09-Autoencoder-keras/keras-ae+feature-reduction-tutorial.ipynb

{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": 24,
   "metadata": {},
   "outputs": [],
   "source": [
    "# -*- coding: utf-8 -*-\n",
    "\n",
    "%matplotlib inline\n",
    "\n",
    "import pandas as pd\n",
    "import numpy as np\n",
    "import matplotlib.pyplot as plt\n",
    "import seaborn as sns\n",
    "\n",
    "import warnings\n",
    "warnings.filterwarnings(\"ignore\")"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "-----"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# AE (Auto Encoder)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 25,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "(60000, 784)\n",
      "(10000, 784)\n"
     ]
    }
   ],
   "source": [
    "from keras.datasets import mnist\n",
    "\n",
    "(x_train, y_train), (x_test, y_test) = mnist.load_data()\n",
    "x_train = x_train.astype('float32') / 255.\n",
    "x_test = x_test.astype('float32') / 255.\n",
    "x_train = x_train.reshape((len(x_train), np.prod(x_train.shape[1:])))\n",
    "x_test = x_test.reshape((len(x_test), np.prod(x_test.shape[1:])))\n",
    "print (x_train.shape)\n",
    "print (x_test.shape)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {},
   "outputs": [],
   "source": [
    "from keras.layers import Input, Dense\n",
    "from keras.models import Model\n",
    "from keras import backend as K\n",
    "\n",
    "# configure\n",
    "encoding_dim = 32\n",
    "input_img = Input(shape=(784,))\n",
    "\n",
    "\n",
    "# layers\n",
    "encoded = Dense(encoding_dim, activation='relu')(input_img)\n",
    "decoded = Dense(784, activation='sigmoid')(encoded)\n",
    "\n",
    "\n",
    "# Models\n",
    "autoencoder = Model(input_img, decoded) # autoencoder\n",
    "\n",
    "encoder = Model(input_img, encoded) # encoder\n",
    "\n",
    "encoded_input = Input(shape=(encoding_dim,))\n",
    "decoder_layer = autoencoder.layers[-1]\n",
    "decoder = Model(encoded_input, decoder_layer(encoded_input)) # decoder"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Train on 60000 samples, validate on 10000 samples\n",
      "Epoch 1/50\n",
      "60000/60000 [==============================] - 2s 35us/step - loss: 0.3601 - rmse: 0.3039 - recall: 0.2390 - val_loss: 0.2703 - val_rmse: 0.2595 - val_recall: 0.2043\n",
      "Epoch 2/50\n",
      "60000/60000 [==============================] - 2s 30us/step - loss: 0.2626 - rmse: 0.2538 - recall: 0.2422 - val_loss: 0.2512 - val_rmse: 0.2460 - val_recall: 0.3215\n",
      "Epoch 3/50\n",
      "60000/60000 [==============================] - 2s 29us/step - loss: 0.2410 - rmse: 0.2373 - recall: 0.3931 - val_loss: 0.2287 - val_rmse: 0.2281 - val_recall: 0.4780\n",
      "Epoch 4/50\n",
      "60000/60000 [==============================] - 2s 30us/step - loss: 0.2205 - rmse: 0.2210 - recall: 0.5166 - val_loss: 0.2101 - val_rmse: 0.2132 - val_recall: 0.5599\n",
      "Epoch 5/50\n",
      "60000/60000 [==============================] - 2s 31us/step - loss: 0.2047 - rmse: 0.2085 - recall: 0.5910 - val_loss: 0.1970 - val_rmse: 0.2027 - val_recall: 0.6182\n",
      "Epoch 6/50\n",
      "60000/60000 [==============================] - 2s 30us/step - loss: 0.1938 - rmse: 0.1998 - recall: 0.6342 - val_loss: 0.1879 - val_rmse: 0.1953 - val_recall: 0.6453\n",
      "Epoch 7/50\n",
      "60000/60000 [==============================] - 2s 30us/step - loss: 0.1856 - rmse: 0.1931 - recall: 0.6626 - val_loss: 0.1806 - val_rmse: 0.1891 - val_recall: 0.6737\n",
      "Epoch 8/50\n",
      "60000/60000 [==============================] - 2s 31us/step - loss: 0.1790 - rmse: 0.1874 - recall: 0.6840 - val_loss: 0.1745 - val_rmse: 0.1838 - val_recall: 0.6990\n",
