ch15/ch15_part1.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {
    "id": "bmTf8Nr0ifCU"
   },
   "source": [
    "# 머신 러닝 교과서 3판"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "id": "urDF4hh_ifCV"
   },
   "source": [
    "# 15장 - 심층 합성곱 신경망으로 이미지 분류 (1/2)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "id": "R5cTA0rXifCV"
   },
   "source": [
    "**아래 링크를 통해 이 노트북을 주피터 노트북 뷰어(nbviewer.jupyter.org)로 보거나 구글 코랩(colab.research.google.com)에서 실행할 수 있습니다.**\n",
    "\n",
    "<table class=\"tfo-notebook-buttons\" align=\"left\">\n",
    "  <td>\n",
    "    <a target=\"_blank\" href=\"https://nbviewer.jupyter.org/github/rickiepark/python-machine-learning-book-3rd-edition/blob/master/ch15/ch15_part1.ipynb\"><img src=\"https://jupyter.org/assets/main-logo.svg\" width=\"28\" />주피터 노트북 뷰어로 보기</a>\n",
    "  </td>\n",
    "  <td>\n",
    "    <a target=\"_blank\" href=\"https://colab.research.google.com/github/rickiepark/python-machine-learning-book-3rd-edition/blob/master/ch15/ch15_part1.ipynb\"><img src=\"https://www.tensorflow.org/images/colab_logo_32px.png\" />구글 코랩(Colab)에서 실행하기</a>\n",
    "  </td>\n",
    "</table>"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "id": "f27nCXbrifCW"
   },
   "source": [
    "### 목차"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "id": "2_7fyte5ifCW"
   },
   "source": [
    "- 합성곱 신경망의 구성 요소\n",
    "    - CNN과 특성 계층 학습\n",
    "    - 이산 합성곱 수행\n",
    "        - 1차원 이산 합성곱 수행\n",
    "        - 출력 특성 맵의 크기를 조절하기 위해 입력에 패딩하기\n",
    "        - 합성곱 출력 크기 계산\n",
    "        - 2D 이산 합성곱 수행\n",
    "    - 서브샘플링\n",
    "- 기본 구성 요소를 사용하여 심층 합성곱 신경망 구성\n",
    "    - 여러 개의 입력 또는 컬러 채널 다루기\n",
    "    - 드롭아웃으로 신경망 규제\n",
    "    - 분류를 위한 손실 함수\n",
    "- 텐서플로를 사용하여 심층 합성곱 신경망 구현\n",
    "    - 다층 CNN 구조\n",
    "    - 데이터 적재와 전처리\n",
    "    - 텐서플로 케라스 API를 사용해 CNN 구현하기\n",
    "        - 케라스에서 CNN 층 설정하기\n",
    "        - 케라스로 CNN 구성하기"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {
    "id": "177002meifCW"
   },
   "outputs": [],
   "source": [
    "from IPython.display import Image"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "id": "QhoWUw-DifCW"
   },
   "source": [
    "## 합성곱 신경망의 구성 요소\n",
    "\n",
    "### CNN과 특성 계층 학습"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {
    "colab": {
     "base_uri": "https://localhost:8080/",
     "height": 304
    },
    "id": "Z1yJcNPqifCW",
    "outputId": "6d14c4aa-960d-49c7-c752-e21b8018b294"
   },
   "outputs": [
    {
     "data": {
      "text/html": [
       "<img src=\"https://git.io/JL5O3\" width=\"700\"/>"
      ],
      "text/plain": [
       "<IPython.core.display.Image object>"
      ]
     },
     "execution_count": 2,
     "metadata": {
      "tags": []
     },
     "output_type": "execute_result"
    }
   ],
   "source": [
    "Image(url='https://git.io/JL5O3', width=700)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "id": "n1CfLNUSifCX"
   },
   "source": [
    "### 이산 합성곱 수행\n",
    "\n",
    "#### 1차원 이산 합성곱 수행"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {
    "colab": {
     "base_uri": "https://localhost:8080/",
     "height": 278
    },
    "id": "GKOpZy2TifCX",
    "outputId": "c6e834ec-e540-4706-f40d-08e484ff11bc"
   },
   "outputs": [
    {
     "data": {
      "text/html": [
       "<img src=\"https://git.io/JL5On\" width=\"700\"/>"
      ],
      "text/plain": [
       "<IPython.core.display.Image object>"
      ]
     },
     "execution_count": 3,
     "metadata": {
      "tags": []
     },
     "output_type": "execute_result"
    }
   ],
   "source": [
    "Image(url='https://git.io/JL5On', width=700)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {
    "colab": {
     "base_uri": "https://localhost:8080/",
     "height": 453
    },
    "id": "E3J3_NTNifCY",
    "outputId": "22098915-c7f7-4733-94c1-1ad63b094ae0"
   },
   "outputs": [
    {
     "data": {
      "text/html": [
       "<img src=\"https://git.io/JL5O8\" width=\"700\"/>"
      ],
      "text/plain": [
       "<IPython.core.display.Image object>"
      ]
     },
     "execution_count": 4,
     "metadata": {
      "tags": []
     },
     "output_type": "execute_result"
    }
   ],
   "source": [
    "Image(url='https://git.io/JL5O8', width=700)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "id": "4FylyB5IifCY"
   },
   "source": [
    "#### 출력 특성 맵의 크기를 조절하기 위해 입력에 패딩하기"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {
    "colab": {
     "base_uri": "https://localhost:8080/",
     "height": 211
    },
    "id": "atIWe4jRifCY",
    "outputId": "9a4d73bb-0d2c-4e8f-ed00-fb19b5283c5f"
   },
   "outputs": [
    {
     "data": {
      "text/html": [
       "<img src=\"https://git.io/JL5Ow\" width=\"700\"/>"
      ],
      "text/plain": [
       "<IPython.core.display.Image object>"
      ]
     },
     "execution_count": 5,
     "metadata": {
      "tags": []
     },
     "output_type": "execute_result"
    }
   ],
   "source": [
    "Image(url='https://git.io/JL5Ow', width=700)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "id": "fjCOUYGWifCY"
