "# モデル生成\n",
    "model = Sequential()\n",
    "\n",
    "# 層の追加\n",
    "layers=[\n",
    "    Dense(128, activation='relu'),\n",
    "    Dropout(0.1),\n",
    "    normalization.BatchNormalization(),\n",
    "    Dense(128, activation='relu'),\n",
    "    Dropout(0.1),\n",
    "    normalization.BatchNormalization(),\n",
    "    Dense(64, activation='relu'),\n",
    "    Dropout(0.1),\n",
    "    normalization.BatchNormalization(),\n",
    "    Dense(3, activation='linear')\n",
    "]\n",
    "for layer in layers:\n",
    "    model.add(layer)\n",
    "\n",
    "# モデルの学習設定\n",
    "model.compile(\n",
    "    loss=losses.mean_squared_error,\n",
    "    optimizer=optimizers.Adam(),\n",
    "    metrics=['acc']\n",
    ")\n",
    "\n",
    "# モデルの学習\n",
    "result = model.fit(\n",
    "    X_train_n,\n",
    "    y_train,\n",
    "    batch_size=32,\n",
    "    epochs=10,\n",
    "    verbose=False,\n",
    "    validation_data=(X_test_n,y_test),\n",
    "    callbacks=[\n",
    "        LossHistory()\n",
    "    ]\n",
    ")"
   ]
  }
 ],
 "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
}