      "Epoch 9/50\n",
      "60000/60000 [==============================] - 2s 32us/step - loss: 0.1734 - rmse: 0.1826 - recall: 0.7013 - val_loss: 0.1695 - val_rmse: 0.1794 - val_recall: 0.7007\n",
      "Epoch 10/50\n",
      "60000/60000 [==============================] - 2s 31us/step - loss: 0.1684 - rmse: 0.1781 - recall: 0.7153 - val_loss: 0.1647 - val_rmse: 0.1751 - val_recall: 0.7153\n",
      "Epoch 11/50\n",
      "60000/60000 [==============================] - 2s 30us/step - loss: 0.1639 - rmse: 0.1740 - recall: 0.7278 - val_loss: 0.1602 - val_rmse: 0.1709 - val_recall: 0.7391\n",
      "Epoch 12/50\n",
      "60000/60000 [==============================] - 2s 31us/step - loss: 0.1596 - rmse: 0.1699 - recall: 0.7396 - val_loss: 0.1561 - val_rmse: 0.1669 - val_recall: 0.7427\n",
      "Epoch 13/50\n",
      "60000/60000 [==============================] - 2s 31us/step - loss: 0.1556 - rmse: 0.1661 - recall: 0.7500 - val_loss: 0.1522 - val_rmse: 0.1631 - val_recall: 0.7518\n",
      "Epoch 14/50\n",
      "60000/60000 [==============================] - 2s 30us/step - loss: 0.1518 - rmse: 0.1624 - recall: 0.7594 - val_loss: 0.1486 - val_rmse: 0.1595 - val_recall: 0.7634\n",
      "Epoch 15/50\n",
      "60000/60000 [==============================] - 2s 30us/step - loss: 0.1483 - rmse: 0.1589 - recall: 0.7678 - val_loss: 0.1451 - val_rmse: 0.1561 - val_recall: 0.7753\n",
      "Epoch 16/50\n",
      "60000/60000 [==============================] - 2s 32us/step - loss: 0.1451 - rmse: 0.1557 - recall: 0.7756 - val_loss: 0.1421 - val_rmse: 0.1530 - val_recall: 0.7777\n",
      "Epoch 17/50\n",
      "60000/60000 [==============================] - 2s 30us/step - loss: 0.1421 - rmse: 0.1527 - recall: 0.7824 - val_loss: 0.1392 - val_rmse: 0.1502 - val_recall: 0.7815\n",
      "Epoch 18/50\n",
      "60000/60000 [==============================] - 2s 30us/step - loss: 0.1393 - rmse: 0.1499 - recall: 0.7884 - val_loss: 0.1367 - val_rmse: 0.1476 - val_recall: 0.7841\n",
      "Epoch 19/50\n",
      "60000/60000 [==============================] - 2s 30us/step - loss: 0.1368 - rmse: 0.1473 - recall: 0.7937 - val_loss: 0.1342 - val_rmse: 0.1450 - val_recall: 0.7908\n",
      "Epoch 20/50\n",
      "60000/60000 [==============================] - 2s 30us/step - loss: 0.1345 - rmse: 0.1449 - recall: 0.7985 - val_loss: 0.1319 - val_rmse: 0.1426 - val_recall: 0.7965\n",
      "Epoch 21/50\n",
      "60000/60000 [==============================] - 2s 31us/step - loss: 0.1323 - rmse: 0.1426 - recall: 0.8028 - val_loss: 0.1297 - val_rmse: 0.1403 - val_recall: 0.8038\n",
      "Epoch 22/50\n",
      "60000/60000 [==============================] - 2s 30us/step - loss: 0.1302 - rmse: 0.1405 - recall: 0.8065 - val_loss: 0.1277 - val_rmse: 0.1381 - val_recall: 0.8124\n",
      "Epoch 23/50\n",
      "60000/60000 [==============================] - 2s 30us/step - loss: 0.1283 - rmse: 0.1384 - recall: 0.8102 - val_loss: 0.1259 - val_rmse: 0.1362 - val_recall: 0.8093\n",
      "Epoch 24/50\n",
      "60000/60000 [==============================] - 2s 31us/step - loss: 0.1265 - rmse: 0.1365 - recall: 0.8132 - val_loss: 0.1241 - val_rmse: 0.1342 - val_recall: 0.8132\n",
      "Epoch 25/50\n",
      "60000/60000 [==============================] - 2s 30us/step - loss: 0.1247 - rmse: 0.1346 - recall: 0.8161 - val_loss: 0.1223 - val_rmse: 0.1323 - val_recall: 0.8191\n",
      "Epoch 26/50\n",
      "60000/60000 [==============================] - 2s 31us/step - loss: 0.1231 - rmse: 0.1328 - recall: 0.8189 - val_loss: 0.1207 - val_rmse: 0.1304 - val_recall: 0.8236\n",
      "Epoch 27/50\n",
      "60000/60000 [==============================] - 2s 31us/step - loss: 0.1215 - rmse: 0.1310 - recall: 0.8215 - val_loss: 0.1192 - val_rmse: 0.1288 - val_recall: 0.8216\n",