   },
   "source": [
    "#### 합성곱 출력 크기 계산"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "metadata": {
    "colab": {
     "base_uri": "https://localhost:8080/"
    },
    "id": "kLmHMacdifCY",
    "outputId": "66f041a9-276b-4087-8c97-1aa2bc7a0ee8"
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "텐서플로 버전: 2.4.0\n",
      "넘파이 버전:  1.19.4\n"
     ]
    }
   ],
   "source": [
    "import tensorflow as tf\n",
    "import numpy as np\n",
    "\n",
    "print('텐서플로 버전:', tf.__version__)\n",
    "print('넘파이 버전: ', np.__version__)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {
    "colab": {
     "base_uri": "https://localhost:8080/"
    },
    "id": "15FDd_AMifCZ",
    "outputId": "4436b866-38df-49f2-d95b-e469bc304e6b"
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Conv1d 구현: [ 5. 14. 16. 26. 24. 34. 19. 22.]\n",
      "넘파이 결과: [ 5 14 16 26 24 34 19 22]\n"
     ]
    }
   ],
   "source": [
    "def conv1d(x, w, p=0, s=1):\n",
    "    w_rot = np.array(w[::-1])\n",
    "    x_padded = np.array(x)\n",
    "    if p > 0:\n",
    "        zero_pad = np.zeros(shape=p)\n",
    "        x_padded = np.concatenate(\n",
    "            [zero_pad, x_padded, zero_pad])\n",
    "    res = []\n",
    "    for i in range(0, int((len(x_padded) - len(w_rot)) / s) + 1, s):\n",
    "        res.append(np.sum(\n",
    "            x_padded[i:i+w_rot.shape[0]] * w_rot))\n",
    "    return np.array(res)\n",
    "\n",
    "\n",
    "## 테스트:\n",
    "x = [1, 3, 2, 4, 5, 6, 1, 3]\n",
    "w = [1, 0, 3, 1, 2]\n",
    "\n",
    "print('Conv1d 구현:',\n",
    "      conv1d(x, w, p=2, s=1))\n",
    "\n",
    "print('넘파이 결과:',\n",
    "      np.convolve(x, w, mode='same')) "
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "id": "22WjlUkgifCZ"
   },
   "source": [
    "#### 2D 이산 합성곱 수행"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "metadata": {
    "colab": {
     "base_uri": "https://localhost:8080/",
     "height": 413
    },
    "id": "vOAuCO9kifCZ",
    "outputId": "dac854c7-fe2c-4240-c215-09afefedce58"
   },
   "outputs": [
    {
     "data": {
      "text/html": [
       "<img src=\"https://git.io/JL5OP\" width=\"700\"/>"
      ],
      "text/plain": [
       "<IPython.core.display.Image object>"
      ]
     },
     "execution_count": 8,
     "metadata": {
      "tags": []
     },
     "output_type": "execute_result"
    }
   ],
   "source": [
    "Image(url='https://git.io/JL5OP', width=700)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "metadata": {
    "colab": {
     "base_uri": "https://localhost:8080/",
     "height": 351
    },
    "id": "il5jQ-mZifCZ",
    "outputId": "2ed0ea1e-4a24-4f14-8018-8fb31c46ff61",
    "scrolled": true
   },
   "outputs": [
    {
     "data": {
      "text/html": [
       "<img src=\"https://git.io/JL5OD\" width=\"600\"/>"
      ],
      "text/plain": [
       "<IPython.core.display.Image object>"
      ]
     },
     "execution_count": 9,
     "metadata": {
      "tags": []
     },
     "output_type": "execute_result"
    }
   ],
   "source": [
    "Image(url='https://git.io/JL5OD', width=600)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "metadata": {
    "colab": {
     "base_uri": "https://localhost:8080/",
     "height": 373
    },
    "id": "TVLqLnMXifCa",
    "outputId": "1ee4584d-ed19-497a-a255-0567a406de0e"
   },
   "outputs": [
    {
     "data": {
      "text/html": [
       "<img src=\"https://git.io/JL5OS\" width=\"800\"/>"
      ],
      "text/plain": [
       "<IPython.core.display.Image object>"
      ]
     },
     "execution_count": 10,
     "metadata": {
      "tags": []
     },
     "output_type": "execute_result"
    }
   ],
   "source": [
    "Image(url='https://git.io/JL5OS', width=800)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 11,
   "metadata": {
    "colab": {
     "base_uri": "https://localhost:8080/"
    },
    "id": "I8VBXnCjifCa",
    "outputId": "b71c36c5-3d82-4da9-93ca-c754c3265e03"
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Conv2d 구현:\n",
      " [[11. 25. 32. 13.]\n",
      " [19. 25. 24. 13.]\n",
      " [13. 28. 25. 17.]\n",
      " [11. 17. 14.  9.]]\n",
      "싸이파이 결과:\n",
      " [[11 25 32 13]\n",
      " [19 25 24 13]\n",
      " [13 28 25 17]\n",
      " [11 17 14  9]]\n"
     ]
    }
   ],
   "source": [
    "import scipy.signal\n",
    "\n",
    "\n",
    "def conv2d(X, W, p=(0, 0), s=(1, 1)):\n",
    "    W_rot = np.array(W)[::-1,::-1]\n",
    "    X_orig = np.array(X)\n",
    "    n1 = X_orig.shape[0] + 2*p[0]\n",
    "    n2 = X_orig.shape[1] + 2*p[1]\n",
    "    X_padded = np.zeros(shape=(n1, n2))\n",
    "    X_padded[p[0]:p[0]+X_orig.shape[0],\n",
    "    p[1]:p[1]+X_orig.shape[1]] = X_orig\n",
    "\n",
    "    res = []\n",
    "    for i in range(0, int((X_padded.shape[0] - \n",
    "                           W_rot.shape[0])/s[0])+1, s[0]):\n",
    "        res.append([])\n",
    "        for j in range(0, int((X_padded.shape[1] - \n",
    "                               W_rot.shape[1])/s[1])+1, s[1]):\n",