      "Epoch 28/50\n",
      "60000/60000 [==============================] - 2s 30us/step - loss: 0.1200 - rmse: 0.1294 - recall: 0.8238 - val_loss: 0.1177 - val_rmse: 0.1271 - val_recall: 0.8275\n",
      "Epoch 29/50\n",
      "60000/60000 [==============================] - 2s 31us/step - loss: 0.1186 - rmse: 0.1278 - recall: 0.8262 - val_loss: 0.1164 - val_rmse: 0.1256 - val_recall: 0.8255\n",
      "Epoch 30/50\n",
      "60000/60000 [==============================] - 2s 31us/step - loss: 0.1173 - rmse: 0.1263 - recall: 0.8281 - val_loss: 0.1151 - val_rmse: 0.1240 - val_recall: 0.8311\n",
      "Epoch 31/50\n",
      "60000/60000 [==============================] - 2s 30us/step - loss: 0.1161 - rmse: 0.1248 - recall: 0.8301 - val_loss: 0.1139 - val_rmse: 0.1226 - val_recall: 0.8340\n",
      "Epoch 32/50\n",
      "60000/60000 [==============================] - 2s 32us/step - loss: 0.1149 - rmse: 0.1235 - recall: 0.8319 - val_loss: 0.1127 - val_rmse: 0.1213 - val_recall: 0.8338\n",
      "Epoch 33/50\n",
      "60000/60000 [==============================] - 2s 31us/step - loss: 0.1138 - rmse: 0.1222 - recall: 0.8335 - val_loss: 0.1117 - val_rmse: 0.1200 - val_recall: 0.8364\n",
      "Epoch 34/50\n",
      "60000/60000 [==============================] - 2s 30us/step - loss: 0.1128 - rmse: 0.1210 - recall: 0.8351 - val_loss: 0.1107 - val_rmse: 0.1189 - val_recall: 0.8356\n",
      "Epoch 35/50\n",
      "60000/60000 [==============================] - 2s 31us/step - loss: 0.1118 - rmse: 0.1198 - recall: 0.8365 - val_loss: 0.1097 - val_rmse: 0.1177 - val_recall: 0.8407\n",
      "Epoch 36/50\n",
      "60000/60000 [==============================] - 2s 31us/step - loss: 0.1109 - rmse: 0.1188 - recall: 0.8379 - val_loss: 0.1089 - val_rmse: 0.1166 - val_recall: 0.8448\n",
      "Epoch 37/50\n",
      "60000/60000 [==============================] - 2s 33us/step - loss: 0.1101 - rmse: 0.1178 - recall: 0.8393 - val_loss: 0.1081 - val_rmse: 0.1157 - val_recall: 0.8445\n",
      "Epoch 38/50\n",
      "60000/60000 [==============================] - 2s 33us/step - loss: 0.1093 - rmse: 0.1169 - recall: 0.8406 - val_loss: 0.1073 - val_rmse: 0.1148 - val_recall: 0.8447\n",
      "Epoch 39/50\n",
      "60000/60000 [==============================] - 2s 31us/step - loss: 0.1086 - rmse: 0.1160 - recall: 0.8416 - val_loss: 0.1066 - val_rmse: 0.1138 - val_recall: 0.8491\n",
      "Epoch 40/50\n",
      "60000/60000 [==============================] - 2s 33us/step - loss: 0.1079 - rmse: 0.1152 - recall: 0.8430 - val_loss: 0.1060 - val_rmse: 0.1131 - val_recall: 0.8485\n",
      "Epoch 41/50\n",
      "60000/60000 [==============================] - 2s 31us/step - loss: 0.1073 - rmse: 0.1144 - recall: 0.8440 - val_loss: 0.1054 - val_rmse: 0.1123 - val_recall: 0.8481\n",
      "Epoch 42/50\n",
      "60000/60000 [==============================] - 2s 31us/step - loss: 0.1067 - rmse: 0.1137 - recall: 0.8450 - val_loss: 0.1048 - val_rmse: 0.1116 - val_recall: 0.8508\n",
      "Epoch 43/50\n",
      "60000/60000 [==============================] - 2s 31us/step - loss: 0.1062 - rmse: 0.1130 - recall: 0.8461 - val_loss: 0.1043 - val_rmse: 0.1110 - val_recall: 0.8502\n",
      "Epoch 44/50\n",
      "60000/60000 [==============================] - 2s 31us/step - loss: 0.1056 - rmse: 0.1123 - recall: 0.8471 - val_loss: 0.1038 - val_rmse: 0.1102 - val_recall: 0.8541\n",
      "Epoch 45/50\n",
      "60000/60000 [==============================] - 2s 33us/step - loss: 0.1051 - rmse: 0.1117 - recall: 0.8481 - val_loss: 0.1033 - val_rmse: 0.1097 - val_recall: 0.8517\n",
      "Epoch 46/50\n",