    "            X_sub = X_padded[i:i+W_rot.shape[0],\n",
    "                             j:j+W_rot.shape[1]]\n",
    "            res[-1].append(np.sum(X_sub * W_rot))\n",
    "    return(np.array(res))\n",
    "\n",
    "X = [[1, 3, 2, 4], [5, 6, 1, 3], [1, 2, 0, 2], [3, 4, 3, 2]]\n",
    "W = [[1, 0, 3], [1, 2, 1], [0, 1, 1]]\n",
    "\n",
    "print('Conv2d 구현:\\n',\n",
    "    conv2d(X, W, p=(1, 1), s=(1, 1)))\n",
    "\n",
    "\n",
    "print('싸이파이 결과:\\n',\n",
    "    scipy.signal.convolve2d(X, W, mode='same'))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "id": "RHePysibifCa"
   },
   "source": [
    "### 서브샘플링"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 12,
   "metadata": {
    "colab": {
     "base_uri": "https://localhost:8080/",
     "height": 365
    },
    "id": "kc1ZMqcLifCa",
    "outputId": "808a9ddc-cbd6-4c12-fcf5-07f7f9c1844e"
   },
   "outputs": [
    {
     "data": {
      "text/html": [
       "<img src=\"https://git.io/JL5OH\" width=\"700\"/>"
      ],
      "text/plain": [
       "<IPython.core.display.Image object>"
      ]
     },
     "execution_count": 12,
     "metadata": {
      "tags": []
     },
     "output_type": "execute_result"
    }
   ],
   "source": [
    "Image(url='https://git.io/JL5OH', width=700)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "id": "57cBF6lxifCb"
   },
   "source": [
    "## 기본 구성 요소를 사용하여 심층 합성곱 신경망 구성\n",
    "\n",
    "### 여러 개의 입력 또는 컬러 채널 다루기"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 13,
   "metadata": {
    "colab": {
     "base_uri": "https://localhost:8080/",
     "height": 421
    },
    "id": "VdJuRn5AifCb",
    "outputId": "fe9a6649-460e-432a-a7d8-c7ff51e1fe94"
   },
   "outputs": [
    {
     "data": {
      "text/html": [
       "<img src=\"https://git.io/JL5O5\" width=\"800\"/>"
      ],
      "text/plain": [
       "<IPython.core.display.Image object>"
      ]
     },
     "execution_count": 13,
     "metadata": {
      "tags": []
     },
     "output_type": "execute_result"
    }
   ],
   "source": [
    "Image(url='https://git.io/JL5O5', width=800)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "id": "ldWAgzDQifCb"
   },
   "source": [
    "**팁: 이미지 파일 읽기**"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 14,
   "metadata": {
    "colab": {
     "base_uri": "https://localhost:8080/"
    },
    "id": "U2_UgtcxvuJj",
    "outputId": "ad42ff96-c402-4b14-bab9-23e8736b1823"
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "--2021-01-02 02:04:04--  https://git.io/JL5Ob\n",
      "Resolving git.io (git.io)... 34.198.20.103, 3.221.203.32, 34.193.233.154, ...\n",
      "Connecting to git.io (git.io)|34.198.20.103|:443... connected.\n",
      "HTTP request sent, awaiting response... 302 Found\n",
      "Location: https://github.com/rickiepark/python-machine-learning-book-3rd-edition/raw/master/ch15/example-image.png [following]\n",
      "--2021-01-02 02:04:05--  https://github.com/rickiepark/python-machine-learning-book-3rd-edition/raw/master/ch15/example-image.png\n",
      "Resolving github.com (github.com)... 140.82.114.4\n",
      "Connecting to github.com (github.com)|140.82.114.4|:443... connected.\n",
      "HTTP request sent, awaiting response... 302 Found\n",
      "Location: https://raw.githubusercontent.com/rickiepark/python-machine-learning-book-3rd-edition/master/ch15/example-image.png [following]\n",
      "--2021-01-02 02:04:05--  https://raw.githubusercontent.com/rickiepark/python-machine-learning-book-3rd-edition/master/ch15/example-image.png\n",
      "Resolving raw.githubusercontent.com (raw.githubusercontent.com)... 151.101.0.133, 151.101.64.133, 151.101.128.133, ...\n",
      "Connecting to raw.githubusercontent.com (raw.githubusercontent.com)|151.101.0.133|:443... connected.\n",
      "HTTP request sent, awaiting response... 200 OK\n",
      "Length: 22283 (22K) [image/png]\n",
      "Saving to: ‘example-image.png’\n",
      "\n",
      "example-image.png   100%[===================>]  21.76K  --.-KB/s    in 0s      \n",
      "\n",
      "2021-01-02 02:04:05 (49.8 MB/s) - ‘example-image.png’ saved [22283/22283]\n",
      "\n"
     ]
    }
   ],
   "source": [
    "# 코랩에서 실행할 경우 이미지를 먼저 다운로드합니다\n",
    "!wget https://git.io/JL5Ob -O example-image.png"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 15,
   "metadata": {
    "colab": {
     "base_uri": "https://localhost:8080/"
    },
    "id": "2lQymxAJifCb",
    "outputId": "ca5496f2-6523-4af2-c66e-b741eb2cc20b"
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "이미지 크기: (252, 221, 3)\n",
      "채널 개수: 3\n",
      "이미지 데이터 타입: <dtype: 'uint8'>\n",
      "tf.Tensor(\n",
      "[[[179 134 110]\n",
      "  [182 136 112]]\n",
      "\n",
      " [[180 135 111]\n",
      "  [182 137 113]]], shape=(2, 2, 3), dtype=uint8)\n"
     ]
    }
   ],
   "source": [
    "import tensorflow as tf\n",
    "\n",
    "\n",
    "img_raw = tf.io.read_file('example-image.png')\n",
    "img = tf.image.decode_image(img_raw)\n",
    "print('이미지 크기:', img.shape)\n",
    "print('채널 개수:', img.shape[2])\n",