      "60000/60000 [==============================] - 2s 32us/step - loss: 0.1047 - rmse: 0.1111 - recall: 0.8491 - val_loss: 0.1028 - val_rmse: 0.1091 - val_recall: 0.8520\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Epoch 47/50\n",
      "60000/60000 [==============================] - 2s 31us/step - loss: 0.1042 - rmse: 0.1106 - recall: 0.8499 - val_loss: 0.1024 - val_rmse: 0.1085 - val_recall: 0.8551\n",
      "Epoch 48/50\n",
      "60000/60000 [==============================] - 2s 31us/step - loss: 0.1038 - rmse: 0.1100 - recall: 0.8509 - val_loss: 0.1020 - val_rmse: 0.1081 - val_recall: 0.8541\n",
      "Epoch 49/50\n",
      "60000/60000 [==============================] - 2s 32us/step - loss: 0.1034 - rmse: 0.1095 - recall: 0.8517 - val_loss: 0.1016 - val_rmse: 0.1075 - val_recall: 0.8561\n",
      "Epoch 50/50\n",
      "60000/60000 [==============================] - 2s 31us/step - loss: 0.1030 - rmse: 0.1091 - recall: 0.8525 - val_loss: 0.1012 - val_rmse: 0.1070 - val_recall: 0.8609\n"
     ]
    }
   ],
   "source": [
    "def rmse(y_true, y_pred):\n",
    "    return K.sqrt(K.mean(K.square(y_pred - y_true), axis=-1))\n",
    "\n",
    "def recall(y_true, y_pred):\n",
    "    y_true_yn = K.round(K.clip(y_true, 0, 1))\n",
    "    y_pred_yn = K.round(K.clip(y_pred, 0, 1))\n",
    "\n",
    "    count_true_positive = K.sum(y_true_yn * y_pred_yn)\n",
    "    count_true_positive_false_negative = K.sum(y_true_yn)\n",
    "    recall = count_true_positive / (count_true_positive_false_negative + K.epsilon())\n",
    "    return recall\n",
    "\n",
    "\n",
    "# train autoencoder\n",
    "autoencoder.compile(optimizer='adadelta', loss='binary_crossentropy', metrics=[rmse, recall])\n",
    "autoencoder.fit(x_train, x_train,\n",
    "                epochs=50,\n",
    "                batch_size=256,\n",
    "                shuffle=True,\n",
    "                validation_data=(x_test, x_test))\n",
    "\n",
    "\n",
    "# encoding result\n",
    "encoded_imgs = encoder.predict(x_test)\n",
    "decoded_imgs = decoder.predict(encoded_imgs)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {},
   "outputs": [],
   "source": [
    "np.set_printoptions(suppress=True)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(10000, 32)"
      ]
     },
     "execution_count": 6,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "encoded_imgs.shape"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "array([[ 6.3763433 , 10.697407  ,  6.32076   , ...,  1.0930234 ,\n",
       "         4.918487  ,  0.96055055],\n",
       "       [10.730854  ,  4.0360937 , 11.519238  , ...,  2.345777  ,\n",
       "         6.7017703 ,  8.453162  ],\n",
       "       [ 0.11287993,  1.6944015 ,  6.056198  , ...,  2.642245  ,\n",
       "         1.2720426 ,  2.938009  ],\n",
       "       ...,\n",
       "       [ 9.286157  ,  4.192349  , 15.971061  , ...,  7.5206766 ,\n",
       "         9.507512  ,  4.18503   ],\n",
       "       [ 9.254978  ,  0.        , 10.685984  , ...,  2.9746552 ,\n",
       "         8.3398485 ,  3.0615594 ],\n",
       "       [ 9.218382  ,  6.188239  , 18.660156  , ...,  9.215693  ,\n",
       "        13.732598  , 18.18651   ]], dtype=float32)"
      ]
     },
     "execution_count": 7,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "encoded_imgs"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "array([[0.00003127, 0.0000094 , 0.00001687, ..., 0.00002388, 0.00000471,\n",
       "        0.00001549],\n",
       "       [0.00000545, 0.00000814, 0.00004065, ..., 0.00006705, 0.00001795,\n",
       "        0.00001101],\n",
       "       [0.00021891, 0.00011071, 0.00018751, ..., 0.00011635, 0.00006125,\n",
       "        0.00013148],\n",
       "       ...,\n",