    "print('이미지 데이터 타입:', img.dtype)\n",
    "print(img[100:102, 100:102, :])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 16,
   "metadata": {
    "colab": {
     "base_uri": "https://localhost:8080/"
    },
    "id": "qo8Ybg9wifCc",
    "outputId": "db8d73b0-77ba-401d-a9a2-5ce8e3334c1a"
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "이미지 크기: (252, 221, 3)\n",
      "채널 개수: 3\n",
      "이미지 데이터 타입: uint8\n",
      "[[[179 134 110]\n",
      "  [182 136 112]]\n",
      "\n",
      " [[180 135 111]\n",
      "  [182 137 113]]]\n"
     ]
    }
   ],
   "source": [
    "import imageio\n",
    "\n",
    "\n",
    "img = imageio.imread('example-image.png')\n",
    "print('이미지 크기:', img.shape)\n",
    "print('채널 개수:', img.shape[2])\n",
    "print('이미지 데이터 타입:', img.dtype)\n",
    "print(img[100:102, 100:102, :])"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "id": "pRudlk4_ifCd"
   },
   "source": [
    "**노트: CNN 입력을 위한 흑백 이미지의 랭크**"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 17,
   "metadata": {
    "colab": {
     "base_uri": "https://localhost:8080/"
    },
    "id": "LeRSnSZEw45S",
    "outputId": "791a36a3-c619-4430-b58c-40048b26880f"
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "--2021-01-02 02:04:06--  https://git.io/JL5Op\n",
      "Resolving git.io (git.io)... 34.198.20.103, 3.221.203.32, 34.193.233.154, ...\n",
      "Connecting to git.io (git.io)|34.198.20.103|:443... connected.\n",
      "HTTP request sent, awaiting response... 302 Found\n",
      "Location: https://github.com/rickiepark/python-machine-learning-book-3rd-edition/raw/master/ch15/example-image-gray.png [following]\n",
      "--2021-01-02 02:04:06--  https://github.com/rickiepark/python-machine-learning-book-3rd-edition/raw/master/ch15/example-image-gray.png\n",
      "Resolving github.com (github.com)... 140.82.113.3\n",
      "Connecting to github.com (github.com)|140.82.113.3|:443... connected.\n",
      "HTTP request sent, awaiting response... 302 Found\n",
      "Location: https://raw.githubusercontent.com/rickiepark/python-machine-learning-book-3rd-edition/master/ch15/example-image-gray.png [following]\n",
      "--2021-01-02 02:04:06--  https://raw.githubusercontent.com/rickiepark/python-machine-learning-book-3rd-edition/master/ch15/example-image-gray.png\n",
      "Resolving raw.githubusercontent.com (raw.githubusercontent.com)... 151.101.0.133, 151.101.64.133, 151.101.128.133, ...\n",
      "Connecting to raw.githubusercontent.com (raw.githubusercontent.com)|151.101.0.133|:443... connected.\n",
      "HTTP request sent, awaiting response... 200 OK\n",
      "Length: 8501 (8.3K) [image/png]\n",
      "Saving to: ‘example-image-gray.png’\n",
      "\n",
      "example-image-gray. 100%[===================>]   8.30K  --.-KB/s    in 0s      \n",
      "\n",
      "2021-01-02 02:04:06 (90.6 MB/s) - ‘example-image-gray.png’ saved [8501/8501]\n",
      "\n"
     ]
    }
   ],
   "source": [
    "# 코랩에서 실행할 경우 이미지를 먼저 다운로드합니다\n",
    "!wget https://git.io/JL5Op -O example-image-gray.png"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 18,
   "metadata": {
    "colab": {
     "base_uri": "https://localhost:8080/"
    },
    "id": "ZyCzEMJOifCd",
    "outputId": "14b03b8a-fdc1-45bc-e666-1432b097dee8"
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "랭크: 3\n",
      "크기: TensorShape([252, 221, 1])\n"
     ]
    }
   ],
   "source": [
    "img_raw = tf.io.read_file('example-image-gray.png')\n",
    "img = tf.image.decode_image(img_raw)\n",
    "tf.print('랭크:', tf.rank(img))\n",
    "tf.print('크기:', img.shape)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 19,
   "metadata": {
    "colab": {
     "base_uri": "https://localhost:8080/"
    },
    "id": "OwY7M7mxifCd",
    "outputId": "ce6b4c50-3060-483b-8cc7-6280af52ff25"
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "랭크: 2\n",
      "크기: (252, 221)\n",
      "새로운 크기: TensorShape([252, 221, 1])\n"
     ]
    }
   ],
   "source": [
    "img = imageio.imread('example-image-gray.png')\n",
    "tf.print('랭크:', tf.rank(img))\n",
    "tf.print('크기:', img.shape)\n",
    "\n",
    "img_reshaped = tf.reshape(img, (img.shape[0], img.shape[1], 1))\n",
    "tf.print('새로운 크기:', img_reshaped.shape)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "id": "Zy4sB8jVifCd"
   },
   "source": [
    "### 드롭아웃으로 신경망 규제"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 20,
   "metadata": {
    "colab": {
     "base_uri": "https://localhost:8080/",
     "height": 379
    },
    "id": "Lrc2CPzwifCe",
    "outputId": "0e5ca32a-381a-4c74-c13d-b457bcb15c42"
   },
   "outputs": [
    {
     "data": {
      "text/html": [
       "<img src=\"https://git.io/JL5Oh\" width=\"700\"/>"
      ],
      "text/plain": [
       "<IPython.core.display.Image object>"
      ]
     },
     "execution_count": 20,
     "metadata": {
      "tags": []
     },
     "output_type": "execute_result"
    }
   ],
   "source": [
    "Image(url='https://git.io/JL5Oh', width=700)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 21,
   "metadata": {
    "id": "Mp-rHuGVifCe"
   },
   "outputs": [],
   "source": [
    "from tensorflow import keras\n",
    "\n",
    "\n",
    "conv_layer = keras.layers.Conv2D(\n",
    "    filters=16, kernel_size=(3, 3),\n",
    "    kernel_regularizer=keras.regularizers.l2(0.001))\n",
    "\n",
    "fc_layer = keras.layers.Dense(\n",
    "    units=16, kernel_regularizer=keras.regularizers.l2(0.001))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "id": "Kv1zkQutifCe"
   },
   "source": [
    "### 분류를 위한 손실 함수\n",
    "\n",
    " * **`BinaryCrossentropy()`**\n",
    "   * `from_logits=False` \n",
    "   * `from_logits=True`\n",
    "\n",
    " * **`CategoricalCrossentropy()`**\n",
    "   * `from_logits=False`\n",
    "   * `from_logits=True`\n",
    "   \n",
    " * **`SparseCategoricalCrossentropy()`**\n",
    "   * `from_logits=False`\n",
    "   * `from_logits=True`"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 22,
   "metadata": {
    "colab": {
     "base_uri": "https://localhost:8080/",
     "height": 357
    },
    "id": "iOAinQEQifCe",
    "outputId": "5314da16-d14d-48c9-97da-2223d4807678"
   },
   "outputs": [
    {
     "data": {
      "text/html": [
       "<img src=\"https://git.io/JL53f\" width=\"800\"/>"
      ],
      "text/plain": [
       "<IPython.core.display.Image object>"
      ]
     },
     "execution_count": 22,
     "metadata": {
      "tags": []
     },
     "output_type": "execute_result"
    }
   ],
   "source": [
    "Image(url='https://git.io/JL53f', width=800)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 23,
   "metadata": {
    "colab": {
     "base_uri": "https://localhost:8080/"
    },
    "id": "B5oj-jDgifCf",
    "outputId": "08140cba-9784-4096-8ad6-a014e30335a3"
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "BCE (확률): 0.3711 (로짓): 0.3711\n",
      "CCE (확률): 0.5996 (로짓): 0.5996\n",
      "Sparse CCE (확률): 0.5996 (로짓): 0.5996\n"
     ]
    }
   ],
   "source": [
    "from distutils.version import LooseVersion as Version\n",
    "\n",
    "\n",
    "####### 이진 크로스 엔트로피\n",
    "bce_probas = tf.keras.losses.BinaryCrossentropy(from_logits=False)\n",
    "bce_logits = tf.keras.losses.BinaryCrossentropy(from_logits=True)\n",
    "\n",
    "logits = tf.constant([0.8])\n",
    "probas = tf.keras.activations.sigmoid(logits)\n",
    "\n",
    "tf.print(\n",
    "    'BCE (확률): {:.4f}'.format(\n",
    "    bce_probas(y_true=[1], y_pred=probas)),\n",
    "    '(로짓): {:.4f}'.format(\n",
    "    bce_logits(y_true=[1], y_pred=logits)))\n",
    "\n",
    "\n",
    "####### 범주형 크로스 엔트로피\n",
    "cce_probas = tf.keras.losses.CategoricalCrossentropy(\n",
    "    from_logits=False)\n",
    "cce_logits = tf.keras.losses.CategoricalCrossentropy(\n",
    "    from_logits=True)\n",
    "\n",
    "logits = tf.constant([[1.5, 0.8, 2.1]])\n",
    "probas = tf.keras.activations.softmax(logits)\n",
    "\n",
    "if Version(tf.__version__) >= '2.3.0':\n",
    "    tf.print(\n",
    "        'CCE (확률): {:.4f}'.format(\n",
    "        cce_probas(y_true=[[0, 0, 1]], y_pred=probas)),\n",
    "        '(로짓): {:.4f}'.format(\n",
    "        cce_logits(y_true=[[0, 0, 1]], y_pred=logits)))\n",
    "else:\n",
    "    tf.print(\n",
    "        'CCE (확률): {:.4f}'.format(\n",
    "        cce_probas(y_true=[0, 0, 1], y_pred=probas)),\n",
    "        '(로짓): {:.4f}'.format(\n",
    "        cce_logits(y_true=[0, 0, 1], y_pred=logits)))\n",
    "\n",
    "####### 희소 범주형 크로스 엔트로피\n",
    "sp_cce_probas = tf.keras.losses.SparseCategoricalCrossentropy(\n",
    "    from_logits=False)\n",
    "sp_cce_logits = tf.keras.losses.SparseCategoricalCrossentropy(\n",
    "    from_logits=True)\n",
    "\n",
    "tf.print(\n",
    "    'Sparse CCE (확률): {:.4f}'.format(\n",
    "    sp_cce_probas(y_true=[2], y_pred=probas)),\n",
    "    '(로짓): {:.4f}'.format(\n",
    "    sp_cce_logits(y_true=[2], y_pred=logits)))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "id": "EcQnpguXifCf"
   },
   "source": [
    "## 텐서플로를 사용하여 심층 합성곱 신경망 구현\n",
    "\n",
    "### 다층 CNN 구조"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 24,
   "metadata": {
    "colab": {
     "base_uri": "https://localhost:8080/",
     "height": 288
    },
    "id": "n893uGaEifCf",
    "outputId": "ddd39a90-f796-4420-80b0-8086036135dd"
   },
   "outputs": [
    {
     "data": {
      "text/html": [
       "<img src=\"https://git.io/JL53U\" width=\"800\"/>"
      ],
      "text/plain": [
       "<IPython.core.display.Image object>"
      ]
     },
     "execution_count": 24,
     "metadata": {
      "tags": []
     },
     "output_type": "execute_result"
    }
   ],
   "source": [
    "Image(url='https://git.io/JL53U', width=800)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "id": "klgI_t62ifCf"
   },
   "source": [
    "### 데이터 적재와 전처리"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 25,
   "metadata": {
    "id": "qGT2_kbKifCf"
   },
   "outputs": [],
   "source": [
    "import tensorflow_datasets as tfds\n",
    "import pandas as pd\n",
    "\n",