       "       [0.00000003, 0.00000001, 0.00000001, ..., 0.00000001, 0.00000001,\n",
       "        0.00000001],\n",
       "       [0.00000085, 0.00000016, 0.00000244, ..., 0.00000174, 0.00000092,\n",
       "        0.00000097],\n",
       "       [0.        , 0.        , 0.00000026, ..., 0.00000004, 0.00000004,\n",
       "        0.        ]], dtype=float32)"
      ]
     },
     "execution_count": 8,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "decoded_imgs"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 1440x288 with 20 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    }
   ],
   "source": [
    "n = 10\n",
    "plt.figure(figsize=(20, 4))\n",
    "for i in range(n):\n",
    "    ax = plt.subplot(2, n, i + 1)\n",
    "    plt.imshow(x_test[i].reshape(28, 28))\n",
    "    plt.gray()\n",
    "    ax.get_xaxis().set_visible(False)\n",
    "    ax.get_yaxis().set_visible(False)\n",
    "\n",
    "    ax = plt.subplot(2, n, i + 1 + n)\n",
    "    plt.imshow(decoded_imgs[i].reshape(28, 28))\n",
    "    plt.gray()\n",
    "    ax.get_xaxis().set_visible(False)\n",
    "    ax.get_yaxis().set_visible(False)\n",
    "plt.show()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(10000, 32)"
      ]
     },
     "execution_count": 10,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "encoded_imgs.shape"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "-----"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Projected vector (feature reduction test)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 11,
   "metadata": {},
   "outputs": [],
   "source": [
    "feature_cols=['x0', 'x1', 'x2', 'x3', 'x4', 'x5', \n",
    "              'x6', 'x7', 'x8', 'x9', 'x10', 'x11', \n",
    "              'x12', 'x13', 'x14', 'x15', 'x16', 'x17', \n",
    "              'x18', 'x19', 'x20', 'x21', 'x22', 'x23', \n",
    "              'x24', 'x25', 'x26', 'x27', 'x28', 'x29', 'x30', 'x31']\n",
    "df = pd.DataFrame(encoded_imgs, columns=feature_cols)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 12,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "k: 1 | cost: 4758724.5801922325\n",
      "k: 2 | cost: 3816152.5562725374\n",
      "k: 3 | cost: 3557510.4202446467\n",
      "k: 4 | cost: 3345392.245199546\n",
      "k: 5 | cost: 3206985.177979756\n",
      "k: 6 | cost: 3079228.072029741\n",
      "k: 7 | cost: 2965349.784726162\n",
      "k: 8 | cost: 2885292.093935602\n",
      "k: 9 | cost: 2814794.611192748\n",
      "k: 10 | cost: 2750466.9586672634\n",
      "k: 11 | cost: 2686893.7413742524\n",
      "k: 12 | cost: 2630499.212320446\n",
      "k: 13 | cost: 2570749.053694904\n",
      "k: 14 | cost: 2522300.11913658\n"
     ]
    },
    {
     "data": {
      "text/plain": [
       "[<matplotlib.lines.Line2D at 0x13a6edef0>]"
      ]
     },
     "execution_count": 12,
     "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": [
    "from sklearn.cluster import KMeans\n",
    "\n",
    "X = df\n",
    "\n",
    "k_list = []\n",
    "cost_list = []\n",
    "for k in range (1, 15):\n",
    "    kmeans = KMeans(n_clusters=k).fit(X)\n",
    "    interia = kmeans.inertia_\n",
    "    print (\"k:\", k, \"| cost:\", interia)\n",
    "    k_list.append(k)\n",
    "    cost_list.append(interia)\n",
    "    \n",
    "plt.plot(k_list, cost_list)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 15,
   "metadata": {},
   "outputs": [],
   "source": [
    "from sklearn.manifold import TSNE\n",
    "\n",
    "model = TSNE(learning_rate=300)\n",
    "transformed = model.fit_transform(X)\n",
    "\n",
    "xs = transformed[:,0]\n",
    "ys = transformed[:,1]\n",
    "plt.scatter(xs,ys)\n",