    "import matplotlib.pyplot as plt"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 26,
   "metadata": {
    "colab": {
     "base_uri": "https://localhost:8080/"
    },
    "id": "csMbYS6PifCf",
    "outputId": "eff76e54-c167-411d-b14c-de506e073b76"
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "dict_keys(['test', 'train'])\n"
     ]
    }
   ],
   "source": [
    "## MNIST 데이터셋\n",
    "\n",
    "mnist_bldr = tfds.builder('mnist')\n",
    "mnist_bldr.download_and_prepare()\n",
    "datasets = mnist_bldr.as_dataset(shuffle_files=False)\n",
    "print(datasets.keys())\n",
    "mnist_train_orig, mnist_test_orig = datasets['train'], datasets['test']"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 27,
   "metadata": {
    "id": "B0b0Q15oifCg"
   },
   "outputs": [],
   "source": [
    "BUFFER_SIZE = 10000\n",
    "BATCH_SIZE = 64\n",
    "NUM_EPOCHS = 20"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 28,
   "metadata": {
    "id": "PnBbEngHifCg"
   },
   "outputs": [],
   "source": [
    "mnist_train = mnist_train_orig.map(\n",
    "    lambda item: (tf.cast(item['image'], tf.float32)/255.0, \n",
    "                  tf.cast(item['label'], tf.int32)))\n",
    "\n",
    "mnist_test = mnist_test_orig.map(\n",
    "    lambda item: (tf.cast(item['image'], tf.float32)/255.0, \n",
    "                  tf.cast(item['label'], tf.int32)))\n",
    "\n",
    "tf.random.set_seed(1)\n",
    "\n",
    "mnist_train = mnist_train.shuffle(buffer_size=BUFFER_SIZE,\n",
    "                                  reshuffle_each_iteration=False)\n",
    "\n",
    "mnist_valid = mnist_train.take(10000).batch(BATCH_SIZE)\n",
    "mnist_train = mnist_train.skip(10000).batch(BATCH_SIZE)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "id": "HIa1GMPZifCg"
   },
   "source": [
    "### 텐서플로 케라스 API를 사용해 CNN 구현하기\n",
    "\n",
    "#### 케라스에서 CNN 층 설정하기\n",
    "\n",
    " * **Conv2D:** `tf.keras.layers.Conv2D`\n",
    "   * `filters`\n",
    "   * `kernel_size`\n",
    "   * `strides`\n",
    "   * `padding`\n",
    "   \n",
    "   \n",
    " * **MaxPool2D:** `tf.keras.layers.MaxPool2D`\n",
    "   * `pool_size`\n",
    "   * `strides`\n",
    "   * `padding`\n",
    "   \n",
    "   \n",
    " * **Dropout** `tf.keras.layers.Dropout2D`\n",
    "   * `rate`"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "id": "Dg1drzu5ifCg"
   },
   "source": [
    "#### 케라스로 CNN 구성하기"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 29,
   "metadata": {
    "id": "oZnaVWxQifCg"
   },
   "outputs": [],
   "source": [
    "model = tf.keras.Sequential()\n",
    "\n",
    "model.add(tf.keras.layers.Conv2D(\n",
    "    filters=32, kernel_size=(5, 5),\n",
    "    strides=(1, 1), padding='same',\n",
    "    data_format='channels_last',\n",
    "    name='conv_1', activation='relu'))\n",
    "\n",
    "model.add(tf.keras.layers.MaxPool2D(\n",
    "    pool_size=(2, 2), name='pool_1'))\n",
    "    \n",
    "model.add(tf.keras.layers.Conv2D(\n",
    "    filters=64, kernel_size=(5, 5),\n",
    "    strides=(1, 1), padding='same',\n",
    "    name='conv_2', activation='relu'))\n",
    "\n",
    "model.add(tf.keras.layers.MaxPool2D(pool_size=(2, 2), name='pool_2'))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 30,
   "metadata": {
    "colab": {
     "base_uri": "https://localhost:8080/"
    },
    "id": "v5b9cESoifCg",
    "outputId": "4fce20e6-051d-454a-b8f9-14930d9f76d3"
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "TensorShape([16, 7, 7, 64])"
      ]
     },
     "execution_count": 30,
     "metadata": {
      "tags": []
     },
     "output_type": "execute_result"
    }
   ],
   "source": [
    "model.compute_output_shape(input_shape=(16, 28, 28, 1))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 31,
   "metadata": {
    "colab": {
     "base_uri": "https://localhost:8080/"
    },
    "id": "CinCB6uKifCg",
    "outputId": "c95db01b-a53b-460b-fbcf-1040d7077f7c"
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "TensorShape([16, 3136])"
      ]
     },
     "execution_count": 31,
     "metadata": {
      "tags": []
     },
     "output_type": "execute_result"
    }
   ],
   "source": [
    "model.add(tf.keras.layers.Flatten())\n",
    "    \n",
    "model.compute_output_shape(input_shape=(16, 28, 28, 1))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 32,
   "metadata": {
    "id": "9zQEJSKTifCh"
   },
   "outputs": [],
   "source": [
    "model.add(tf.keras.layers.Dense(\n",
    "    units=1024, name='fc_1', \n",
    "    activation='relu'))\n",
    "\n",
    "model.add(tf.keras.layers.Dropout(\n",
    "    rate=0.5))\n",
    "    \n",
    "model.add(tf.keras.layers.Dense(\n",
    "    units=10, name='fc_2',\n",
    "    activation='softmax'))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 33,
   "metadata": {
    "colab": {
     "base_uri": "https://localhost:8080/"
    },
    "id": "xd3wAxU2ifCh",
    "outputId": "436ffafb-8be3-4b77-f036-6cf2afd80089"
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "TensorShape([16, 10])"
      ]
     },
     "execution_count": 33,
     "metadata": {
      "tags": []
     },
     "output_type": "execute_result"
    }
   ],
   "source": [
    "tf.random.set_seed(1)\n",
    "model.build(input_shape=(None, 28, 28, 1))\n",
    "\n",
    "model.compute_output_shape(input_shape=(16, 28, 28, 1))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 34,
   "metadata": {
    "colab": {
     "base_uri": "https://localhost:8080/"
    },
    "id": "sKsAwaKtifCh",
    "outputId": "62a6d9bb-18f1-4655-e087-b3ea6ea775ef"
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Model: \"sequential\"\n",
      "_________________________________________________________________\n",
      "Layer (type)                 Output Shape              Param #   \n",
      "=================================================================\n",
      "conv_1 (Conv2D)              (None, 28, 28, 32)        832       \n",
      "_________________________________________________________________\n",
      "pool_1 (MaxPooling2D)        (None, 14, 14, 32)        0         \n",
      "_________________________________________________________________\n",
      "conv_2 (Conv2D)              (None, 14, 14, 64)        51264     \n",
      "_________________________________________________________________\n",
      "pool_2 (MaxPooling2D)        (None, 7, 7, 64)          0         \n",
      "_________________________________________________________________\n",
      "flatten (Flatten)            (None, 3136)              0         \n",
      "_________________________________________________________________\n",
      "fc_1 (Dense)                 (None, 1024)              3212288   \n",
      "_________________________________________________________________\n",
      "dropout (Dropout)            (None, 1024)              0         \n",
      "_________________________________________________________________\n",
      "fc_2 (Dense)                 (None, 10)                10250     \n",
      "=================================================================\n",
      "Total params: 3,274,634\n",
      "Trainable params: 3,274,634\n",
      "Non-trainable params: 0\n",
      "_________________________________________________________________\n"
     ]
    }
   ],
   "source": [
    "model.summary()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 35,
   "metadata": {
    "colab": {
     "base_uri": "https://localhost:8080/"
    },
    "id": "9-yPq6g9ifCh",
    "outputId": "4e636dc8-b050-4354-8cc2-9c158a1b0d0e"
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Epoch 1/20\n",
      "782/782 [==============================] - 12s 10ms/step - loss: 0.3113 - accuracy: 0.9011 - val_loss: 0.0481 - val_accuracy: 0.9844\n",
      "Epoch 2/20\n",
      "782/782 [==============================] - 6s 7ms/step - loss: 0.0480 - accuracy: 0.9849 - val_loss: 0.0417 - val_accuracy: 0.9872\n",
      "Epoch 3/20\n",
      "782/782 [==============================] - 6s 7ms/step - loss: 0.0304 - accuracy: 0.9901 - val_loss: 0.0493 - val_accuracy: 0.9858\n",
      "Epoch 4/20\n",
      "782/782 [==============================] - 6s 7ms/step - loss: 0.0253 - accuracy: 0.9921 - val_loss: 0.0393 - val_accuracy: 0.9892\n",
      "Epoch 5/20\n",
      "782/782 [==============================] - 6s 7ms/step - loss: 0.0184 - accuracy: 0.9939 - val_loss: 0.0441 - val_accuracy: 0.9890\n",
      "Epoch 6/20\n",
      "782/782 [==============================] - 6s 7ms/step - loss: 0.0175 - accuracy: 0.9944 - val_loss: 0.0422 - val_accuracy: 0.9894\n",
      "Epoch 7/20\n",
      "782/782 [==============================] - 6s 7ms/step - loss: 0.0110 - accuracy: 0.9963 - val_loss: 0.0428 - val_accuracy: 0.9901\n",
      "Epoch 8/20\n",
      "782/782 [==============================] - 6s 7ms/step - loss: 0.0115 - accuracy: 0.9962 - val_loss: 0.0456 - val_accuracy: 0.9892\n",
      "Epoch 9/20\n",
      "782/782 [==============================] - 6s 7ms/step - loss: 0.0089 - accuracy: 0.9971 - val_loss: 0.0485 - val_accuracy: 0.9899\n",
      "Epoch 10/20\n",
      "782/782 [==============================] - 6s 7ms/step - loss: 0.0105 - accuracy: 0.9965 - val_loss: 0.0421 - val_accuracy: 0.9898\n",
      "Epoch 11/20\n",
      "782/782 [==============================] - 6s 7ms/step - loss: 0.0093 - accuracy: 0.9970 - val_loss: 0.0446 - val_accuracy: 0.9916\n",
      "Epoch 12/20\n",
      "782/782 [==============================] - 6s 7ms/step - loss: 0.0077 - accuracy: 0.9975 - val_loss: 0.0578 - val_accuracy: 0.9890\n",
      "Epoch 13/20\n",
      "782/782 [==============================] - 6s 7ms/step - loss: 0.0073 - accuracy: 0.9982 - val_loss: 0.0508 - val_accuracy: 0.9918\n",
      "Epoch 14/20\n",
      "782/782 [==============================] - 7s 8ms/step - loss: 0.0046 - accuracy: 0.9983 - val_loss: 0.0688 - val_accuracy: 0.9890\n",
      "Epoch 15/20\n",
      "782/782 [==============================] - 6s 7ms/step - loss: 0.0062 - accuracy: 0.9980 - val_loss: 0.0548 - val_accuracy: 0.9907\n",
      "Epoch 16/20\n",