    "plt.show()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "-----"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Classification using reducted feature"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 32,
   "metadata": {},
   "outputs": [],
   "source": [
    "from keras.models import Sequential\n",
    "from keras.layers import Dense\n",
    "from keras import layers\n",
    "from keras.utils import np_utils\n",
    "import keras\n",
    "\n",
    "y_train = np_utils.to_categorical(y_train)\n",
    "y_test = np_utils.to_categorical(y_test)\n",
    "\n",
    "reducted_x_train = encoder.predict(x_train)\n",
    "reducted_x_test = encoder.predict(x_test)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 33,
   "metadata": {},
   "outputs": [],
   "source": [
    "model=Sequential()\n",
    "model.add(Dense(64, activation='relu', input_dim=encoding_dim))\n",
    "model.add(layers.Dropout(0.2))\n",
    "model.add(Dense(32, activation='relu'))\n",
    "model.add(layers.Dropout(0.2))\n",
    "model.add(Dense(10, activation='softmax'))\n",
    "model.compile(loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy'])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 34,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Train on 48000 samples, validate on 12000 samples\n",
      "Epoch 1/10\n",
      "48000/48000 [==============================] - 1s 25us/step - loss: 1.3601 - acc: 0.5836 - val_loss: 0.4691 - val_acc: 0.8634\n",
      "Epoch 2/10\n",
      "48000/48000 [==============================] - 1s 17us/step - loss: 0.6316 - acc: 0.7998 - val_loss: 0.3434 - val_acc: 0.8994\n",
      "Epoch 3/10\n",
      "48000/48000 [==============================] - 1s 15us/step - loss: 0.5076 - acc: 0.8419 - val_loss: 0.2789 - val_acc: 0.9142\n",
      "Epoch 4/10\n",
      "48000/48000 [==============================] - 1s 16us/step - loss: 0.4422 - acc: 0.8626 - val_loss: 0.2475 - val_acc: 0.9253\n",
      "Epoch 5/10\n",
      "48000/48000 [==============================] - 1s 15us/step - loss: 0.4038 - acc: 0.8777 - val_loss: 0.2302 - val_acc: 0.9323\n",
      "Epoch 6/10\n",
      "48000/48000 [==============================] - 1s 15us/step - loss: 0.3703 - acc: 0.8868 - val_loss: 0.2115 - val_acc: 0.9371\n",
      "Epoch 7/10\n",
      "48000/48000 [==============================] - 1s 16us/step - loss: 0.3475 - acc: 0.8932 - val_loss: 0.1963 - val_acc: 0.9395\n",
      "Epoch 8/10\n",
      "48000/48000 [==============================] - 1s 16us/step - loss: 0.3268 - acc: 0.9009 - val_loss: 0.1923 - val_acc: 0.9411\n",
      "Epoch 9/10\n",
      "48000/48000 [==============================] - 1s 16us/step - loss: 0.3108 - acc: 0.9033 - val_loss: 0.1849 - val_acc: 0.9448\n",
      "Epoch 10/10\n",
      "48000/48000 [==============================] - 1s 16us/step - loss: 0.3034 - acc: 0.9075 - val_loss: 0.1765 - val_acc: 0.9443\n",
      "10000/10000 [==============================] - 0s 5us/step\n",
      "Test Loss and Accuracy -> [0.17936846714699642, 0.9452000051736832]\n"
     ]
    }
   ],
   "source": [
    "history = model.fit(reducted_x_train, y_train, epochs=10, batch_size=100, validation_split=0.2)\n",
    "performance_test = model.evaluate(reducted_x_test, y_test, batch_size=100)\n",
    "print('Test Loss and Accuracy ->', performance_test)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": []
  }
 ],
 "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": "3.7.2"
  },
  "toc": {
   "base_numbering": 1,
   "nav_menu": {},
   "number_sections": true,
   "sideBar": true,
   "skip_h1_title": false,
   "title_cell": "Table of Contents",
   "title_sidebar": "Contents",
   "toc_cell": false,
   "toc_position": {},
   "toc_section_display": true,
   "toc_window_display": false
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}