      "782/782 [==============================] - 6s 7ms/step - loss: 0.0100 - accuracy: 0.9970 - val_loss: 0.0592 - val_accuracy: 0.9911\n",
      "Epoch 17/20\n",
      "782/782 [==============================] - 6s 7ms/step - loss: 0.0046 - accuracy: 0.9984 - val_loss: 0.0600 - val_accuracy: 0.9910\n",
      "Epoch 18/20\n",
      "782/782 [==============================] - 6s 7ms/step - loss: 0.0064 - accuracy: 0.9980 - val_loss: 0.0704 - val_accuracy: 0.9897\n",
      "Epoch 19/20\n",
      "782/782 [==============================] - 6s 7ms/step - loss: 0.0052 - accuracy: 0.9983 - val_loss: 0.0638 - val_accuracy: 0.9899\n",
      "Epoch 20/20\n",
      "782/782 [==============================] - 6s 7ms/step - loss: 0.0047 - accuracy: 0.9984 - val_loss: 0.0553 - val_accuracy: 0.9917\n"
     ]
    }
   ],
   "source": [
    "model.compile(optimizer=tf.keras.optimizers.Adam(),\n",
    "              loss=tf.keras.losses.SparseCategoricalCrossentropy(),\n",
    "              metrics=['accuracy']) # same as `tf.keras.metrics.SparseCategoricalAccuracy(name='accuracy')`\n",
    "\n",
    "history = model.fit(mnist_train, epochs=NUM_EPOCHS, \n",
    "                    validation_data=mnist_valid, \n",
    "                    shuffle=True)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 36,
   "metadata": {
    "colab": {
     "base_uri": "https://localhost:8080/",
     "height": 286
    },
    "id": "a06czk7ZifCh",
    "outputId": "a2661698-7225-4b51-f6a2-7b7167b78088"
   },
   "outputs": [
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 864x288 with 2 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light",
      "tags": []
     },
     "output_type": "display_data"
    }
   ],
   "source": [
    "hist = history.history\n",
    "x_arr = np.arange(len(hist['loss'])) + 1\n",
    "\n",
    "fig = plt.figure(figsize=(12, 4))\n",
    "ax = fig.add_subplot(1, 2, 1)\n",
    "ax.plot(x_arr, hist['loss'], '-o', label='Train loss')\n",
    "ax.plot(x_arr, hist['val_loss'], '--<', label='Validation loss')\n",
    "ax.set_xlabel('Epoch', size=15)\n",
    "ax.set_ylabel('Loss', size=15)\n",
    "ax.legend(fontsize=15)\n",
    "ax = fig.add_subplot(1, 2, 2)\n",
    "ax.plot(x_arr, hist['accuracy'], '-o', label='Train acc.')\n",
    "ax.plot(x_arr, hist['val_accuracy'], '--<', label='Validation acc.')\n",
    "ax.legend(fontsize=15)\n",
    "ax.set_xlabel('Epoch', size=15)\n",
    "ax.set_ylabel('Accuracy', size=15)\n",
    "\n",
    "#plt.savefig('images/15_12.png', dpi=300)\n",
    "plt.show()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 37,
   "metadata": {
    "colab": {
     "base_uri": "https://localhost:8080/"
    },
    "id": "d3dx-LhlifCi",
    "outputId": "59ccc88f-5f41-48cf-a9ca-e388aa329c4c"
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "500/500 [==============================] - 2s 3ms/step - loss: 0.0449 - accuracy: 0.9927\n",
      "\n",
      "테스트 정확도 99.27%\n"
     ]
    }
   ],
   "source": [
    "test_results = model.evaluate(mnist_test.batch(20))\n",
    "print('\\n테스트 정확도 {:.2f}%'.format(test_results[1]*100))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 38,
   "metadata": {
    "colab": {
     "base_uri": "https://localhost:8080/",
     "height": 283
    },
    "id": "T7kb_hRaifCi",
    "outputId": "361038a5-2432-4307-d0f8-26f16df2a92d",
    "scrolled": true
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "TensorShape([12, 10])\n",
      "tf.Tensor([2 0 4 8 7 6 0 6 3 1 8 0], shape=(12,), dtype=int64)\n"
     ]
    },
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 864x288 with 12 Axes>"
      ]
     },
     "metadata": {
      "tags": []
     },
     "output_type": "display_data"
    }
   ],
   "source": [
    "batch_test = next(iter(mnist_test.batch(12)))\n",
    "\n",
    "preds = model(batch_test[0])\n",
    "\n",
    "tf.print(preds.shape)\n",
    "preds = tf.argmax(preds, axis=1)\n",
    "print(preds)\n",
    "\n",
    "fig = plt.figure(figsize=(12, 4))\n",
    "for i in range(12):\n",
    "    ax = fig.add_subplot(2, 6, i+1)\n",
    "    ax.set_xticks([]); ax.set_yticks([])\n",
    "    img = batch_test[0][i, :, :, 0]\n",
    "    ax.imshow(img, cmap='gray_r')\n",
    "    ax.text(0.9, 0.1, '{}'.format(preds[i]), \n",
    "            size=15, color='blue',\n",
    "            horizontalalignment='center',\n",
    "            verticalalignment='center', \n",
    "            transform=ax.transAxes)\n",
    "    \n",
    "#plt.savefig('images/15_13.png', dpi=300)\n",
    "plt.show()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 39,
   "metadata": {
    "id": "5sLLwpHLifCi"
   },
   "outputs": [],
   "source": [
    "import os\n",
    "\n",
    "if not os.path.exists('models'):\n",
    "    os.mkdir('models')\n",
    "\n",
    "\n",
    "model.save('models/mnist-cnn.h5')"
   ]
  }
 ],
 "metadata": {
  "accelerator": "GPU",
  "colab": {
   "name": "ch15_part1.ipynb",
   "provenance": []
  },
  "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": "3.7.3"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 4
}